LM audio understanding & sample gen

acestep/constants.py

@@ -69,6 +69,7 @@ TASK_TYPES_BASE = ["text2music", "repaint", "cover", "extract", "lego", "complet
 # Default instructions
 DEFAULT_DIT_INSTRUCTION = "Fill the audio semantic mask based on the given conditions:"
 DEFAULT_LM_INSTRUCTION = "Generate audio semantic tokens based on the given conditions:"
+DEFAULT_LM_UNDERSTAND_INSTRUCTION = "Understand the given musical conditions and describe the audio semantics accordingly:"
 
 # Instruction templates for each task type
 # Note: Some instructions use placeholders like {TRACK_NAME} or {TRACK_CLASSES}

acestep/constrained_logits_processor.py

@@ -35,9 +35,6 @@ class FSMState(Enum):
     DURATION_NAME = auto()       # Generating "duration: "
     DURATION_VALUE = auto()      # Generating numeric value 10-600
     NEWLINE_AFTER_DURATION = auto()
-    GENRES_NAME = auto()         # Generating "genres: "
-    GENRES_VALUE = auto()        # Generating any non-empty string
-    NEWLINE_AFTER_GENRES = auto()
     KEYSCALE_NAME = auto()       # Generating "keyscale: "
     KEYSCALE_VALUE = auto()      # Generating keyscale pattern
     NEWLINE_AFTER_KEYSCALE = auto()
@@ -77,35 +74,28 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         tokenizer: AutoTokenizer,
         enabled: bool = True,
         debug: bool = False,
-        genres_vocab_path: Optional[str] = None,
-        skip_genres: bool = True,
+        **kwargs: Any,
     ):
         """
         Initialize the constrained logits processor.
         
         This processor should be initialized once when loading the LLM and reused
-        for all generations. Use update_caption() before each generation to update
-        the caption-based genre filtering.
-        
+        for all generations.
         Args:
             tokenizer: The tokenizer to use for encoding/decoding
             enabled: Whether to enable constrained decoding
             debug: Whether to print debug information
-            genres_vocab_path: Path to genres vocabulary file (one genre per line)
-                              If None, defaults to "acestep/genres_vocab.txt"
-            skip_genres: Whether to skip genres generation in metadata (default True)
         """
         self.tokenizer = tokenizer
         self.enabled = enabled
         self.debug = debug
-        self.skip_genres = skip_genres
         self.skip_caption = False  # Set to True to skip caption field generation
         self.skip_language = False  # Set to True to skip language field generation
         self.caption: Optional[str] = None  # Set via update_caption() before each generation
         
         # User-provided metadata fields (optional)
         # If provided, these fields will be used directly instead of generating
-        # Format: {"bpm": "120", "caption": "...", "duration": "234", "keyscale": "G major", "language": "en", "timesignature": "4", "genres": "Pop Rock"}
+        # Format: {"bpm": "120", "caption": "...", "duration": "234", "keyscale": "G major", "language": "en", "timesignature": "4"}
         self.user_provided_metadata: Dict[str, Optional[str]] = {
             "bpm": None,
             "caption": None,
@@ -113,7 +103,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             "keyscale": None,
             "language": None,
             "timesignature": None,
-            "genres": None,
         }
         
         # Temperature settings for different generation phases (set per-generation)
@@ -131,6 +120,10 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Stop at reasoning flag - if True, stop generation after </think> tag
         self.stop_at_reasoning: bool = False
         
+        # Generation phase - "cot" or "codes"
+        # Used to determine FSM behavior when prompt already contains CoT
+        self.generation_phase: str = "cot"
+        
         # Current state
         self.state = FSMState.THINK_TAG
         self.position_in_state = 0  # Position within current state's fixed string
@@ -140,6 +133,8 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Caption generation state tracking
         self.caption_after_newline = False  # Track if we're right after a newline in caption
         self.caption_token_count = 0  # Track token count for caption (max 512)
+        self.caption_ending = False  # Track if caption is ending (after detecting non-indented line)
+        self.pending_field_name = ""  # Accumulate field name tokens when caption is ending
         
         # Token queue for user-provided fields (injected directly without generation)
         self.user_field_token_queue: List[int] = []
@@ -147,16 +142,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         
         # Pre-compute token IDs for efficiency
         self._precompute_tokens()
-        
-        # Genres vocabulary for constrained decoding
-        self.genres_vocab_path = genres_vocab_path or os.path.join(
-            os.path.dirname(os.path.abspath(__file__)), "genres_vocab.txt"
-        )
-        self.genres_vocab: List[str] = []  # Full vocab
-        self.genres_vocab_mtime: float = 0.0
-        self.genres_trie: Dict = {}  # Trie for full vocab (fallback)
-        self.caption_genres_trie: Dict = {}  # Trie for caption-matched genres (priority)
-        self.caption_matched_genres: List[str] = []  # Genres matched from caption
+
         self._char_to_tokens: Dict[str, set] = {}  # Precomputed char -> token IDs mapping
         
         # Precompute token mappings once (O(vocab_size), runs once at init)
@@ -199,11 +185,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         
         # Build language prefix tree (similar to keyscale but for language codes)
         self.language_prefix_tree = self._build_language_prefix_tree()
-        
-        self._load_genres_vocab()
-        
-        # Note: Caption-based genre filtering is initialized via update_caption() before each generation
-        
+
         # Fixed strings for each state
         # IMPORTANT: Do NOT include trailing space after colon - tokenizer will handle spacing
         # All matching should be done at token level, not string level
@@ -214,14 +196,13 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             FSMState.BPM_NAME: "bpm:",
             FSMState.CAPTION_NAME: "caption:",
             FSMState.DURATION_NAME: "duration:",
-            FSMState.GENRES_NAME: "genres:",
             FSMState.KEYSCALE_NAME: "keyscale:",
             FSMState.LANGUAGE_NAME: "language:",
             FSMState.TIMESIG_NAME: "timesignature:",
             FSMState.THINK_END_TAG: "</think>",
         }
         
-        # State transitions - build dynamically based on skip_genres
+        # State transitions
         self._build_state_transitions()
     
     def _get_next_field_state(self, current_field: str) -> Optional[FSMState]:
@@ -230,19 +211,17 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         even if the field is user-provided (we still need to generate the field name).
         
         Args:
-            current_field: Current field name ("bpm", "caption", "duration", "genres", "keyscale", "language", "timesignature")
+            current_field: Current field name ("bpm", "caption", "duration", "keyscale", "language", "timesignature")
             
         Returns:
             Next FSMState (NAME state of next field), or THINK_END_TAG if no more fields
         """
         # New field order: bpm -> caption -> duration -> keyscale -> language -> timesignature
-        # genres is optional and can be skipped
-        field_order = ["bpm", "caption", "duration", "genres", "keyscale", "language", "timesignature"]
+        field_order = ["bpm", "caption", "duration","keyscale", "language", "timesignature"]
         field_to_state = {
             "bpm": FSMState.BPM_NAME,
             "caption": FSMState.CAPTION_NAME,
             "duration": FSMState.DURATION_NAME,
-            "genres": FSMState.GENRES_NAME,
             "keyscale": FSMState.KEYSCALE_NAME,
             "language": FSMState.LANGUAGE_NAME,
             "timesignature": FSMState.TIMESIG_NAME,
@@ -256,10 +235,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Find next field in order
         for i in range(current_idx + 1, len(field_order)):
             field = field_order[i]
-            
-            # Skip fields based on flags
-            if field == "genres" and self.skip_genres:
-                continue
+
             if field == "caption" and self.skip_caption:
                 continue
             if field == "language" and self.skip_language:
@@ -272,7 +248,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         return FSMState.THINK_END_TAG
     
     def _build_state_transitions(self):
-        """Build state transition map based on skip_genres and user-provided metadata."""
+        """Build state transition map based on user-provided metadata."""
         self.next_state = {
             FSMState.THINK_TAG: FSMState.NEWLINE_AFTER_THINK,
             FSMState.NEWLINE_AFTER_THINK: FSMState.BPM_NAME,  # Always start with BPM
@@ -281,7 +257,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         }
         
         # Build transitions for all fields (even if user-provided, we still need to generate field name)
-        # Field order: bpm -> caption -> duration -> genres -> keyscale -> language -> timesignature
+        # Field order: bpm -> caption -> duration -> keyscale -> language -> timesignature
         
         # BPM field: NAME -> VALUE -> next field (caption or duration)
         self.next_state[FSMState.BPM_NAME] = FSMState.BPM_VALUE
@@ -296,11 +272,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         self.next_state[FSMState.DURATION_NAME] = FSMState.DURATION_VALUE
         self.next_state[FSMState.DURATION_VALUE] = self._get_next_field_state("duration")
         
-        # Genres field (only if not skipped): NAME -> VALUE -> next field
-        if not self.skip_genres:
-            self.next_state[FSMState.GENRES_NAME] = FSMState.GENRES_VALUE
-            self.next_state[FSMState.GENRES_VALUE] = self._get_next_field_state("genres")
-        
         # Keyscale field: NAME -> VALUE -> next field (language or timesignature)
         self.next_state[FSMState.KEYSCALE_NAME] = FSMState.KEYSCALE_VALUE
         self.next_state[FSMState.KEYSCALE_VALUE] = self._get_next_field_state("keyscale")
@@ -314,11 +285,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         self.next_state[FSMState.TIMESIG_NAME] = FSMState.TIMESIG_VALUE
         self.next_state[FSMState.TIMESIG_VALUE] = FSMState.THINK_END_TAG
     
-    def set_skip_genres(self, skip: bool):
-        """Set whether to skip genres generation and rebuild state transitions."""
-        self.skip_genres = skip
-        self._build_state_transitions()
-    
     def set_skip_caption(self, skip: bool):
         """Set whether to skip caption generation and rebuild state transitions."""
         self.skip_caption = skip
@@ -372,6 +338,21 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         """
         self.stop_at_reasoning = stop
     
+    def set_generation_phase(self, phase: str):
+        """
+        Set the generation phase.
+        
+        Args:
+            phase: "cot" for CoT metadata generation, "codes" for audio codes generation,
+                   or "understand" for audio understanding (codes → metadata + lyrics).
+                   When phase is "codes" and the input prompt already contains </think>,
+                   the FSM will skip metadata generation and go directly to codes generation.
+                   When phase is "understand", generate CoT metadata then free-form lyrics.
+        """
+        if phase not in ("cot", "codes", "understand"):
+            raise ValueError(f"Invalid generation phase: {phase!r}. Must be 'cot', 'codes', or 'understand'")
+        self.generation_phase = phase
+    
     def set_user_metadata(self, metadata: Optional[Dict[str, Optional[str]]] = None):
         """
         Set user-provided metadata fields. Fields that are provided will be used directly
@@ -385,14 +366,13 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                 - "keyscale": Optional[str] - e.g., "G major"
                 - "language": Optional[str] - e.g., "en"
                 - "timesignature": Optional[str] - e.g., "4"
-                - "genres": Optional[str] - e.g., "Pop Rock"
                 If None, clears all user-provided metadata.
         """
         if metadata is None:
             metadata = {}
         
         # Update user-provided metadata
-        for field in ["bpm", "caption", "duration", "keyscale", "language", "timesignature", "genres"]:
+        for field in ["bpm", "caption", "duration", "keyscale", "language", "timesignature"]:
             if field in metadata:
                 self.user_provided_metadata[field] = metadata[field]
             else:
@@ -458,10 +438,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Vocab size
         self.vocab_size = len(self.tokenizer)
         
-        # Comma token for multi-genre support
-        comma_tokens = self.tokenizer.encode(",", add_special_tokens=False)
-        self.comma_token = comma_tokens[-1] if comma_tokens else None
-        
         # EOS token for duration-constrained codes generation
         self.eos_token_id = self.tokenizer.eos_token_id
         
@@ -484,6 +460,8 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Precompute audio code mask for efficient blocking (O(1) instead of O(n))
         # This mask will be added to scores during caption generation
         self.audio_code_mask: Optional[torch.Tensor] = None
+        # Inverse mask: block all non-audio-code tokens (for CODES_GENERATION state)
+        self.non_audio_code_mask: Optional[torch.Tensor] = None
         self._build_audio_code_mask()
         
         # Build valid keyscales set (prefix tree will be built after _char_to_tokens is initialized)
@@ -519,9 +497,13 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         This mask can be added to scores in O(1) time instead of O(n) loop.
         
         The mask is [1, vocab_size] tensor with -inf at audio code token positions.
+        
+        Also builds the inverse mask (non_audio_code_mask) for CODES_GENERATION state,
+        which blocks all non-audio-code tokens.
         """
         if not self.audio_code_token_ids:
             self.audio_code_mask = None
+            self.non_audio_code_mask = None
             return
         
         # Create mask tensor: 0 everywhere, -inf at audio code positions
@@ -536,8 +518,19 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         
         self.audio_code_mask = mask
         
+        # Build inverse mask: -inf everywhere EXCEPT at audio code positions
+        # This is used in CODES_GENERATION state to only allow audio codes
+        inverse_mask = torch.full((1, self.vocab_size), float('-inf'), dtype=torch.float32)
+        inverse_mask[0, audio_code_indices] = 0
+        
+        # Also allow EOS token in codes generation (will be controlled by duration constraint)
+        if self.eos_token_id is not None:
+            inverse_mask[0, self.eos_token_id] = 0
+        
+        self.non_audio_code_mask = inverse_mask
+        
         if self.debug:
-            logger.debug(f"Built audio code mask for {len(self.audio_code_token_ids)} tokens")
+            logger.debug(f"Built audio code masks for {len(self.audio_code_token_ids)} tokens")
     
     def _build_keyscale_prefix_tree(self) -> Dict[Tuple[int, ...], Set[int]]:
         """
@@ -814,133 +807,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             print(f"  {repr(ks)}")
         
         print("=" * 60)
-    
-    def _load_genres_vocab(self):
-        """
-        Load genres vocabulary from file. Supports hot reload by checking file mtime.
-        File format: one genre per line, lines starting with # are comments.
-        """
-        if not os.path.exists(self.genres_vocab_path):
-            if self.debug:
-                logger.debug(f"Genres vocab file not found: {self.genres_vocab_path}")
-            return
-        
-        try:
-            mtime = os.path.getmtime(self.genres_vocab_path)
-            if mtime <= self.genres_vocab_mtime:
-                return  # File hasn't changed
-            
-            with open(self.genres_vocab_path, 'r', encoding='utf-8') as f:
-                genres = []
-                for line in f:
-                    line = line.strip()
-                    if line and not line.startswith('#'):
-                        genres.append(line.lower())
-                
-                self.genres_vocab = genres
-                self.genres_vocab_mtime = mtime
-                self._build_genres_trie()
-                
-                if self.debug:
-                    logger.debug(f"Loaded {len(self.genres_vocab)} genres from {self.genres_vocab_path}")
-        except Exception as e:
-            logger.warning(f"Failed to load genres vocab: {e}")
-    
-    def _build_genres_trie(self):
-        """
-        Build a trie (prefix tree) from genres vocabulary for efficient prefix matching.
-        Each node is a dict with:
-          - '_end': True if this node represents a complete genre
-          - other keys: next characters in the trie
-        """
-        self.genres_trie = {}
-        
-        for genre in self.genres_vocab:
-            node = self.genres_trie
-            for char in genre:
-                if char not in node:
-                    node[char] = {}
-                node = node[char]
-            node['_end'] = True  # Mark end of a complete genre
-        
-        if self.debug:
-            logger.debug(f"Built genres trie with {len(self.genres_vocab)} entries")
-    
-    def _extract_caption_genres(self, caption: str):
-        """
-        Extract genres from the user's caption that match entries in the vocabulary.
-        This creates a smaller trie for faster and more relevant genre generation.
-        
-        Strategy (optimized - O(words * max_genre_len) instead of O(vocab_size)):
-        1. Extract words/phrases from caption
-        2. For each word, use trie to find all vocab entries that START with this word
-        3. Build a separate trie from matched genres
-        """
-        if not caption or not self.genres_vocab:
-            return
-        
-        caption_lower = caption.lower()
-        matched_genres = set()
-        
-        # Extract words from caption (split by common delimiters)
-        import re
-        words = re.split(r'[,\s\-_/\\|]+', caption_lower)
-        words = [w.strip() for w in words if w.strip() and len(w.strip()) >= 2]
-        
-        # For each word, find genres in trie that start with this word
-        for word in words:
-            # Find all genres starting with this word using trie traversal
-            node = self._get_genres_trie_node(word)
-            if node is not None:
-                # Collect all complete genres under this node
-                self._collect_complete_genres(node, word, matched_genres)
-        
-        # Also check if any word appears as a substring in short genres (< 20 chars)
-        # This is a quick check for common single-word genres
-        genres_set = set(self.genres_vocab)
-        for word in words:
-            if word in genres_set:
-                matched_genres.add(word)
-        
-        if not matched_genres:
-            if self.debug:
-                logger.debug(f"No genres matched in caption, using full vocab")
-            return
-        
-        # Build a trie from matched genres
-        self.caption_matched_genres = list(matched_genres)
-        self.caption_genres_trie = {}
-        
-        for genre in matched_genres:
-            node = self.caption_genres_trie
-            for char in genre:
-                if char not in node:
-                    node[char] = {}
-                node = node[char]
-            node['_end'] = True
-        
-        if self.debug:
-            logger.debug(f"Matched {len(matched_genres)} genres from caption: {list(matched_genres)[:5]}...")
-    
-    def _collect_complete_genres(self, node: Dict, prefix: str, result: set, max_depth: int = 50):
-        """
-        Recursively collect all complete genres under a trie node.
-        Limited depth to avoid too many matches.
-        """
-        if max_depth <= 0:
-            return
-        
-        if node.get('_end', False):
-            result.add(prefix)
-        
-        # Limit total collected genres to avoid slowdown
-        if len(result) >= 100:
-            return
-        
-        for char, child_node in node.items():
-            if char not in ('_end', '_tokens'):
-                self._collect_complete_genres(child_node, prefix + char, result, max_depth - 1)
-    
+
     def _precompute_char_token_mapping(self):
         """
         Precompute mapping from characters to token IDs and token decoded texts.
@@ -992,36 +859,8 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         
         if self.debug:
             logger.debug(f"Precomputed char->token mapping for {len(self._char_to_tokens)} unique characters")
-    
-    def _try_reload_genres_vocab(self):
-        """Check if genres vocab file has been updated and reload if necessary."""
-        if not os.path.exists(self.genres_vocab_path):
-            return
-        
-        try:
-            mtime = os.path.getmtime(self.genres_vocab_path)
-            if mtime > self.genres_vocab_mtime:
-                self._load_genres_vocab()
-        except Exception:
-            pass  # Ignore errors during hot reload check
-    
-    def _get_genres_trie_node(self, prefix: str) -> Optional[Dict]:
-        """
-        Get the trie node for a given prefix.
-        Returns None if the prefix is not valid (no genres start with this prefix).
-        """
-        node = self.genres_trie
-        for char in prefix.lower():
-            if char not in node:
-                return None
-            node = node[char]
-        return node
-    
-    def _is_complete_genre(self, text: str) -> bool:
-        """Check if the given text is a complete genre in the vocabulary."""
-        node = self._get_genres_trie_node(text.strip())
-        return node is not None and node.get('_end', False)
-    
+
+
     def _get_trie_node_from_trie(self, trie: Dict, prefix: str) -> Optional[Dict]:
         """Get a trie node from a specific trie (helper for caption vs full trie)."""
         node = trie
@@ -1030,109 +869,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                 return None
             node = node[char]
         return node
-    
-    def _get_allowed_genres_tokens(self) -> List[int]:
-        """
-        Get allowed tokens for genres field based on trie matching.
-        
-        The entire genres string (including commas) must match a complete entry in the vocab.
-        For example, if vocab contains "pop, rock, jazz", the generated string must exactly
-        match that entry - we don't treat commas as separators for individual genres.
-        
-        Strategy:
-        1. If caption-matched genres exist, use that smaller trie first (faster + more relevant)
-        2. If no caption matches or prefix not in caption trie, fallback to full vocab trie
-        3. Get valid next characters from current trie node
-        4. For each candidate token, verify the full decoded text forms a valid trie prefix
-        """
-        if not self.genres_vocab:
-            # No vocab loaded, allow all except newline if empty
-            return []
-        
-        # Use the full accumulated value (don't split by comma - treat as single entry)
-        accumulated = self.accumulated_value.lower()
-        current_genre_prefix = accumulated.strip()
-        
-        # Determine which trie to use: caption-matched (priority) or full vocab (fallback)
-        use_caption_trie = False
-        current_node = None
-        
-        # Try caption-matched trie first if available
-        if self.caption_genres_trie:
-            if current_genre_prefix == "":
-                current_node = self.caption_genres_trie
-                use_caption_trie = True
-            else:
-                current_node = self._get_trie_node_from_trie(self.caption_genres_trie, current_genre_prefix)
-                if current_node is not None:
-                    use_caption_trie = True
-        
-        # Fallback to full vocab trie
-        if current_node is None:
-            if current_genre_prefix == "":
-                current_node = self.genres_trie
-            else:
-                current_node = self._get_genres_trie_node(current_genre_prefix)
-        
-        if current_node is None:
-            # Invalid prefix, force newline to end
-            if self.newline_token:
-                return [self.newline_token]
-            return []
-        
-        # Get valid next characters from trie node
-        valid_next_chars = set(k for k in current_node.keys() if k not in ('_end', '_tokens'))
-        
-        # If current value is a complete genre, allow newline to end
-        is_complete = current_node.get('_end', False)
-        
-        if not valid_next_chars:
-            # No more characters to match, only allow newline if complete
-            allowed = set()
-            if is_complete and self.newline_token:
-                allowed.add(self.newline_token)
-            return list(allowed)
-        
-        # Collect candidate tokens based on first character
-        candidate_tokens = set()
-        for char in valid_next_chars:
-            if char in self._char_to_tokens:
-                candidate_tokens.update(self._char_to_tokens[char])
-        
-        # Select the appropriate trie for validation
-        active_trie = self.caption_genres_trie if use_caption_trie else self.genres_trie
-        
-        # Validate each candidate token: check if prefix + decoded_token is a valid trie prefix
-        allowed = set()
-        for token_id in candidate_tokens:
-            # Use precomputed decoded text (already normalized)
-            decoded_normalized = self._token_to_text.get(token_id, "")
-            
-            if not decoded_normalized or not decoded_normalized.strip():
-                # Token decodes to empty or only whitespace - allow if space/comma is a valid next char
-                if ' ' in valid_next_chars or ',' in valid_next_chars:
-                    allowed.add(token_id)
-                continue
-            
-            # Build new prefix by appending decoded token
-            # Handle space-prefixed tokens (e.g., " rock" from "pop rock")
-            if decoded_normalized.startswith(' ') or decoded_normalized.startswith(','):
-                # Token has leading space/comma - append directly
-                new_prefix = current_genre_prefix + decoded_normalized
-            else:
-                new_prefix = current_genre_prefix + decoded_normalized
-            
-            # Check if new_prefix is a valid prefix in the active trie
-            new_node = self._get_trie_node_from_trie(active_trie, new_prefix)
-            if new_node is not None:
-                allowed.add(token_id)
-        
-        # If current value is a complete genre, also allow newline
-        if is_complete and self.newline_token:
-            allowed.add(self.newline_token)
-        
-        return list(allowed)
-    
+
     def reset(self):
         """Reset the processor state for a new generation."""
         self.state = FSMState.THINK_TAG
@@ -1144,6 +881,8 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         self.current_user_field = None  # Reset current user field
         self.caption_after_newline = False  # Reset caption newline tracking
         self.caption_token_count = 0  # Reset caption token count
+        self.caption_ending = False  # Reset caption ending tracking
+        self.pending_field_name = ""  # Reset pending field name
     
     def set_target_duration(self, duration: Optional[float]):
         """
@@ -1163,27 +902,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             if self.debug:
                 logger.debug("Target duration cleared, no duration constraint")
     
-    def update_caption(self, caption: Optional[str]):
-        """
-        Update the caption and rebuild the caption-matched genres trie.
-        Call this before each generation to prioritize genres from the new caption.
-        
-        Args:
-            caption: User's input caption. If None or empty, clears caption matching.
-        """
-        # Check for hot reload of genres vocabulary
-        self._try_reload_genres_vocab()
-        
-        self.caption = caption
-        self.caption_genres_trie = {}
-        self.caption_matched_genres = []
-        
-        if caption:
-            self._extract_caption_genres(caption)
-        
-        # Also reset FSM state for new generation
-        self.reset()
-    
     def _get_allowed_tokens_for_fixed_string(self, fixed_str: str) -> List[int]:
         """
         Get the token IDs that can continue the fixed string from current position.
@@ -1342,26 +1060,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             logger.debug(f"Numeric field decision: newline_prob={newline_prob:.4f}, max_digit_prob={max_digit_prob:.4f}")
         
         return newline_prob > max_digit_prob
-    
-    def _should_end_text_field(self, logits: torch.Tensor) -> bool:
-        """
-        Determine if we should end a text field (genres).
-        Returns True if P(newline) > P(any other token) AND we have some content.
-        """
-        if not self.accumulated_value.strip():
-            return False  # Need at least some content
-        
-        probs = torch.softmax(logits, dim=-1)
-        newline_prob = probs[0, self.newline_token].item() if self.newline_token else 0
-        
-        # Get max probability among non-newline tokens
-        masked_probs = probs.clone()
-        if self.newline_token:
-            masked_probs[0, self.newline_token] = 0
-        max_other_prob = masked_probs[0].max().item()
-        
-        return newline_prob > max_other_prob
-    
+
     def _get_allowed_keyscale_tokens(self) -> List[int]:
         """
         Get allowed tokens for keyscale field using the precomputed prefix tree.
@@ -1435,9 +1134,32 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             return self._apply_temperature_scaling(scores)
         
         if self.state == FSMState.COMPLETED:
+            # In understanding phase, block audio codes during lyrics generation (COMPLETED state)
+            if self.generation_phase == "understand" and self.audio_code_mask is not None:
+                # Move mask to same device/dtype as scores if needed
+                if self.audio_code_mask.device != scores.device or self.audio_code_mask.dtype != scores.dtype:
+                    self.audio_code_mask = self.audio_code_mask.to(device=scores.device, dtype=scores.dtype)
+                scores = scores + self.audio_code_mask
             return self._apply_temperature_scaling(scores)
         
+        # For codes phase, detect if input already contains </think> and skip to CODES_GENERATION
+        if self.generation_phase == "codes" and self.state == FSMState.THINK_TAG:
+            # Check if input contains </think> token sequence
+            if self._input_contains_think_end_tag(input_ids):
+                # Skip metadata generation, go directly to codes generation
+                self.state = FSMState.CODES_GENERATION
+                self.codes_count = 0
+                if self.debug:
+                    logger.debug("Codes phase: detected </think> in input, skipping to CODES_GENERATION")
+        
         if self.state == FSMState.CODES_GENERATION:
+            # Block all non-audio-code tokens (only allow audio codes and EOS)
+            if self.non_audio_code_mask is not None:
+                # Move mask to same device/dtype as scores if needed
+                if self.non_audio_code_mask.device != scores.device or self.non_audio_code_mask.dtype != scores.dtype:
+                    self.non_audio_code_mask = self.non_audio_code_mask.to(device=scores.device, dtype=scores.dtype)
+                scores = scores + self.non_audio_code_mask
+            
             # Apply duration constraint in codes generation phase
             if self.target_codes is not None and self.eos_token_id is not None:
                 if self.codes_count < self.target_codes:
@@ -1464,6 +1186,31 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         # Apply temperature scaling after constraint masking
         return self._apply_temperature_scaling(scores)
     
+    def _input_contains_think_end_tag(self, input_ids: torch.LongTensor) -> bool:
+        """
+        Check if input contains the </think> closing tag.
+        
+        Args:
+            input_ids: [batch_size, seq_len] input token IDs
+            
+        Returns:
+            True if </think> is found in the input (any sequence in batch)
+        """
+        # Tokenize </think> to get its token sequence
+        think_end_tokens = self.tokenizer.encode("</think>", add_special_tokens=False)
+        if not think_end_tokens:
+            return False
+        
+        # Check each sequence in batch
+        for b in range(input_ids.shape[0]):
+            seq = input_ids[b].tolist()
+            # Search for the token sequence in the input
+            for i in range(len(seq) - len(think_end_tokens) + 1):
+                if seq[i:i+len(think_end_tokens)] == think_end_tokens:
+                    return True
+        
+        return False
+    
     def _apply_temperature_scaling(self, scores: torch.FloatTensor) -> torch.FloatTensor:
         """
         Apply temperature scaling based on current generation phase.
@@ -1501,7 +1248,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
         Uses the same tokenization logic as prefix tree building.
         
         Args:
-            field_name: Field name ("bpm", "caption", "duration", "keyscale", "language", "timesignature", "genres")
+            field_name: Field name ("bpm", "caption", "duration", "keyscale", "language", "timesignature")
             
         Returns:
             List of token IDs for the complete field, or None if field is not provided
@@ -1518,7 +1265,6 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             "keyscale": "keyscale: ",
             "language": "language: ",
             "timesignature": "timesignature: ",
-            "genres": "genres: ",
         }
         prefix = field_to_prefix[field_name]
         full_text = f"{prefix}{value}\n"
@@ -1673,15 +1419,25 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                 
                 # If top token does NOT start with space/tab, it's a new field (like "duration:")
                 if len(top_token_text) > 0 and top_token_text[0] not in ' \t':
-                    # Caption is ending, transition to next field
+                    # Caption is ending - LM is generating next field name
+                    # Instead of forcing state transition to DURATION_NAME (which would regenerate the field name),
+                    # we enter a "caption_ending" mode where we allow free generation until we detect the field value
                     self.caption_after_newline = False
-                    self._transition_to_next_state()
-                    # Process with new state (DURATION_NAME)
-                    return self._process_single_sequence(input_ids, scores)
+                    self.caption_ending = True  # Start tracking field name
+                    self.pending_field_name = ""  # Reset pending field name
+                    # Allow free generation (no constraints) so LM can generate field name naturally
+                    return scores
                 else:
                     # It's indentation, continue caption
                     self.caption_after_newline = False
             
+            # If caption is ending (LM generating next field name), allow free generation
+            # and track the field name until we see colon
+            if self.caption_ending:
+                # Allow any token (free generation)
+                # The field name detection will happen in update_state()
+                return scores
+            
             # Block backticks (code blocks)
             if self.backtick_token is not None:
                 scores[0, self.backtick_token] = float('-inf')
@@ -1749,55 +1505,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                     mask[0, self.newline_token] = 0
                 
                 scores = scores + mask
-        
-        elif self.state == FSMState.GENRES_VALUE:
-            # Check if field is user-provided and we haven't started injecting yet
-            if self.user_provided_metadata["genres"] is not None and not self.user_field_token_queue and not self.accumulated_value:
-                # Initialize token queue with field value tokens (value + newline)
-                value = self.user_provided_metadata["genres"]
-                value_text = f" {value}\n"
-                value_tokens = self.tokenizer.encode(value_text, add_special_tokens=False)
-                if value_tokens:
-                    self.user_field_token_queue = value_tokens
-                    self.current_user_field = "genres"
-                    # Inject first token
-                    mask[0, value_tokens[0]] = 0
-                    scores = scores + mask
-                    return scores
-            
-            # Try to hot-reload genres vocab if file has changed
-            self._try_reload_genres_vocab()
-            
-            # Get allowed tokens based on genres vocabulary
-            allowed = self._get_allowed_genres_tokens()
-            
-            if allowed:
-                # Use vocabulary-constrained decoding
-                for t in allowed:
-                    mask[0, t] = 0
-                scores = scores + mask
-            elif self.genres_vocab:
-                # Vocab is loaded but no valid continuation found
-                # Force newline to end the field
-                if self.newline_token:
-                    mask[0, self.newline_token] = 0
-                    if self.debug:
-                        logger.debug(f"No valid genre continuation for '{self.accumulated_value}', forcing newline")
-                scores = scores + mask
-            else:
-                # Fallback: no vocab loaded, use probability-based ending
-                if self._should_end_text_field(scores):
-                    if self.newline_token:
-                        mask[0, self.newline_token] = 0
-                        self._transition_to_next_state()
-                    scores = scores + mask
-                else:
-                    # Allow any token except newline if we don't have content yet
-                    if not self.accumulated_value.strip():
-                        if self.newline_token:
-                            scores[0, self.newline_token] = float('-inf')
-                    # Otherwise, don't constrain (fallback behavior)
-        
+
         elif self.state == FSMState.KEYSCALE_VALUE:
             # Check if field is user-provided and we haven't started injecting yet
             if self.user_provided_metadata["keyscale"] is not None and not self.user_field_token_queue and not self.accumulated_token_ids:
@@ -1835,7 +1543,10 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                     scores = scores + mask
         
         elif self.state == FSMState.LANGUAGE_VALUE:
-            # Language field: similar to keyscale, uses prefix tree
+            # Language field: Use top-1 probability language (greedy selection)
+            # Unlike other fields, we don't use prefix tree sampling.
+            # Instead, we select the highest probability language at the start,
+            # then force generate the rest of that language code.
             
             # Check if field is user-provided and we haven't started injecting yet
             if self.user_provided_metadata["language"] is not None and not self.user_field_token_queue and not self.accumulated_token_ids:
@@ -1851,25 +1562,63 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                     scores = scores + mask
                     return scores
             
-            # Check if current token sequence is complete (allows newline)
-            token_prefix = tuple(self.accumulated_token_ids)
-            if token_prefix in self.language_prefix_tree and self.newline_token in self.language_prefix_tree[token_prefix]:
-                # Complete language, allow newline
-                if self.newline_token:
-                    mask[0, self.newline_token] = 0
-                scores = scores + mask
-            else:
-                # Not complete, allow valid continuation tokens
-                allowed = self._get_allowed_language_tokens()
-                if allowed:
-                    for t in allowed:
-                        mask[0, t] = 0
-                    scores = scores + mask
+            # If we haven't started generating language yet (empty accumulated_token_ids),
+            # select the top-1 probability token from all valid first tokens
+            if not self.accumulated_token_ids:
+                # Get all possible first tokens for all languages
+                empty_prefix = tuple()
+                if empty_prefix in self.language_prefix_tree:
+                    candidate_tokens = list(self.language_prefix_tree[empty_prefix])
+                    
+                    if candidate_tokens:
+                        # Find the token with highest probability (top-1)
+                        # Create a mask that blocks all tokens except candidates
+                        temp_mask = torch.full_like(scores, float('-inf'))
+                        for t in candidate_tokens:
+                            temp_mask[0, t] = 0
+                        temp_scores = scores + temp_mask
+                        
+                        # Get the highest probability token among candidates
+                        top_token_id = torch.argmax(temp_scores[0]).item()
+                        
+                        # Only allow this top-1 token, block all others (including other language tokens)
+                        mask[0, top_token_id] = 0
+                        scores = scores + mask
+                        
+                        if self.debug:
+                            top_token_text = self.tokenizer.decode([top_token_id])
+                            logger.debug(f"Language field: selected top-1 token {top_token_id} ({repr(top_token_text)}) from {len(candidate_tokens)} candidates")
+                    else:
+                        # No valid first tokens found - force newline
+                        if self.newline_token:
+                            mask[0, self.newline_token] = 0
+                        scores = scores + mask
                 else:
-                    # No valid tokens found - force newline to end field
+                    # Empty prefix not in tree - force newline
                     if self.newline_token:
                         mask[0, self.newline_token] = 0
                     scores = scores + mask
+            else:
+                # We've started generating a language, continue with prefix tree constraints
+                # Check if current token sequence is complete (allows newline)
+                token_prefix = tuple(self.accumulated_token_ids)
+                if token_prefix in self.language_prefix_tree and self.newline_token in self.language_prefix_tree[token_prefix]:
+                    # Complete language, allow newline
+                    if self.newline_token:
+                        mask[0, self.newline_token] = 0
+                    scores = scores + mask
+                else:
+                    # Not complete, allow valid continuation tokens
+                    allowed = self._get_allowed_language_tokens()
+                    if allowed:
+                        for t in allowed:
+                            mask[0, t] = 0
+                        scores = scores + mask
+                    else:
+                        # No valid tokens found - force newline to end field
+                        if self.newline_token:
+                            mask[0, self.newline_token] = 0
+                        scores = scores + mask
         
         elif self.state == FSMState.TIMESIG_VALUE:
             # Check if field is user-provided and we haven't started injecting yet
@@ -1908,12 +1657,18 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             old_state = self.state
             next_state = self.next_state[self.state]
             
-            # If stop_at_reasoning is True and we're transitioning from THINK_END_TAG,
-            # skip CODES_GENERATION and go directly to COMPLETED
-            if self.stop_at_reasoning and old_state == FSMState.THINK_END_TAG:
-                next_state = FSMState.COMPLETED
-                if self.debug:
-                    logger.debug(f"stop_at_reasoning=True: skipping CODES_GENERATION, going directly to COMPLETED")
+            # Handle different cases at THINK_END_TAG based on generation phase
+            # NOTE: Do NOT override next_state here when stop_at_reasoning=True
+            # because we need to transition to the tag state first to generate </think>,
+            # then handle stop_at_reasoning in update_state() AFTER the tag is complete
+            if old_state == FSMState.THINK_END_TAG:
+                if self.generation_phase == "understand":
+                    # Understanding mode: allow free-form lyrics after metadata
+                    # No more constrained decoding after </think>
+                    next_state = FSMState.COMPLETED
+                    if self.debug:
+                        logger.debug(f"generation_phase='understand': allowing free-form lyrics after </think>")
+                # else: default to CODES_GENERATION (for "codes" phase) or respect stop_at_reasoning flag
             
             self.state = next_state
             self.position_in_state = 0
@@ -1921,6 +1676,8 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             self.accumulated_token_ids = []  # Reset token ID sequence for new field
             self.caption_after_newline = False  # Reset caption newline tracking
             self.caption_token_count = 0  # Reset caption token count
+            self.caption_ending = False  # Reset caption ending tracking
+            self.pending_field_name = ""  # Reset pending field name
             if self.debug:
                 logger.debug(f"FSM transition: {old_state.name} -> {self.state.name}")
     
@@ -1991,23 +1748,23 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
             
             # Check if we've completed the fixed string
             if self.position_in_state >= len(fixed_str):
-                self._transition_to_next_state()
+                # Special handling for THINK_END_TAG with stop_at_reasoning
+                if self.state == FSMState.THINK_END_TAG and self.stop_at_reasoning:
+                    # </think> tag is complete, now we can stop generation
+                    # Force transition to COMPLETED instead of CODES_GENERATION
+                    old_state = self.state
+                    self.state = FSMState.COMPLETED
+                    self.position_in_state = 0
+                    self.accumulated_value = ""
+                    self.accumulated_token_ids = []
+                    if self.debug:
+                        logger.debug(f"FSM transition (stop_at_reasoning): {old_state.name} -> {self.state.name}")
+                else:
+                    self._transition_to_next_state()
         
         elif self.state in [FSMState.BPM_VALUE, FSMState.DURATION_VALUE, FSMState.TIMESIG_VALUE]:
             # Accumulate numeric value using token ID sequence
             if generated_token_id == self.newline_token:
-                # if self.state == FSMState.DURATION_VALUE and self.accumulated_value:
-                #     try:
-                #         generated_duration = int(self.accumulated_value)
-                #         if self.target_codes is None and generated_duration > 0:
-                #             self.target_codes = int(generated_duration * 5)
-                #             if self.debug:
-                #                 logger.debug(f"Synced duration: {generated_duration}s -> Set target_codes limit to {self.target_codes}")
-                #     except ValueError:
-                #         if self.debug:
-                #             logger.warning(f"Could not parse duration value: {self.accumulated_value}")
-                # Newline ends the field
-                # Save old state before transition
                 old_state = self.state
                 self._transition_to_next_state()
                 # IMPORTANT: After state transition, if new state is a fixed_strings state,
@@ -2022,21 +1779,7 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                 # Also update legacy accumulated_value for compatibility
                 if token_str.strip().isdigit():
                     self.accumulated_value += token_str.strip()
-        
-        elif self.state == FSMState.GENRES_VALUE:
-            if generated_token_id == self.newline_token:
-                # Newline ends the field
-                self._transition_to_next_state()
-                # IMPORTANT: After state transition, if new state is a fixed_strings state,
-                # we should NOT update position_in_state with the newline token length,
-                # because that token belongs to the old state, not the new state.
-                # Return early to avoid the fixed_strings update logic below.
-                if self.state in self.fixed_strings:
-                    return
-            else:
-                # Genres still uses string-based trie, so keep accumulated_value
-                self.accumulated_value += token_str
-        
+
         elif self.state == FSMState.CAPTION_VALUE:
             # Track token count for 512 limit
             self.caption_token_count += 1
@@ -2049,9 +1792,51 @@ class MetadataConstrainedLogitsProcessor(LogitsProcessor):
                 # Mark that we need to check next token for field transition
                 self.caption_after_newline = True
             else:
-                # Not a newline - if we were after newline and this is not space, 
+                # Not a newline - if we were after newline and this is not space,
                 # transition already happened in _process_single_sequence
                 self.caption_after_newline = False
+            
+            # If caption is ending, accumulate field name and detect field completion
+            if self.caption_ending:
+                self.pending_field_name += token_str
+                
+                # Check if we've completed a field name (detected colon)
+                if ':' in token_str or token_str.strip() == ':':
+                    # Extract field name (before colon)
+                    field_name_full = self.pending_field_name.strip()
+                    # Remove trailing colon if present
+                    field_name = field_name_full.rstrip(':').strip().lower()
+                    
+                    if self.debug:
+                        logger.debug(f"Detected field name after caption: {repr(field_name)}")
+                    
+                    # Map field name to VALUE state
+                    field_name_to_value_state = {
+                        "duration": FSMState.DURATION_VALUE,
+                        "keyscale": FSMState.KEYSCALE_VALUE,
+                        "language": FSMState.LANGUAGE_VALUE,
+                        "timesignature": FSMState.TIMESIG_VALUE,
+                    }
+                    
+                    if field_name in field_name_to_value_state:
+                        # Transition directly to the field's VALUE state
+                        old_state = self.state
+                        self.state = field_name_to_value_state[field_name]
+                        self.position_in_state = 0
+                        self.accumulated_value = ""
+                        self.accumulated_token_ids = []
+                        self.caption_ending = False
+                        self.pending_field_name = ""
+                        
+                        if self.debug:
+                            logger.debug(f"FSM transition (caption ending): {old_state.name} -> {self.state.name}")
+                    else:
+                        # Unknown field name, force transition to next field
+                        if self.debug:
+                            logger.warning(f"Unknown field name after caption: {repr(field_name)}, forcing transition")
+                        self.caption_ending = False
+                        self.pending_field_name = ""
+                        self._transition_to_next_state()
         
         elif self.state == FSMState.KEYSCALE_VALUE:
             if generated_token_id == self.newline_token:

acestep/gradio_ui.py

@@ -363,12 +363,20 @@ def create_generation_section(dit_handler, llm_handler, init_params=None) -> dic
                         
                     # Audio Codes for text2music
                     with gr.Accordion("🎼 LM Codes Hints", open=False, visible=True) as text2music_audio_codes_group:
-                        text2music_audio_code_string = gr.Textbox(
-                            label="LM Codes Hints",
-                            placeholder="<|audio_code_10695|><|audio_code_54246|>...",
-                            lines=6,
-                            info="Paste LM codes hints for text2music generation"
-                        )
+                        with gr.Row(equal_height=True):
+                            text2music_audio_code_string = gr.Textbox(
+                                label="LM Codes Hints",
+                                placeholder="<|audio_code_10695|><|audio_code_54246|>...",
+                                lines=6,
+                                info="Paste LM codes hints for text2music generation",
+                                scale=9,
+                            )
+                            transcribe_btn = gr.Button(
+                                "Transcribe",
+                                variant="secondary",
+                                size="sm",
+                                scale=1,
+                            )
                     
                     # Repainting controls
                     with gr.Group(visible=False) as repainting_group:
@@ -639,6 +647,7 @@ def create_generation_section(dit_handler, llm_handler, init_params=None) -> dic
         "src_audio": src_audio,
         "convert_src_to_codes_btn": convert_src_to_codes_btn,
         "text2music_audio_code_string": text2music_audio_code_string,
+        "transcribe_btn": transcribe_btn,
         "text2music_audio_codes_group": text2music_audio_codes_group,
         "lm_temperature": lm_temperature,
         "lm_cfg_scale": lm_cfg_scale,
@@ -679,6 +688,9 @@ def create_results_section(dit_handler) -> dict:
     with gr.Group():
         gr.HTML('<div class="section-header"><h3>🎧 Generated Results</h3></div>')
         
+        # Hidden state to store LM-generated metadata
+        lm_metadata_state = gr.State(value=None)
+        
         status_output = gr.Textbox(label="Generation Status", interactive=False)
         
         with gr.Row():
@@ -725,6 +737,7 @@ def create_results_section(dit_handler) -> dict:
                     align_plot_2 = gr.Plot(label="Attention Focus Score Heatmap (Sample 2)")
     
     return {
+        "lm_metadata_state": lm_metadata_state,
         "status_output": status_output,
         "generated_audio_1": generated_audio_1,
         "generated_audio_2": generated_audio_2,
@@ -751,7 +764,7 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
             task_type: The task type (e.g., "text2music")
             
         Returns:
-            Tuple of (caption_value, lyrics_value, think_value) for updating UI components
+            Tuple of (caption, lyrics, think, bpm, duration, keyscale, language, timesignature) for updating UI components
         """
         try:
             # Get the project root directory
@@ -764,14 +777,14 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
             # Check if directory exists
             if not os.path.exists(examples_dir):
                 gr.Warning(f"Examples directory not found: examples/{task_type}/")
-                return "", "", True
+                return "", "", True, None, None, "", "", ""
             
             # Find all JSON files in the directory
             json_files = glob.glob(os.path.join(examples_dir, "*.json"))
             
             if not json_files:
                 gr.Warning(f"No JSON files found in examples/{task_type}/")
-                return "", "", True
+                return "", "", True, None, None, "", "", ""
             
             # Randomly select one file
             selected_file = random.choice(json_files)
@@ -796,19 +809,111 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
                 if not isinstance(think_value, bool):
                     think_value = True
                 
-                gr.Info(f"Loaded example from {os.path.basename(selected_file)}")
-                return caption_value, lyrics_value, think_value
+                # Extract optional metadata fields
+                bpm_value = None
+                if 'bpm' in data and data['bpm'] not in [None, "N/A", ""]:
+                    try:
+                        bpm_value = int(data['bpm'])
+                    except (ValueError, TypeError):
+                        pass
+                
+                duration_value = None
+                if 'duration' in data and data['duration'] not in [None, "N/A", ""]:
+                    try:
+                        duration_value = float(data['duration'])
+                    except (ValueError, TypeError):
+                        pass
+                
+                keyscale_value = data.get('keyscale', '')
+                if keyscale_value in [None, "N/A"]:
+                    keyscale_value = ''
+                
+                language_value = data.get('language', '')
+                if language_value in [None, "N/A"]:
+                    language_value = ''
+                
+                timesignature_value = data.get('timesignature', '')
+                if timesignature_value in [None, "N/A"]:
+                    timesignature_value = ''
+                
+                gr.Info(f"📁 Loaded example from {os.path.basename(selected_file)}")
+                return caption_value, lyrics_value, think_value, bpm_value, duration_value, keyscale_value, language_value, timesignature_value
                 
             except json.JSONDecodeError as e:
                 gr.Warning(f"Failed to parse JSON file {os.path.basename(selected_file)}: {str(e)}")
-                return "", "", True
+                return "", "", True, None, None, "", "", ""
             except Exception as e:
                 gr.Warning(f"Error reading file {os.path.basename(selected_file)}: {str(e)}")
-                return "", "", True
+                return "", "", True, None, None, "", "", ""
                 
         except Exception as e:
             gr.Warning(f"Error loading example: {str(e)}")
-            return "", "", True
+            return "", "", True, None, None, "", "", ""
+    
+    def sample_example_smart(task_type: str):
+        """Smart sample function that uses LM if initialized, otherwise falls back to examples
+        
+        Args:
+            task_type: The task type (e.g., "text2music")
+            
+        Returns:
+            Tuple of (caption, lyrics, think, bpm, duration, keyscale, language, timesignature) for updating UI components
+        """
+        # Check if LM is initialized
+        if llm_handler.llm_initialized:
+            # Use LM to generate example
+            try:
+                # Generate example using LM with empty input (NO USER INPUT)
+                metadata, status = llm_handler.understand_audio_from_codes(
+                    audio_codes="NO USER INPUT",
+                    use_constrained_decoding=True,
+                    temperature=0.85,
+                )
+                
+                if metadata:
+                    caption_value = metadata.get('caption', '')
+                    lyrics_value = metadata.get('lyrics', '')
+                    think_value = True  # Always enable think when using LM-generated examples
+                    
+                    # Extract optional metadata fields
+                    bpm_value = None
+                    if 'bpm' in metadata and metadata['bpm'] not in [None, "N/A", ""]:
+                        try:
+                            bpm_value = int(metadata['bpm'])
+                        except (ValueError, TypeError):
+                            pass
+                    
+                    duration_value = None
+                    if 'duration' in metadata and metadata['duration'] not in [None, "N/A", ""]:
+                        try:
+                            duration_value = float(metadata['duration'])
+                        except (ValueError, TypeError):
+                            pass
+                    
+                    keyscale_value = metadata.get('keyscale', '')
+                    if keyscale_value in [None, "N/A"]:
+                        keyscale_value = ''
+                    
+                    language_value = metadata.get('language', '')
+                    if language_value in [None, "N/A"]:
+                        language_value = ''
+                    
+                    timesignature_value = metadata.get('timesignature', '')
+                    if timesignature_value in [None, "N/A"]:
+                        timesignature_value = ''
+                    
+                    gr.Info("🤖 Generated example using LM (Language Model)")
+                    return caption_value, lyrics_value, think_value, bpm_value, duration_value, keyscale_value, language_value, timesignature_value
+                else:
+                    gr.Warning("Failed to generate example using LM, falling back to examples directory")
+                    return load_random_example(task_type)
+                    
+            except Exception as e:
+                gr.Warning(f"Error generating example with LM: {str(e)}, falling back to examples directory")
+                return load_random_example(task_type)
+        else:
+            # LM not initialized, use examples directory
+            return load_random_example(task_type)
     
     def update_init_status(status_msg, enable_btn):
         """Update initialization status and enable/disable generate button"""
@@ -1105,8 +1210,6 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
                 metadata_lines.append(f"- **User Query Rewritten Caption:** {lm_generated_metadata['caption']}")
             if lm_generated_metadata.get('duration'):
                 metadata_lines.append(f"- **Duration:** {lm_generated_metadata['duration']} seconds")
-            if lm_generated_metadata.get('genres'):
-                metadata_lines.append(f"- **Genres:** {lm_generated_metadata['genres']}")
             if lm_generated_metadata.get('keyscale'):
                 metadata_lines.append(f"- **KeyScale:** {lm_generated_metadata['keyscale']}")
             if lm_generated_metadata.get('language'):
@@ -1134,7 +1237,8 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
             align_score_2,
             align_text_2,
             align_plot_2,
-            updated_audio_codes  # Update audio codes in UI
+            updated_audio_codes,  # Update audio codes in UI
+            lm_generated_metadata  # Store metadata for "Send to src audio" buttons
         )
     
     generation_section["generate_btn"].click(
@@ -1186,7 +1290,8 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
             results_section["align_score_2"],
             results_section["align_text_2"],
             results_section["align_plot_2"],
-            generation_section["text2music_audio_code_string"]  # Update audio codes display
+            generation_section["text2music_audio_code_string"],  # Update audio codes display
+            results_section["lm_metadata_state"]  # Store metadata
         ]
     )
     
@@ -1306,33 +1411,214 @@ def setup_event_handlers(demo, dit_handler, llm_handler, dataset_handler, datase
         ]
     )
     
-    # Send generated audio to src_audio
-    def send_audio_to_src(audio_file):
-        """Send generated audio file to src_audio input"""
+    # Send generated audio to src_audio and populate metadata
+    def send_audio_to_src_with_metadata(audio_file, lm_metadata):
+        """Send generated audio file to src_audio input and populate metadata fields
+        
+        Args:
+            audio_file: Audio file path
+            lm_metadata: Dictionary containing LM-generated metadata
+            
+        Returns:
+            Tuple of (audio_file, bpm, caption, duration, key_scale, language, time_signature)
+        """
         if audio_file is None:
-            return None
-        return audio_file
+            return None, None, None, None, None, None, None
+        
+        # Extract metadata fields if available
+        bpm_value = None
+        caption_value = None
+        duration_value = None
+        key_scale_value = None
+        language_value = None
+        time_signature_value = None
+        
+        if lm_metadata:
+            # BPM
+            if lm_metadata.get('bpm'):
+                bpm_str = lm_metadata.get('bpm')
+                if bpm_str and bpm_str != "N/A":
+                    try:
+                        bpm_value = int(bpm_str)
+                    except (ValueError, TypeError):
+                        pass
+            
+            # Caption (Rewritten Caption)
+            if lm_metadata.get('caption'):
+                caption_value = lm_metadata.get('caption')
+            
+            # Duration
+            if lm_metadata.get('duration'):
+                duration_str = lm_metadata.get('duration')
+                if duration_str and duration_str != "N/A":
+                    try:
+                        duration_value = float(duration_str)
+                    except (ValueError, TypeError):
+                        pass
+            
+            # KeyScale
+            if lm_metadata.get('keyscale'):
+                key_scale_str = lm_metadata.get('keyscale')
+                if key_scale_str and key_scale_str != "N/A":
+                    key_scale_value = key_scale_str
+            
+            # Language
+            if lm_metadata.get('language'):
+                language_str = lm_metadata.get('language')
+                if language_str and language_str != "N/A":
+                    language_value = language_str
+            
+            # Time Signature
+            if lm_metadata.get('timesignature'):
+                time_sig_str = lm_metadata.get('timesignature')
+                if time_sig_str and time_sig_str != "N/A":
+                    time_signature_value = time_sig_str
+        
+        return (
+            audio_file,
+            bpm_value,
+            caption_value,
+            duration_value,
+            key_scale_value,
+            language_value,
+            time_signature_value
+        )
     
     results_section["send_to_src_btn_1"].click(
-        fn=send_audio_to_src,
-        inputs=[results_section["generated_audio_1"]],
-        outputs=[generation_section["src_audio"]]
+        fn=send_audio_to_src_with_metadata,
+        inputs=[
+            results_section["generated_audio_1"],
+            results_section["lm_metadata_state"]
+        ],
+        outputs=[
+            generation_section["src_audio"],
+            generation_section["bpm"],
+            generation_section["captions"],
+            generation_section["audio_duration"],
+            generation_section["key_scale"],
+            generation_section["vocal_language"],
+            generation_section["time_signature"]
+        ]
     )
     
     results_section["send_to_src_btn_2"].click(
-        fn=send_audio_to_src,
-        inputs=[results_section["generated_audio_2"]],
-        outputs=[generation_section["src_audio"]]
+        fn=send_audio_to_src_with_metadata,
+        inputs=[
+            results_section["generated_audio_2"],
+            results_section["lm_metadata_state"]
+        ],
+        outputs=[
+            generation_section["src_audio"],
+            generation_section["bpm"],
+            generation_section["captions"],
+            generation_section["audio_duration"],
+            generation_section["key_scale"],
+            generation_section["vocal_language"],
+            generation_section["time_signature"]
+        ]
     )
     
-    # Sample button - load random example
+    # Sample button - smart sample (uses LM if initialized, otherwise examples)
     generation_section["sample_btn"].click(
-        fn=load_random_example,
+        fn=sample_example_smart,
         inputs=[generation_section["task_type"]],
         outputs=[
             generation_section["captions"],
             generation_section["lyrics"],
-            generation_section["think_checkbox"]
+            generation_section["think_checkbox"],
+            generation_section["bpm"],
+            generation_section["audio_duration"],
+            generation_section["key_scale"],
+            generation_section["vocal_language"],
+            generation_section["time_signature"],
+        ]
+    )
+    
+    # Transcribe audio codes to metadata (or generate example if empty)
+    def transcribe_audio_codes(audio_code_string):
+        """
+        Transcribe audio codes to metadata using LLM understanding.
+        If audio_code_string is empty, generate a sample example instead.
+        
+        Args:
+            audio_code_string: String containing audio codes (or empty for example generation)
+            
+        Returns:
+            Tuple of (status_message, caption, lyrics, bpm, duration, keyscale, language, timesignature)
+        """
+        if not llm_handler.llm_initialized:
+            return "❌ 5Hz LM not initialized. Please initialize it first.", "", "", None, None, "", "", ""
+        
+        # If codes are empty, this becomes a "generate example" task
+        # Use "NO USER INPUT" as the input to generate a sample
+        if not audio_code_string or not audio_code_string.strip():
+            audio_code_string = "NO USER INPUT"
+        
+        # Call LLM understanding
+        metadata, status = llm_handler.understand_audio_from_codes(audio_codes=audio_code_string, use_constrained_decoding=True)
+        
+        # Extract fields for UI update
+        caption = metadata.get('caption', '')
+        lyrics = metadata.get('lyrics', '')
+        bpm = metadata.get('bpm')
+        duration = metadata.get('duration')
+        keyscale = metadata.get('keyscale', '')
+        language = metadata.get('language', '')
+        timesignature = metadata.get('timesignature', '')
+        
+        # Convert to appropriate types
+        try:
+            bpm = int(bpm) if bpm and bpm != 'N/A' else None
+        except:
+            bpm = None
+        
+        try:
+            duration = float(duration) if duration and duration != 'N/A' else None
+        except:
+            duration = None
+        
+        return (
+            status,
+            caption,
+            lyrics,
+            bpm,
+            duration,
+            keyscale,
+            language,
+            timesignature
+        )
+    
+    # Update transcribe button text based on whether codes are present
+    def update_transcribe_button_text(audio_code_string):
+        """
+        Update the transcribe button text based on input content.
+        If empty: "Generate Example"
+        If has content: "Transcribe"
+        """
+        if not audio_code_string or not audio_code_string.strip():
+            return gr.update(value="Generate Example")
+        else:
+            return gr.update(value="Transcribe")
+    
+    # Update button text when codes change
+    generation_section["text2music_audio_code_string"].change(
+        fn=update_transcribe_button_text,
+        inputs=[generation_section["text2music_audio_code_string"]],
+        outputs=[generation_section["transcribe_btn"]]
+    )
+    
+    generation_section["transcribe_btn"].click(
+        fn=transcribe_audio_codes,
+        inputs=[generation_section["text2music_audio_code_string"]],
+        outputs=[
+            results_section["status_output"],       # Show status
+            generation_section["captions"],         # Update caption field
+            generation_section["lyrics"],           # Update lyrics field
+            generation_section["bpm"],              # Update BPM field
+            generation_section["audio_duration"],   # Update duration field
+            generation_section["key_scale"],        # Update keyscale field
+            generation_section["vocal_language"],   # Update language field
+            generation_section["time_signature"],   # Update time signature field
         ]
     )
     

acestep/llm_inference.py

@@ -8,6 +8,7 @@ import time
 from typing import Optional, Dict, Any, Tuple, List
 from contextlib import contextmanager
 
+import yaml
 import torch
 from loguru import logger
 from transformers import AutoTokenizer, AutoModelForCausalLM
@@ -17,7 +18,7 @@ from transformers.generation.logits_process import (
     RepetitionPenaltyLogitsProcessor,
 )
 from acestep.constrained_logits_processor import MetadataConstrainedLogitsProcessor
-from acestep.constants import DEFAULT_LM_INSTRUCTION
+from acestep.constants import DEFAULT_LM_INSTRUCTION, DEFAULT_LM_UNDERSTAND_INSTRUCTION
 
 
 class LLMHandler:
@@ -247,6 +248,10 @@ class LLMHandler:
         stop_at_reasoning: bool = False,
         skip_caption: bool = False,
         skip_language: bool = False,
+        generation_phase: str = "cot",
+        caption: str = "",
+        lyrics: str = "",
+        cot_text: str = "",
     ) -> str:
         """Shared vllm path: accept prebuilt formatted prompt and return text."""
         from nanovllm import SamplingParams
@@ -258,12 +263,14 @@ class LLMHandler:
         # Use shared constrained processor if enabled
         constrained_processor = None
         if use_constrained_decoding or use_phase_temperatures:
+            # Reset processor state for new generation
+            self.constrained_processor.reset()
+            
             # Use shared processor, just update caption and settings
             self.constrained_processor.enabled = use_constrained_decoding
             self.constrained_processor.debug = constrained_decoding_debug
             self.constrained_processor.metadata_temperature = metadata_temperature if use_phase_temperatures else None
             self.constrained_processor.codes_temperature = codes_temperature if use_phase_temperatures else None
-            self.constrained_processor.update_caption(formatted_prompt)  # Use formatted prompt for genre extraction
             self.constrained_processor.set_target_duration(target_duration)
             # Always call set_user_metadata to ensure previous settings are cleared if None
             self.constrained_processor.set_user_metadata(user_metadata)
@@ -271,6 +278,8 @@ class LLMHandler:
             # Set skip_caption and skip_language based on flags
             self.constrained_processor.set_skip_caption(skip_caption)
             self.constrained_processor.set_skip_language(skip_language)
+            # Set generation phase for phase-aware processing
+            self.constrained_processor.set_generation_phase(generation_phase)
             
             constrained_processor = self.constrained_processor
 
@@ -286,7 +295,21 @@ class LLMHandler:
         )
 
         if cfg_scale > 1.0:
-            formatted_unconditional_prompt = self.build_formatted_prompt(negative_prompt, is_negative_prompt=True)
+            # Build unconditional prompt based on generation phase
+            if generation_phase == "codes":
+                # Codes phase: use empty CoT in unconditional prompt
+                # formatted_prompt was built with build_formatted_prompt_with_cot(caption, lyrics, cot_text)
+                # For unconditional, we use empty CoT: build_formatted_prompt_with_cot(caption, lyrics, cot_text, is_negative_prompt=True, negative_prompt=...)
+                formatted_unconditional_prompt = self.build_formatted_prompt_with_cot(
+                    caption, lyrics, cot_text, is_negative_prompt=True, negative_prompt=negative_prompt
+                )
+            else:
+                # CoT phase: unconditional prompt
+                # If negative_prompt is provided, use it as caption; otherwise remove caption and keep only lyrics
+                formatted_unconditional_prompt = self.build_formatted_prompt(
+                    caption, lyrics, is_negative_prompt=True, generation_phase="cot", negative_prompt=negative_prompt
+                )
+            
             outputs = self.llm.generate(
                 [formatted_prompt],
                 sampling_params,
@@ -326,6 +349,10 @@ class LLMHandler:
         stop_at_reasoning: bool = False,
         skip_caption: bool = False,
         skip_language: bool = False,
+        generation_phase: str = "cot",
+        caption: str = "",
+        lyrics: str = "",
+        cot_text: str = "",
     ) -> str:
         """Shared PyTorch path: accept prebuilt formatted prompt and return text."""
         inputs = self.llm_tokenizer(
@@ -338,10 +365,12 @@ class LLMHandler:
         # Use shared constrained processor if enabled
         constrained_processor = None
         if use_constrained_decoding:
+            # Reset processor state for new generation
+            self.constrained_processor.reset()
+            
             # Use shared processor, just update caption and settings
             self.constrained_processor.enabled = use_constrained_decoding
             self.constrained_processor.debug = constrained_decoding_debug
-            self.constrained_processor.update_caption(formatted_prompt)  # Use formatted prompt for genre extraction
             self.constrained_processor.set_target_duration(target_duration)
             # Always call set_user_metadata to ensure previous settings are cleared if None
             self.constrained_processor.set_user_metadata(user_metadata)
@@ -349,6 +378,8 @@ class LLMHandler:
             # Set skip_caption and skip_language based on flags
             self.constrained_processor.set_skip_caption(skip_caption)
             self.constrained_processor.set_skip_language(skip_language)
+            # Set generation phase for phase-aware processing
+            self.constrained_processor.set_generation_phase(generation_phase)
             
             constrained_processor = self.constrained_processor
 
@@ -366,7 +397,18 @@ class LLMHandler:
                 logits_processor.append(RepetitionPenaltyLogitsProcessor(penalty=repetition_penalty))
 
             if cfg_scale > 1.0:
-                formatted_unconditional_prompt = self.build_formatted_prompt(negative_prompt, is_negative_prompt=True)
+                # Build unconditional prompt based on generation phase
+                if generation_phase == "codes":
+                    # Codes phase: use empty CoT in unconditional prompt
+                    formatted_unconditional_prompt = self.build_formatted_prompt_with_cot(
+                        caption, lyrics, cot_text, is_negative_prompt=True, negative_prompt=negative_prompt
+                    )
+                else:
+                    # CoT phase: unconditional prompt
+                    # If negative_prompt is provided, use it as caption; otherwise remove caption and keep only lyrics
+                    formatted_unconditional_prompt = self.build_formatted_prompt(
+                        caption, lyrics, is_negative_prompt=True, generation_phase="cot", negative_prompt=negative_prompt
+                    )
                 
                 # Tokenize both prompts together to ensure same length (with left padding)
                 # Left padding is important for generation tasks
@@ -464,9 +506,33 @@ class LLMHandler:
         """Check if all required metadata are present."""
         if user_metadata is None:
             return False
-        if 'bpm' in user_metadata and 'keyscale' in user_metadata and 'timesignature' in user_metadata and 'duration' in user_metadata and 'genres' in user_metadata:
+        if 'bpm' in user_metadata and 'keyscale' in user_metadata and 'timesignature' in user_metadata and 'duration' in user_metadata:
             return True
         return False
+    
+    def _format_metadata_as_cot(self, metadata: Dict[str, Any]) -> str:
+        """
+        Format parsed metadata as CoT text using YAML format (matching training format).
+        
+        Args:
+            metadata: Dictionary with keys: bpm, caption, duration, keyscale, language, timesignature
+            
+        Returns:
+            Formatted CoT text: "<think>\n{yaml_content}\n</think>"
+        """
+        # Build cot_items dict with only non-None values
+        cot_items = {}
+        for key in ['bpm', 'caption', 'duration', 'keyscale', 'language', 'timesignature']:
+            if key in metadata and metadata[key] is not None:
+                cot_items[key] = metadata[key]
+        
+        # Format as YAML (sorted keys, unicode support)
+        if len(cot_items) > 0:
+            cot_yaml = yaml.dump(cot_items, allow_unicode=True, sort_keys=True).strip()
+        else:
+            cot_yaml = ""
+        
+        return f"<think>\n{cot_yaml}\n</think>"
 
     def generate_with_stop_condition(
         self,
@@ -486,10 +552,10 @@ class LLMHandler:
         use_cot_caption: bool = True,
         use_cot_language: bool = True,
     ) -> Tuple[Dict[str, Any], str, str]:
-        """Feishu-compatible LM generation.
+        """Two-phase LM generation: CoT generation followed by audio codes generation.
 
-        - infer_type='dit': stop at </think> and return metas only (no audio codes)
-        - infer_type='llm_dit': normal generation (metas + audio codes)
+        - infer_type='dit': Phase 1 only - generate CoT and return metas (no audio codes)
+        - infer_type='llm_dit': Phase 1 + Phase 2 - generate CoT then audio codes
         
         Args:
             target_duration: Target duration in seconds for codes generation constraint.
@@ -503,17 +569,21 @@ class LLMHandler:
         if infer_type not in {"dit", "llm_dit"}:
             return {}, "", f"❌ invalid infer_type: {infer_type!r} (expected 'dit' or 'llm_dit')"
 
-        # Build formatted prompt
-        formatted_prompt = self.build_formatted_prompt(caption, lyrics)
-
-        # Determine stop condition
-        stop_at_reasoning = (infer_type == "dit")
-        has_all_metas = self.has_all_metas(user_metadata)
+        metadata = {}
         audio_codes = ""
+        has_all_metas = self.has_all_metas(user_metadata)
         
-        if not has_all_metas or not stop_at_reasoning:
-            # For llm_dit mode: use normal generation (stops at EOS)
-            output_text, status = self.generate_from_formatted_prompt(
+        # ========== PHASE 1: CoT Generation ==========
+        # Always generate CoT unless all metadata are user-provided
+        if not has_all_metas:
+            logger.info("Phase 1: Generating CoT metadata...")
+            
+            # Build formatted prompt for CoT phase
+            formatted_prompt = self.build_formatted_prompt(caption, lyrics, generation_phase="cot")
+
+            logger.info(f"generate_with_stop_condition: formatted_prompt={formatted_prompt}")
+            # Generate CoT (stop at </think>)
+            cot_output_text, status = self.generate_from_formatted_prompt(
                 formatted_prompt=formatted_prompt,
                 cfg={
                     "temperature": temperature,
@@ -522,39 +592,121 @@ class LLMHandler:
                     "top_k": top_k,
                     "top_p": top_p,
                     "repetition_penalty": repetition_penalty,
-                    "target_duration": target_duration,
+                    "target_duration": None,  # No duration constraint for CoT phase
                     "user_metadata": user_metadata,
                     "skip_caption": not use_cot_caption,
                     "skip_language": not use_cot_language,
+                    "generation_phase": "cot",
+                    # Pass context for building unconditional prompt in CoT phase
+                    "caption": caption,
+                    "lyrics": lyrics,
                 },
                 use_constrained_decoding=use_constrained_decoding,
                 constrained_decoding_debug=constrained_decoding_debug,
-                stop_at_reasoning=stop_at_reasoning,
+                stop_at_reasoning=True,  # Always stop at </think> in Phase 1
             )
-            if not output_text:
+            
+            if not cot_output_text:
                 return {}, "", status
-
-            # Parse output
-            metadata, audio_codes = self.parse_lm_output(output_text)
-
+            
+            # Parse metadata from CoT output
+            metadata, _ = self.parse_lm_output(cot_output_text)
+            logger.info(f"Phase 1 completed. Generated metadata: {list(metadata.keys())}")
+        else:
+            # Use user-provided metadata
+            logger.info("Phase 1: Using user-provided metadata (skipping generation)")
+            metadata = {k: v for k, v in user_metadata.items() if v is not None}
+        
+        # If infer_type is 'dit', stop here and return only metadata
+        if infer_type == "dit":
+            status_msg = f"✅ Generated CoT metadata successfully\nFields: {', '.join(metadata.keys())}"
+            return metadata, "", status_msg
+        
+        # ========== PHASE 2: Audio Codes Generation ==========
+        logger.info("Phase 2: Generating audio codes...")
+        
+        # Format metadata as CoT using YAML (matching training format)
+        cot_text = self._format_metadata_as_cot(metadata)
+        
+        # Build formatted prompt with CoT for codes generation phase
+        formatted_prompt_with_cot = self.build_formatted_prompt_with_cot(caption, lyrics, cot_text)
+        logger.info(f"generate_with_stop_condition: formatted_prompt_with_cot={formatted_prompt_with_cot}")
+        # Generate audio codes
+        codes_output_text, status = self.generate_from_formatted_prompt(
+            formatted_prompt=formatted_prompt_with_cot,
+            cfg={
+                "temperature": temperature,
+                "cfg_scale": cfg_scale,
+                "negative_prompt": negative_prompt,
+                "top_k": top_k,
+                "top_p": top_p,
+                "repetition_penalty": repetition_penalty,
+                "target_duration": target_duration,
+                "user_metadata": None,  # No user metadata injection in Phase 2
+                "skip_caption": True,  # Skip caption since CoT is already included
+                "skip_language": True,  # Skip language since CoT is already included
+                "generation_phase": "codes",
+                # Pass context for building unconditional prompt in codes phase
+                "caption": caption,
+                "lyrics": lyrics,
+                "cot_text": cot_text,
+            },
+            use_constrained_decoding=use_constrained_decoding,
+            constrained_decoding_debug=constrained_decoding_debug,
+            stop_at_reasoning=False,  # Generate codes until EOS
+        )
+        
+        if not codes_output_text:
+            return metadata, "", status
+        
+        # Parse audio codes from output (metadata should be same as Phase 1)
+        _, audio_codes = self.parse_lm_output(codes_output_text)
+        
         codes_count = len(audio_codes.split('<|audio_code_')) - 1 if audio_codes else 0
-        status_msg = f"✅ Generated successfully\nOutput length: {len(output_text)} chars\nCodes count: {codes_count}"
+        logger.info(f"Phase 2 completed. Generated {codes_count} audio codes")
+        
+        status_msg = f"✅ Generated successfully (2-phase)\nPhase 1: CoT metadata\nPhase 2: {codes_count} audio codes"
         return metadata, audio_codes, status_msg
 
-    def build_formatted_prompt(self, caption: str, lyrics: str = "", is_negative_prompt: bool = False) -> str:
+    def build_formatted_prompt(self, caption: str, lyrics: str = "", is_negative_prompt: bool = False, generation_phase: str = "cot", negative_prompt: str = "NO USER INPUT") -> str:
         """
         Build the chat-formatted prompt for 5Hz LM from caption/lyrics.
         Raises a ValueError if the tokenizer is not initialized.
 
+        Args:
+            caption: Caption text
+            lyrics: Lyrics text
+            is_negative_prompt: If True, builds unconditional prompt for CFG
+            generation_phase: "cot" or "codes" - affects unconditional prompt format
+            negative_prompt: Negative prompt for CFG (used when is_negative_prompt=True)
+            
         Example:
             prompt = handler.build_formatted_prompt("calm piano", "hello world")
         """
         if self.llm_tokenizer is None:
             raise ValueError("LLM tokenizer is not initialized. Call initialize() first.")
+        
         if is_negative_prompt:
-            prompt = caption
+            # Unconditional prompt for CFG
+            # Check if user provided a meaningful negative prompt (not the default)
+            has_negative_prompt = negative_prompt and negative_prompt.strip() and negative_prompt.strip() != "NO USER INPUT"
+            
+            if generation_phase == "cot":
+                # CoT phase unconditional prompt
+                if has_negative_prompt:
+                    # If negative prompt provided, use it as caption
+                    prompt = f"# Caption\n{negative_prompt}\n\n# Lyric\n{lyrics}\n"
+                else:
+                    # No negative prompt: remove caption, keep only lyrics
+                    prompt = f"# Lyric\n{lyrics}\n"
+            else:
+                # Codes phase: will be handled by build_formatted_prompt_with_cot
+                # For backward compatibility, use simple caption as before
+                prompt = caption
         else:
+            # Conditional prompt: include both caption and lyrics
             prompt = f"# Caption\n{caption}\n\n# Lyric\n{lyrics}\n"
+        
         return self.llm_tokenizer.apply_chat_template(
             [
                 {"role": "system", "content": f"# Instruction\n{DEFAULT_LM_INSTRUCTION}\n\n"},
@@ -563,7 +715,258 @@ class LLMHandler:
             tokenize=False,
             add_generation_prompt=True,
         )
-
+    
+    def build_formatted_prompt_with_cot(self, caption: str, lyrics: str, cot_text: str, is_negative_prompt: bool = False, negative_prompt: str = "NO USER INPUT") -> str:
+        """
+        Build the chat-formatted prompt for codes generation phase with pre-generated CoT.
+        
+        Args:
+            caption: Caption text
+            lyrics: Lyrics text  
+            cot_text: Pre-generated CoT text (e.g., "<think>\\nbpm: 120\\n...\\n</think>")
+            is_negative_prompt: If True, uses empty CoT for CFG unconditional prompt
+            negative_prompt: Negative prompt for CFG (used when is_negative_prompt=True)
+            
+        Returns:
+            Formatted prompt string
+            
+        Example:
+            cot = "<think>\\nbpm: 120\\ncaption: calm piano\\n...\\n</think>"
+            prompt = handler.build_formatted_prompt_with_cot("calm piano", "hello", cot)
+        """
+        if self.llm_tokenizer is None:
+            raise ValueError("LLM tokenizer is not initialized. Call initialize() first.")
+        
+        if is_negative_prompt:
+            # Unconditional prompt for codes phase
+            # Check if user provided a meaningful negative prompt
+            has_negative_prompt = negative_prompt and negative_prompt.strip() and negative_prompt.strip() != "NO USER INPUT"
+            
+            # Use empty CoT for unconditional
+            cot_for_prompt = "<think>\n</think>"
+            
+            if has_negative_prompt:
+                # If negative prompt provided, use it as caption
+                caption_for_prompt = negative_prompt
+            else:
+                # No negative prompt: use original caption
+                caption_for_prompt = caption
+        else:
+            # Conditional prompt: use the full CoT and original caption
+            cot_for_prompt = cot_text
+            caption_for_prompt = caption
+        
+        # Build user prompt with caption and lyrics ONLY (no COT)
+        # COT should be in the assistant's message, not user's
+        user_prompt = f"# Caption\n{caption_for_prompt}\n\n# Lyric\n{lyrics}\n"
+        
+        # Build the chat with assistant message containing the COT
+        # The model will continue generation after the COT
+        formatted = self.llm_tokenizer.apply_chat_template(
+            [
+                {"role": "system", "content": f"# Instruction\n{DEFAULT_LM_INSTRUCTION}\n\n"},
+                {"role": "user", "content": user_prompt},
+                {"role": "assistant", "content": cot_for_prompt},
+            ],
+            tokenize=False,
+            add_generation_prompt=False,  # Don't add generation prompt, COT is already in assistant
+        )
+        
+        # Add a newline after </think> so model generates audio codes on next line
+        if not formatted.endswith('\n'):
+            formatted += '\n'
+        
+        return formatted
+    
+    def build_formatted_prompt_for_understanding(
+        self,
+        audio_codes: str,
+        is_negative_prompt: bool = False,
+        negative_prompt: str = "NO USER INPUT"
+    ) -> str:
+        """
+        Build the chat-formatted prompt for audio understanding from codes.
+        
+        This is the reverse of generation: given audio codes, generate metadata and lyrics.
+        
+        Args:
+            audio_codes: Audio code string (e.g., "<|audio_code_123|><|audio_code_456|>...")
+            is_negative_prompt: If True, builds unconditional prompt for CFG
+            negative_prompt: Negative prompt for CFG (used when is_negative_prompt=True)
+            
+        Returns:
+            Formatted prompt string
+            
+        Example:
+            codes = "<|audio_code_18953|><|audio_code_13833|>..."
+            prompt = handler.build_formatted_prompt_for_understanding(codes)
+        """
+        if self.llm_tokenizer is None:
+            raise ValueError("LLM tokenizer is not initialized. Call initialize() first.")
+        
+        # For understanding task, user provides audio codes
+        # Unconditional prompt uses negative_prompt or empty string
+        if is_negative_prompt:
+            user_content = negative_prompt if negative_prompt and negative_prompt.strip() else ""
+        else:
+            user_content = audio_codes
+        
+        return self.llm_tokenizer.apply_chat_template(
+            [
+                {
+                    "role": "system",
+                    "content": f"# Instruction\n{DEFAULT_LM_UNDERSTAND_INSTRUCTION}\n\n"
+                },
+                {
+                    "role": "user",
+                    "content": user_content
+                },
+            ],
+            tokenize=False,
+            add_generation_prompt=True,
+        )
+    
+    def understand_audio_from_codes(
+        self,
+        audio_codes: str,
+        temperature: float = 0.3,
+        cfg_scale: float = 1.0,
+        negative_prompt: str = "NO USER INPUT",
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+        repetition_penalty: float = 1.0,
+        use_constrained_decoding: bool = True,
+        constrained_decoding_debug: bool = False,
+    ) -> Tuple[Dict[str, Any], str]:
+        """
+        Understand audio codes and generate metadata + lyrics.
+        
+        This is the reverse of the normal generation flow:
+        - Input: Audio codes
+        - Output: Metadata (bpm, caption, duration, etc.) + Lyrics
+        
+        Args:
+            audio_codes: String of audio code tokens (e.g., "<|audio_code_123|><|audio_code_456|>...")
+            temperature: Sampling temperature for generation
+            cfg_scale: Classifier-Free Guidance scale (1.0 = no CFG, >1.0 = use CFG)
+            negative_prompt: Negative prompt for CFG
+            top_k: Top-K sampling (None = disabled)
+            top_p: Top-P (nucleus) sampling (None = disabled)
+            repetition_penalty: Repetition penalty (1.0 = no penalty)
+            use_constrained_decoding: Whether to use FSM-based constrained decoding for metadata
+            constrained_decoding_debug: Whether to enable debug logging for constrained decoding
+            
+        Returns:
+            Tuple of (metadata_dict, status_message)
+            metadata_dict contains:
+                - bpm: int or str
+                - caption: str
+                - duration: int or str
+                - genres: str
+                - keyscale: str
+                - language: str
+                - timesignature: str
+                - lyrics: str (extracted from output after </think>)
+        
+        Example:
+            codes = "<|audio_code_18953|><|audio_code_13833|>..."
+            metadata, status = handler.understand_audio_from_codes(codes)
+            print(metadata['caption'])  # "A cinematic orchestral piece..."
+            print(metadata['lyrics'])   # "[Intro: ...]\\n..."
+        """
+        if not getattr(self, "llm_initialized", False):
+            return {}, "❌ 5Hz LM not initialized. Please initialize it first."
+        
+        if not audio_codes or not audio_codes.strip():
+            return {}, "❌ No audio codes provided. Please paste audio codes first."
+        
+        logger.info(f"Understanding audio codes (length: {len(audio_codes)} chars)")
+        
+        # Build formatted prompt for understanding
+        formatted_prompt = self.build_formatted_prompt_for_understanding(audio_codes)
+        print(f"formatted_prompt: {formatted_prompt}")
+        # Generate using constrained decoding (understand phase)
+        # We want to generate metadata first (CoT), then lyrics (natural text)
+        output_text, status = self.generate_from_formatted_prompt(
+            formatted_prompt=formatted_prompt,
+            cfg={
+                "temperature": temperature,
+                "cfg_scale": cfg_scale,
+                "negative_prompt": negative_prompt,
+                "top_k": top_k,
+                "top_p": top_p,
+                "repetition_penalty": repetition_penalty,
+                "target_duration": None,  # No duration constraint for understanding
+                "user_metadata": None,  # No user metadata injection
+                "skip_caption": False,  # Generate caption
+                "skip_language": False,  # Generate language
+                "generation_phase": "understand",  # Understanding phase: generate CoT metadata, then free-form lyrics
+                # Context for building unconditional prompt
+                "caption": "",
+                "lyrics": "",
+            },
+            use_constrained_decoding=use_constrained_decoding,
+            constrained_decoding_debug=constrained_decoding_debug,
+            stop_at_reasoning=False,  # Continue after </think> to generate lyrics
+        )
+        
+        if not output_text:
+            return {}, status
+        
+        # Parse metadata and extract lyrics
+        metadata, _ = self.parse_lm_output(output_text)
+        
+        # Extract lyrics section (everything after </think>)
+        lyrics = self._extract_lyrics_from_output(output_text)
+        if lyrics:
+            metadata['lyrics'] = lyrics
+        
+        logger.info(f"Understanding completed. Generated {len(metadata)} metadata fields")
+        if constrained_decoding_debug:
+            logger.debug(f"Generated metadata: {list(metadata.keys())}")
+            logger.debug(f"Output text preview: {output_text[:200]}...")
+        
+        status_msg = f"✅ Understanding completed successfully\nGenerated fields: {', '.join(metadata.keys())}"
+        return metadata, status_msg
+    
+    def _extract_lyrics_from_output(self, output_text: str) -> str:
+        """
+        Extract lyrics section from LLM output.
+        
+        The lyrics appear after the </think> tag and typically start with "# Lyric"
+        or directly with lyric content.
+        
+        Args:
+            output_text: Full LLM output text
+            
+        Returns:
+            Extracted lyrics string, or empty string if no lyrics found
+        """
+        import re
+        
+        # Find the </think> tag
+        think_end_pattern = r'</think>'
+        match = re.search(think_end_pattern, output_text)
+        
+        if not match:
+            # No </think> tag found, no lyrics
+            return ""
+        
+        # Extract everything after </think>
+        after_think = output_text[match.end():].strip()
+        
+        if not after_think:
+            return ""
+        
+        # Remove "# Lyric" header if present
+        lyric_header_pattern = r'^#\s*Lyri[c|cs]?\s*\n'
+        after_think = re.sub(lyric_header_pattern, '', after_think, flags=re.IGNORECASE)
+        
+        # Remove <|im_end|> tag at the end if present
+        after_think = re.sub(r'<\|im_end\|>\s*$', '', after_think)
+        
+        return after_think.strip()
+    
     def generate_from_formatted_prompt(
         self,
         formatted_prompt: str,
@@ -583,6 +986,7 @@ class LLMHandler:
                 - negative_prompt (str) used when cfg_scale > 1
                 - top_k (int), top_p (float), repetition_penalty (float)
                 - target_duration (float): Target duration in seconds for codes generation
+                - generation_phase (str): "cot" or "codes" for phase-aware CFG
             use_constrained_decoding: Whether to use FSM-based constrained decoding
             constrained_decoding_debug: Whether to enable debug logging for constrained decoding
             stop_at_reasoning: If True, stop generation immediately after </think> tag (no audio codes)
@@ -610,6 +1014,11 @@ class LLMHandler:
         user_metadata = cfg.get("user_metadata")  # User-provided metadata fields
         skip_caption = cfg.get("skip_caption", False)  # Skip caption generation in CoT
         skip_language = cfg.get("skip_language", False)  # Skip language generation in CoT
+        generation_phase = cfg.get("generation_phase", "cot")  # "cot" or "codes"
+        # Additional context for codes phase unconditional prompt building
+        caption = cfg.get("caption", "")
+        lyrics = cfg.get("lyrics", "")
+        cot_text = cfg.get("cot_text", "")
 
         try:
             if self.llm_backend == "vllm":
@@ -628,6 +1037,10 @@ class LLMHandler:
                     stop_at_reasoning=stop_at_reasoning,
                     skip_caption=skip_caption,
                     skip_language=skip_language,
+                    generation_phase=generation_phase,
+                    caption=caption,
+                    lyrics=lyrics,
+                    cot_text=cot_text,
                 )
                 return output_text, f"✅ Generated successfully (vllm) | length={len(output_text)}"
 
@@ -647,6 +1060,10 @@ class LLMHandler:
                 stop_at_reasoning=stop_at_reasoning,
                 skip_caption=skip_caption,
                 skip_language=skip_language,
+                generation_phase=generation_phase,
+                caption=caption,
+                lyrics=lyrics,
+                cot_text=cot_text,
             )
             return output_text, f"✅ Generated successfully (pt) | length={len(output_text)}"
 

examples/text2music/example_01.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "pop, rap, electronic, blues, hip-house, rhythm and blues",
-    "lyrics": "[verse]\n我走过深夜的街道\n冷风吹乱思念的漂亮外套\n你的微笑像星光很炫耀\n照亮了我孤独的每分每秒\n\n[chorus]\n愿你是风吹过我的脸\n带我飞过最远最遥远的山间\n愿你是风轻触我的梦\n停在心头不再飘散无迹无踪\n\n[verse]\n一起在喧哗避开世俗的骚动\n独自在天台探望月色的朦胧\n你说爱像音乐带点重节奏\n一拍一跳让我忘了心的温度多空洞\n\n[bridge]\n唱起对你的想念不隐藏\n像诗又像画写满藏不了的渴望\n你的影子挥不掉像风的倔强\n追着你飞扬穿越云海一样泛光\n\n[chorus]\n愿你是风吹过我的手\n暖暖的触碰像春日细雨温柔\n愿你是风盘绕我的身\n深情万万重不会有一天走远走\n\n[verse]\n深夜的钢琴弹起动人的旋律\n低音鼓砸进心底的每一次呼吸\n要是能将爱化作歌声传递\n你是否会听见我心里的真心实意"
+    "caption": "An explosive, high-energy pop-rock track with a strong anime theme song feel. The song kicks off with a catchy, synthesized brass fanfare over a driving rock beat with punchy drums and a solid bassline. A powerful, clear male vocal enters with a theatrical and energetic delivery, soaring through the verses and hitting powerful high notes in the chorus. The arrangement is dense and dynamic, featuring rhythmic electric guitar chords, brief instrumental breaks with synth flourishes, and a consistent, danceable groove throughout. The overall mood is triumphant, adventurous, and exhilarating.",
+    "lyrics": "[Intro - Synth Brass Fanfare]\n\n[Verse 1]\n黑夜里的风吹过耳畔\n甜蜜时光转瞬即万\n脚步飘摇在星光上\n心追节奏心跳狂乱\n耳边传来电吉他呼唤\n手指轻触碰点流点燃\n梦在云端任它蔓延\n疯狂跳跃自由无间\n\n[Chorus]\n心电感应在震动间\n拥抱未来勇敢冒险\n那旋律在心中无限\n世界变得如此耀眼\n\n[Instrumental Break - Synth Brass Melody]\n\n[Verse 2]\n鼓点撞击黑夜的底端\n跳动节拍连接你我俩\n在这里让灵魂发光\n燃尽所有不留遗憾\n\n[Instrumental Break - Synth Brass Melody]\n\n[Bridge]\n光影交错彼此的视线\n霓虹之下夜空的蔚蓝\n月光洒下温热心田\n追逐梦想它不会遥远\n\n[Chorus]\n心电感应在震动间\n拥抱未来勇敢冒险\n那旋律在心中无限\n世界变得如此耀眼\n\n[Outro - Instrumental with Synth Brass Melody]\n[Song ends abruptly]",
+    "bpm": 100,
+    "duration": 160,
+    "keyscale": "B minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_02.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "country rock, folk rock, southern rock, bluegrass, country pop",
-    "lyrics": "[verse]\nWoke up to the sunrise glow\nTook my heart and hit the road\nWheels hummin' the only tune I know\nStraight to where the wildflowers grow\n\n[verse]\nGot that old map all wrinkled and torn\nDestination unknown but I'm reborn\nWith a smile that the wind has worn\nChasin' dreams that can't be sworn\n\n[chorus]\nRidin' on a highway to sunshine\nGot my shades and my radio on fine\nLeave the shadows in the rearview rhyme\nHeart's racing as we chase the time\n\n[verse]\nMet a girl with a heart of gold\nTold stories that never get old\nHer laugh like a tale that's been told\nA melody so bold yet uncontrolled\n\n[bridge]\nClouds roll by like silent ghosts\nAs we drive along the coast\nWe toast to the days we love the most\nFreedom's song is what we post\n\n[chorus]\nRidin' on a highway to sunshine\nGot my shades and my radio on fine\nLeave the shadows in the rearview rhyme\nHeart's racing as we chase the time"
+    "caption": "A melancholic Latin trap track built on a foundation of deep 808 sub-bass and crisp, rolling hi-hats from a drum machine. A somber synth pad provides an atmospheric backdrop for the emotional male lead vocal, which is treated with noticeable auto-tune and spacious reverb. The chorus introduces layered vocals for added intensity and features prominent echoed ad-libs that drift through the mix. The arrangement includes a brief breakdown where the beat recedes to emphasize the raw vocal delivery before returning to the full instrumental for a final section featuring melodic synth lines over the main groove.",
+    "lyrics": "[Intro]\nYeah\nMe dicen ven, pero yo ya no quiero na'\nTe cansaste de estar pa' mí, ya no hay marcha atrás\nTu esencia se fue sin avisar\nPero este mundo no me va a borrar\n\n[Verse 1]\nNo hay más pa' ti ni de mami, ahora ves\nNo vengas con drama, lo sabes\nMe dijeron pa' siempre, pero mentían\nComo si tú no estuvieras\nSi yo fui tu juego y tú fuiste mi cárcel\nAhora no lloro aunque cierre la cana\nMe arde el aire, no hay vuelta, no hay miedo\nHoy dejo atrás todo lo que me da miedo\nTú me botaste, te fuiste del puerto\nYo me quedé solo, fue pura derrota\n\n[Pre-Chorus]\nNo era de por ti, lo entiendo, mi pana\nTe vas con otros aunque duela la mañana\nY si vuelvo, mira donde he llegado\nYa no hay vuelta, no\nNi tu plan en la mano\n\n[Chorus]\nMe dicen ven, pero yo ya no quiero na'\nNo quiero\nTe cansaste de estar pa' mí, ya no hay marcha atrás\nTu esencia se fue sin avisar\nPero este mundo no me va a borrar\nNo me va a borrar\nSigo solo, sigo vacío, sigo abajo de mi tierra\nAhora dejo atrás lo que me diste\nEn un año me despido, te miro a los ojos\nYa no soy tu tonto\nTe lo avisé, pero nunca intentaste conmigo\n\n[Verse 2]\nTe vi delirar, te vi mirarte\nAhora entiendo que no fuiste amante\nPero ya no me duele, ya no es rabia\nYa no soy el tonto que dejaste en tu cama\n\n[Chorus]\nMe dicen ven, pero yo ya no quiero na'\nNo quiero\nTe cansaste de estar pa' mí, ya no hay marcha atrás\nTu esencia se fue sin avisar\nPero este mundo no me va a borrar\n\n[Outro]\n[Instrumental fades out]",
+    "bpm": 100,
+    "duration": 159,
+    "keyscale": "G# minor",
+    "language": "es",
+    "timesignature": "4"
 }

examples/text2music/example_03.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "hip-house, funk",
-    "lyrics": "[verse]\n哎呀跳起来，脚尖踩节拍 (oo-yeah!)\n灯光闪烁像星星盛开 (uh-huh!)\n人人都醒来，把烦恼踹开 (get it!)\n热血沸腾，汗水自己安排\n\n[chorus]\n嘿，你还等啥？快抓住节拍 (come on!)\n光芒指引，让心都不存在 (whoa!)\n点燃热火，我们一起飙high (let’s go!)\n跳入午夜的狂欢时代\n\n[bridge]\n咚咚鼓声啊，让你的灵魂起飞 (woo!)\n手心拍一拍，能量翻倍 (ah-hah!)\n键盘响起来，如宇宙的交汇 (oh yeah!)\n就是这感觉，兄弟姐妹都陶醉\n\n[verse]\n灵魂从不睡，只想继续燃烧 (woo!)\n节奏像热浪，席卷这街道 (ow!)\n大伙儿涌上楼台，满面微笑 (yeah!)\n这一刻属于我们，无可替代\n\n[chorus]\n嘿，你还等啥？快抓住节拍 (come on!)\n光芒指引，让心都不存在 (whoa!)\n点燃热火，我们一起飙high (let’s go!)\n跳入午夜的狂欢时代\n\n[verse]\n世界多精彩，握紧把它打开 (alright!)\n每一步都像星球在摇摆 (uh-huh!)\n无边无际的律动像大海 (oo-yeah!)\n跟着光芒之舞，一起澎湃"
+    "caption": "A dark, atmospheric trap track driven by a deep 808 sub-bass line and crisp, rolling hi-hats from a drum machine. A male vocalist delivers assertive French rap verses with a confident flow over moody synth pads that create an urban, nocturnal soundscape. The chorus features layered vocals for emphasis, enhancing the hypnotic feel. The arrangement includes sections where vocal chops and ad-libs are used rhythmically alongside subtle sound effects like record scratches, culminating in a filtered outro that fades into silence before reprising the main hook one last time.",
+    "lyrics": "[Intro]\nAh\nLa rue est froide\nEh, eh, eh\n\n[Chorus]\nLa nuit tombe et je rentre, je fume sous la lune\nLes ombres m'appellent, je suis pas seul, je continue\nLes rêves sous la pluie, mon cœur sous la béton\nJe trace ma route même si la route est longue\n\n[Verse 1]\nLes murs ont des oreilles, les pavés des secrets\nUn faux sourire et t'es déjà trop proche de ta mère\nJ'ai des chaînes invisibles mais elles claquent plus fort\nChaque épreuve c'est une chance, faut qu'on en profite encore\nJ'entends les sirènes, mais je suis déjà en bas\nLe silence est dur mais ça me tient le bras\nTrop de choses à dire, je te jure c'est une blague\nJe sais que le monde tourne mais j'avance sans drague\nFaut que ça pète, pas de place pour les faux\nLa rue c'est un jeu, faut pas se prendre trop au sérieux\nLes rêves c'est du béton, pas des gratte-ciels\nJe fais mon truc, je suis un soldat, pas un pasteur\nLes vautours tournent, je suis dans mon délire\nChaque erreur me fait rire, chaque choix me fait frémir\nJ'ai le cœur glacé mais le mental est solide\nLa rue m'apprend qu'il faut jamais on la liquide\nJ'entends des voix mais elles ne disent rien\nLe silence est une arme, faut que je prenne un destin\nFaut que ça tourne, laisse-les rêver ou c'est du passé\nC'est la vie, faut jamais lâcher\n\n[Chorus]\nLa nuit tombe et je rentre, je fume sous la lune\nLes ombres m'appellent, je suis pas seul, je continue\nLes rêves sous la pluie, mon cœur sous la béton\nJe trace ma route même si la route est longue\nLa nuit tombe et je rentre, je fume sous la lune\nLes ombres m'appellent, je suis pas seul, je continue\nLes rêves sous la pluie, mon cœur sous la béton\nJe trace ma route même si la route est longue\n\n[Verse 2]\nTrop de gens cherchent une main, je trouve que des pièges\nChaque pas est un combat, faut que je lâche\nLe monde est lent mais les vrais sont lucides\nJ'ai gravé mon nom dans la peur et l'acide\nLa ville dort mais je garde mon miroir\nLes faux veulent ma peau mais le regard est noir\nLa rue c'est un livre, j'y crache ma libération\nJe suis là pour tout prendre, pas pour la direction\nLes ombres murmurent mais je suis déjà loin\nLe silence m'éteint, je garde la lumière au poignet\nLes regards sont froids, je vois des âmes en veille\nJe fais mon taf, je vois le ciel qui surveille\n\n[Chorus]\nLa nuit tombe et je rentre, je fume sous la lune\nLes ombres m'appellent, je suis pas seul, je continue\nLes rêves sous la pluie, mon cœur sous la béton\nJe trace ma route même si la route est longue\n\n[Outro]\nLa nuit tombe\nEh, eh, eh\nLa nuit tombe\nEh, eh, eh\nLa nuit tombe et je rentre, je fume sous la lune\nLes ombres m'appellent, je suis pas seul, je continue\nLes rêves sous la pluie, mon cœur sous la béton\nJe trace ma route même si la route est longue\n[Beat fades out]",
+    "bpm": 100,
+    "duration": 142,
+    "keyscale": "E minor",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_04.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "funk, pop, soul, rock, melodic, guitar, drums, bass, keyboard, percussion, 105 BPM, energetic, upbeat, groovy, vibrant, dynamic",
-    "lyrics": "[verse]\nNeon lights they flicker bright\nCity hums in dead of night\nRhythms pulse through concrete veins\nLost in echoes of refrains\n\n[verse]\nBassline groovin' in my chest\nHeartbeats match the city's zest\nElectric whispers fill the air\nSynthesized dreams everywhere\n\n[chorus]\nTurn it up and let it flow\nFeel the fire let it grow\nIn this rhythm we belong\nHear the night sing out our song\n\n[verse]\nGuitar strings they start to weep\nWake the soul from silent sleep\nEvery note a story told\nIn this night we’re bold and gold\n\n[bridge]\nVoices blend in harmony\nLost in pure cacophony\nTimeless echoes timeless cries\nSoulful shouts beneath the skies\n\n[verse]\nKeyboard dances on the keys\nMelodies on evening breeze\nCatch the tune and hold it tight\nIn this moment we take flight"
+    "caption": "A mid-tempo Mandopop ballad built on a steady electronic drum machine groove and a clean synth bassline. The arrangement features layered synthesizers providing chordal pads and melodic counterpoints, complemented by a clean electric guitar playing arpeggiated figures. The emotional male lead vocal, sung in Mandarin, soars through the verses and choruses, often reaching into a powerful falsetto. The track includes an expressive, melodic electric guitar solo and a dynamic bridge that builds tension before a final, passionate chorus filled with vocal ad-libs and a climactic guitar flourish.",
+    "lyrics": "[Intro]\nYeah\nOne\n\n[Verse  1]\n[zh] ye4 wan3 de5 feng1   qing1 qing1 chui1 guo4\n[zh] ni3 de5 xiao4 rong2   zai4 wo3 xin1 zhong1 shan3 shuo4\n[zh] yi1 ju4 hua4   shuo1 bu4 chu1 kou3 de5 zhi2 zhuo2\n[zh] xiang4 xing1 guang1 sa3 man3   wu2 jin4 de5 ye4 kong1\n\n\n[Chorus]\n[zh] ni3 shi4 wo3 xin1 di3   wu2 fa3 mo3 qu4 de5 meng4\n[zh] zai4 mei3 yi2 ge4 shun4 jian1   dou1 rang4 wo3 chen2 zui4 tong4\n[zh] ai4 zai4 xuan2 lv4 zhong1   wo3 de5 ling2 hun2 xiang1 tong1\n[zh] gen1 sui2 jie2 pai1   yi1 qie4 dou1 shi4 xin1 dong4\n\n\n[Verse  2]\n[zh] jie1 jiao3 de5 deng1 ying3   ying4 chu1 bi3 ci3 yan3 shen2\n[zh] xin1 tiao4 de5 jie2 zou4   bian4 de2 geng4 mi2 ren2\n[zh] yong1 bao4 de5 wen1 du4   rang4 shi4 jie4 bian4 qing1 chen2\n[zh] ai4 xiang4 yi1 shou3 wu2 fa3 ting2 zhi3 de5 qu1 ben3\n\n\n[Pre-Chorus]\n[zh] ni3 de5 sheng1 yin1   zai4 yin1 fu2 li3 piao1 you2\n[zh] mei3 yi2 ge4 shun4 jian1   dou1 rang4 wo3 xin1 gan3 shou4\n[zh] ji2 ta1 de5 sheng1 bo1   hua4 chu1 ni3 de5 xiao4 rong2\n[zh] he2 sheng1 qi3 he2 gu3 dian3 dou1 zai4 zhi4 re4 you2 dong4\n\n\n[Chorus]\n[zh] ni3 shi4 wo3 xin1 di3   wu2 fa3 mo3 qu4 de5 meng4\n[zh] zai4 mei3 yi2 ge4 shun4 jian1   dou1 rang4 wo3 chen2 zui4 tong4\n[zh] ai4 zai4 xuan2 lv4 zhong1   wo3 de5 ling2 hun2 xiang1 tong1\n[zh] gen1 sui2 jie2 pai1   yi1 qie4 dou1 shi4 xin1 dong4\n\n\n[Guitar Solo]\nOh\nYeah\n\n[Bridge]\n[zh] dang1 dian4 liu2 de5 shan3 shuo4   rang4 ye4 wan3 fei4 teng2\n[zh] wo3 men5 de5 xin1 ling2   ru2 gu3 sheng1 ban1 fan1 gun3\n[zh] jiu4 rang4 zhe4 yi1 ke4   sui2 yin1 yue4 geng4 zhen1\n[zh] ni3 shi4 wo3 de5 yin1 yue4   shi4 yong3 heng2 de5 wen1 hen2\n\n\n[Outro]\n[Vocal ad-libs]\nOh\nOh\nOh\nOh\n[Final guitar chord and fade out]",
+    "bpm": 100,
+    "duration": 235,
+    "keyscale": "G# minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_05.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "electronic rap",
-    "lyrics": "[verse]\nWaves on the bass, pulsing in the speakers,\nTurn the dial up, we chasing six-figure features,\nGrinding on the beats, codes in the creases,\nDigital hustler, midnight in sneakers.\n\n[chorus]\nElectro vibes, hearts beat with the hum,\nUrban legends ride, we ain't ever numb,\nCircuits sparking live, tapping on the drum,\nLiving on the edge, never succumb.\n\n[verse]\nSynthesizers blaze, city lights a glow,\nRhythm in the haze, moving with the flow,\nSwagger on stage, energy to blow,\nFrom the blocks to the booth, you already know.\n\n[bridge]\nNight's electric, streets full of dreams,\nBass hits collective, bursting at seams,\nHustle perspective, all in the schemes,\nRise and reflective, ain't no in-betweens.\n\n[verse]\nVibin' with the crew, sync in the wire,\nGot the dance moves, fire in the attire,\nRhythm and blues, soul's our supplier,\nRun the digital zoo, higher and higher.\n\n[chorus]\nElectro vibes, hearts beat with the hum,\nUrban legends ride, we ain't ever numb,\nCircuits sparking live, tapping on the drum,\nLiving on the edge, never succumb."
+    "caption": "An explosive, high-energy J-rock and pop-rock anthem driven by a powerful, punchy drum beat and a thick wall of distorted, chugging rhythm guitars. A passionate male lead vocal, delivered in a mix of Japanese and English, soars over the top with layered harmonies and echoed ad-libs that amplify the song's anthemic quality. The arrangement is dynamic, featuring a blistering, melodic guitar solo and a brief, atmospheric bridge with synth pads that provides a moment of contrast before slamming back into the high-octane chorus. The track concludes with a final, powerful guitar chord and a lingering sense of energy.",
+    "lyrics": "[Intro]\nOne more time, just one more time\nMou modorenai, mou modorenai\nThe GURU, my dream come true, my dream come true\n\n[Guitar Riff]\n\n[Verse 1]\nNani ga seigi ka, nani ga koukai ka\nHakase tsuzukeru, hashirinukeru\nNo life's a game, no rules to follow\nDakedo arata na chikara ga yomigaeru\n\n[Pre-Chorus]\nNo, this is our new life\nAisuru chikara, ima koso\nWe believe in ourselves\nMou atomodori dekinai\n\n[Chorus]\nEikou no hikari\nSono saki ni iru jiyuu ga aru\nOne more time, just one more time\nMou modorenai, mou modorenai\nThe GURU, my dream come true, my dream come true\n\n[Verse 2]\nKizutsuita tsubasa mo aruita michi mo\nOnaji hibi no naka de susumu\nNo stage to the hill, sore demo susumu\nMayowanai de susumu, susumu\n\n[Pre-Chorus]\nAisuru chikara, ima koso\nWe believe in ourselves\nMou atomodori dekinai\n\n[Chorus]\nEikou no hikari\nSono saki ni iru jiyuu ga aru\nOne more time, just one more time\nMou modorenai, mou modorenai\nThe GURU, my dream come true, my dream come true\n\n[Guitar Solo]\n\n[Bridge]\nDonna arashi ni matsu ka? I know\nNando mo yami ni shizumu ka?\nOre wa yuku, ore wa yuku\nChikara tsuyoku, yume wo motte\nChikara tsuyoku, yume wo motte\nTooi hikari, moeagaru\nKono kokoro ni hibiku takami\nWe are in the zone, we are in the zone\n\n[Breakdown]\nSono saki ni iru mirai ga aru\nOne more time, just one more time\nMou modorenai, mou modorenai\nThe GURU, my dream come true, my dream come true\n\n[Outro]\nKizutsuite mo\nMakenai yume wo\nNo stage to the hill, sore demo\nOre wa susumu, ore wa susumu\nMata tachiagaru, mata tachiagaru\n\n[Final Guitar Riff and Fade Out]",
+    "bpm": 100,
+    "duration": 200,
+    "keyscale": "E♭ minor",
+    "language": "ja",
+    "timesignature": "4"
 }

examples/text2music/example_06.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "electronic, house, electro house, synthesizer, drums, bass, percussion, fast, energetic, uplifting, exciting",
-    "lyrics": "[verse]\n霓虹灯下我们追逐\n人群跃动像潮水满布\n热浪袭来吹散孤独\n跳进节奏不如停下脚步\n\n[pre-chorus]\n脚尖触电快点感受\n迎着风声释放自由\n心跳节拍配合节奏\n一切烦恼请靠边游\n\n[chorus]\n夏夜狂奔没有尽头\n星光闪烁舞池不朽\n尽情挥洒所有节奏\n无边热情把你包裹哦\n\n[verse]\n天空翻滚黑云入夜\n每颗星星像音乐律贴\n耳边回响那低音线\n环绕耳际如梦境般甜\n\n[pre-chorus]\n脚尖触电快点感受\n迎着风声释放自由\n心跳节拍配合节奏\n一切烦恼请靠边游\n\n[chorus]\n夏夜狂奔没有尽头\n星光闪烁舞池不朽\n尽情挥洒所有节奏\n无边热情把你包裹哦"
+    "caption": "An energetic and uplifting J-pop track with strong anime theme song characteristics. The song opens with a bright piano melody and shimmering synth arpeggios before launching into a driving pop-rock arrangement. A clear, powerful male vocal leads the track, supported by a tight rhythm section of punchy drums and a melodic bassline. Clean and slightly overdriven electric guitars provide rhythmic chords and soaring lead lines, particularly during the anthemic chorus and a brief, expressive guitar solo. The structure includes a reflective bridge that momentarily softens the dynamics, followed by a powerful final chorus and an outro featuring melodic guitar licks and vocal ad-libs that fade to a clean finish.",
+    "lyrics": "[Intro - Synth Arpeggio & Guitar Riff]\n\n[Verse 1]\n空に浮かぶ雲の切れ間に\n願い事をそっと浮かべて\n今日がまた新しい始まり\n今この瞬間を生きていこう\n\n[Pre-Chorus]\nほら 聞こえるよ\n心の奥の声が\n何か始まる予感を抱いて\n今すぐに飛び出そう\n\n[Chorus]\n飛び出そう 明日へと\n風をまとって進むだけさ\nどんな日もが物語るから\n輝く未来を掴もう\n光の中へ踊ろう\n\n[Instrumental Break]\n\n[Verse 2]\n朝日が差し込む君の影\n迷わずに進んでいけばいい\n不安も希望も抱きしめて\n強く踏み出してゆくよ\n\n[Pre-Chorus]\nほら 聞こえるよ\n胸の奥の鼓動が\n夢の扉を開いてゆく\n輝きを捕まえる\n\n[Chorus]\n飛び出そう 明日へと\n風をまとって進むだけさ\nどんな日もが物語るから\n輝く未来を掴もう\n光の中へ踊ろう\n\n[Guitar Solo]\n\n[Bridge]\nもしも道に迷って立ち止まる日が来ても\nこの道の先は笑顔が待ってるから\n信じてみようよ\n未来は輝いてる\n一つの星に生まれ変わるんだ\n\n[Chorus]\n飛び出そう 明日へと\n風をまとって進むだけさ\nどんな日もが物語るから\n輝く未来を掴もう\n光の中へ踊ろう\n\n[Outro]\n音楽と共に飛ぶよ\n[Synth arpeggio fades out]",
+    "bpm": 100,
+    "duration": 205,
+    "keyscale": "F# major",
+    "language": "ja",
+    "timesignature": "2"
 }

examples/text2music/example_07.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "synth-pop, electronic, pop, synthesizer, drums, bass, piano, 128 BPM, energetic, uplifting, modern",
-    "lyrics": "[verse]\nWoke up in a city that's always alive\nNeon lights they shimmer they thrive\nElectric pulses beat they drive\nMy heart races just to survive\n\n[chorus]\nOh electric dreams they keep me high\nThrough the wires I soar and fly\nMidnight rhythms in the sky\nElectric dreams together we’ll defy\n\n[verse]\nLost in the labyrinth of screens\nVirtual love or so it seems\nIn the night the city gleams\nDigital faces haunted by memes\n\n[chorus]\nOh electric dreams they keep me high\nThrough the wires I soar and fly\nMidnight rhythms in the sky\nElectric dreams together we’ll defy\n\n[bridge]\nSilent whispers in my ear\nPixelated love serene and clear\nThrough the chaos find you near\nIn electric dreams no fear\n\n[verse]\nBound by circuits intertwined\nLove like ours is hard to find\nIn this world we’re truly blind\nBut electric dreams free the mind"
+    "caption": "An energetic electro-pop track driven by a squelchy, funky synth bassline and a punchy drum machine beat. The song opens with a distinctive, pitch-bending synth lead that serves as a recurring hook. The male vocals are delivered in a confident, rhythmic, almost rap-like cadence during the verses, transitioning to a more melodic and anthemic style in the chorus. The production is clean and modern, with layered synths and dynamic builds. A brief, atmospheric bridge offers a moment of contrast with sustained synth pads before the track slams back into a high-energy instrumental section and a final, powerful chorus.",
+    "lyrics": "[Intro - Synth Bass Riff]\n\n[Verse 1]\nChaque jour, c'est la retraite, c'est le plein de repas\nOn est déjà en vacances, on est là pour ça\nLes copains ont tout ce qu'on veut, on est prêts à bosser\nMais on revient à pied, on va tout explodir\nLes clients regardent, faut qu'on les contacte\nC'est pas une course, c'est un moment de spectacle\nOn attend, on sourit, on fait la loi\nMais on préfère l'action, pas l'effort du choix\n\n[Chorus]\nOn est là pour marquer, là pour tout exploser\nLe feu, la patate, faut vite tout chambouler\nMême en retard, faut bouger\nParce qu'on a plus peur d'exister\n\n[Instrumental Break - Synth Bass Riff]\n\n[Verse 2]\nPas besoin d'armure, faut du capital\nOn est les mêmes, on est dans l'agenda digital\nOn se frotte les mains, c'est l'énergie pure\nMême s'il y a bientôt dix jours à défendre sans murmure\nLes affiches tremblent, les commandes explosent\nL'agence, c'est notre trône, pas juste une cause\nOn fait le taf, on fait bouger le pays\nMais ensemble on a tout, c'est pour rassoumer ici\nLes copains, les meufs, c'est la règle de tout\nSi on part en morceaux, c'est qu'on y va jusqu'au bout\nLa patate en flammes, l'action dans la main\nOn est là pour marquer, faut pas lâcher notre refrain\n\n[Chorus]\nOn est là pour marquer, là pour tout exploser\nLe feu, la patate, faut vite tout chambouler\nMême en retard, faut bouger\nParce qu'on a plus peur d'exister\n\n[Bridge]\nLe stress monte, la galère s'installe\nOn est là pour gratter tout, peu importe le débat fatal\nOn rêve grand, on part à l'aventure\nLa patate au niveau haut, c'est notre vraie mur\nAlors levons nos verres, remplissons notre esprit\nLes coups de malaxe, on\n'est pas à se trahir\nC'est l'ambiance qu'on déclare, pas besoin d'armure\nOn va marquer l'équipe, on est là, on assure\n\n[Instrumental Break - Synth Bass Riff]\n\n[Outro]\nLe stress monte, la galère s'installe\nOn est là pour gratter tout, peu importe le débat fatal\nOn rêve grand, on part à l'aventure\nLa patate au niveau haut, c'est notre vraie mur\nAlors levons nos verres, remplissons notre esprit\nLes coups de malaxe, on\n'est pas à se trahir\nC'est l'ambiance qu'on déclare, pas besoin d'armure\nOn va marquer l'équipe, on est là, on assure\n[Synth bass riff fades out]",
+    "bpm": 100,
+    "duration": 180,
+    "keyscale": "F minor",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_08.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Cuban music, salsa, son, Afro-Cuban, traditional Cuban",
-    "lyrics": "[verse]\nSun dips low the night ignites\nBassline hums with gleaming lights\nElectric guitar singing tales so fine\nIn the rhythm we all intertwine\n\n[verse]\nDrums beat steady calling out\nPercussion guides no room for doubt\nElectric pulse through every vein\nDance away every ounce of pain\n\n[chorus]\nFeel the rhythm feel the flow\nLet the music take control\nBassline deep electric hum\nIn this night we're never numb\n\n[bridge]\nStars above they start to glow\nEchoes of the night's soft glow\nElectric strings weave through the air\nIn this moment none compare\n\n[verse]\nHeartbeats sync with every tone\nLost in music never alone\nElectric tales of love and peace\nIn this groove we find release\n\n[chorus]\nFeel the rhythm feel the flow\nLet the music take control\nBassline deep electric hum\nIn this night we're never numb"
+    "caption": "An upbeat and celebratory Brazilian samba-pop track driven by a vibrant rhythm section. A nimble nylon-string acoustic guitar lays down the chordal foundation, while a lively brass section, featuring trumpets and trombones, punctuates the arrangement with festive fanfares and melodic lines. The lead male vocal is clear and joyful, singing over a tight groove provided by a clean electric bass and a full drum kit. The song's structure is built for dancing, with an infectious chorus and an instrumental break showcasing a spirited brass and guitar interplay.",
+    "lyrics": "[Intro - Brass Section Melody]\n\n[Verse 1]\nOlha a morena dançando na rua\nVestido vermelho é pura magia\nParece rainha de chapéu de bua\nMeu coração bate na tua melodia\n\n[Chorus]\nParabéns, malagueira morena\nPra todos os cães do nosso Brasil\nCaminho no pôr do sol, pequena\nEntri no nosso show\nFeliz sinal\n\n[Instrumental Break - Brass Section]\n\n[Verse 2]\nMorena gira com sua piquenino\nE os olhos d'água refletindo no céu\nCabelo caído com o rostinho fino\nDesperta no ritmo do carrossel\n\n[Chorus]\nParabéns, malagueira morena\nPra todos os cães do nosso Brasil\nCaminho no pôr do sol, pequena\nEntri no nosso show\nFeliz sinal\n\n[Instrumental Break - Brass Section]\n\n[Bridge]\nLuzes estreladas iluminam a noite\nReflete nos becos, nas ruas, nos bares\nVem brincar no canto, sorrir no derroche\nEssa morena dança onde as estrelas caem\n\n[Chorus]\nParabéns, malagueira morena\nPra todos os cães do nosso Brasil\nCaminho no pôr do sol, pequena\nEntri no nosso show\nFeliz sinal\n\n[Outro - Brass Section Melody]\n[Final chord fades out]",
+    "bpm": 100,
+    "duration": 174,
+    "keyscale": "C major",
+    "language": "pt",
+    "timesignature": "4"
 }

examples/text2music/example_09.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "pop, piano, rap, dark, atmospheric",
-    "lyrics": "[verse]\n月光爬上窗 染白冷的床\n心跳的方向 带我入迷惘\n黑夜吞噬光 命运的纸张\n爱是血色霜 邪恶又芬芳\n\n[chorus]\n你是猎人的欲望 我是迷途的小羊\n深陷你眼眸的荒 唐突献出心脏\n我在夜里回荡 是谁给我希望\n黑暗风中飘荡 假装不再受伤\n\n[verse]\n心锁在门外 谁会解开关怀\n温柔的手拍 藏着冷酷杀害\n思绪如尘埃 撞击爱的霹雳\n灵魂的独白 为你沾满血迹\n\n[bridge]\n你是噩梦的歌唱 是灵魂的捆绑\n绝望中带着光 悬崖边的渴望\n心跳被你鼓掌 恶魔也痴痴想\n渐渐没了抵抗 古老诡计流淌\n\n[chorus]\n你是猎人的欲望 我是迷途的小羊\n深陷你眼眸的荒 唐突献出心脏\n我在夜里回荡 是谁给我希望\n黑暗风中飘荡 假装不再受伤\n\n[outro]\n爱如月黑无光 渗进梦的战场\n逃入无声的场 放手或心嚷嚷\n隐秘的极端 爱是极致风浪\n灵魂彻底交偿 你是终极虚妄"
+    "caption": "An explosive, high-energy J-pop track with strong hyperpop and chiptune influences. The song is driven by a relentless four-on-the-floor electronic drum beat and a punchy synth bassline. A high-pitched, digitally processed female vocal, reminiscent of Vocaloid, delivers rapid-fire melodies over layers of bright, arpeggiated synths and video game-style sound effects. The structure is dynamic, featuring intense verses and anthemic choruses that build into a frenetic instrumental break filled with complex synth runs and glitched-out fills. The track concludes with a dramatic tape-stop effect, deconstructing the beat into a wash of fading synth noise.",
+    "lyrics": "[Intro]\nKimi no koe ga kikoeru\nMotto motto tooku tooku\nAnata no ai wa eien ni\nSugiru omoi koto ni kaete\n\n[Instrumental Break]\n\n[Verse 1]\nYume no naka de kawasu kotoba\nTada no kakehiki\nKokoro no naka de narihibiku\nAnata e no ai koi suru\n\n[Chorus]\nAnata no ai\nItsumademo yume mitai\nKienai omoi\nItsumademo tsunagatteru\n\n[Post-Chorus]\nKimi to deatta ano hi\nSubete ga chigatta\nAnata no ai de tsunagatte\nOwaranai you ni kono mama\n\n[Instrumental Break]\n\n[Verse 2]\nKurikaesu hibi no naka de\nAi ga afuredasu\nAnata no nukumori o kanjiru\nYukuate no nai koi\n\n[Chorus]\nAnata no ai\nItsumademo yume mitai\nKienai omoi\nItsumademo tsunagatteru\n\n[Instrumental Solo]\n\n[Bridge]\nKurikaesu hibi no naka de\nAi ga afuredasu\nAnata no nukumori o kanjiru\nYukuate no nai koi\n\n[Chorus]\nAnata no ai\nItsumademo yume mitai\nKienai omoi\nItsumademo tsunagatteru\n\n[Outro]\n[Instrumental Breakdown]\n[Synth arpeggio fades out]",
+    "bpm": 100,
+    "duration": 167,
+    "keyscale": "E♭ minor",
+    "language": "ja",
+    "timesignature": "2"
 }

examples/text2music/example_10.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "surf music",
-    "lyrics": "[verse]\nSunshine on the boulevard the beach is calling loud\nWaves are dancing golden sand under a cotton cloud\nElectric heartbeat pounding fast the tide is on our side\nCatch a wave and feel alive we’ll take it for a ride\n\n[verse]\nPalm trees swaying left to right they know where we belong\nFeel the rhythm of the night it keeps us moving strong\nSea spray kisses salty air we’re flying with the breeze\nChampagne states of mind we ride we do just as we please\n\n[chorus]\nWe’re riding waves of life together hand in hand\nWith every beat we chase the beat it’s our own wonderland\nFeel the music take you higher as the shorelines blur\nThis is our world our endless summer as we live and learn\n\n[bridge]\nMoonlight paints the ocean blue reflections in our eyes\nStars align to light our path we’re surfing through the skies\nEvery moment like a song we sing it loud and clear\nEvery day’s a new adventure with you always near\n\n[verse]\nNeon lights and city sounds they blend with ocean views\nWe’re unstoppable tonight no way that we can lose\nDreams are written in the sand they sparkle in the sun\nTogether we’re a masterpiece our story’s just begun\n\n[chorus]\nWe’re riding waves of life together hand in hand\nWith every beat we chase the beat it’s our own wonderland\nFeel the music take you higher as the shorelines blur\nThis is our world our endless summer as we live and learn"
+    "caption": "A smooth, contemporary R&B track built on a foundation of clean, chorused electric guitar chords and a steady, unobtrusive drum machine beat. A round synth bass provides a warm low end. The lead male vocal is clear and emotive, delivered in a smooth tenor with occasional falsetto ad-libs and layered harmonies that swell during the anthemic chorus. The arrangement features a brief, intimate bridge with whispered vocals before transitioning into a melodic guitar-led instrumental break. The track concludes with a layered vocal outro, combining sung melodies with whispered phrases that fade to a clean finish.",
+    "lyrics": "[Intro]\nYeah\nOh\nYeah\n\n[Verse 1]\nMidnight whispers in my ear\nStars above, they're shining clear\nCity lights, they start to fade\nUnderneath, our plans we've made\nMoonlit shadows twist and turn\nHearts on fire, passions burn\nLaughter echoes through the night\nHolding hands, we take flight\n\n[Chorus]\nYou're my heartbeat symphony\nEchoes of sweet destiny\nEvery beat a symphony\nYou and me, we're wild and free\n\n[Verse 2]\nNeon dreams and city haze\nLost in love's enchanting maze\nWhisper secrets in the breeze\nTime just stops as we appease\n\n[Chorus]\nYou're my heartbeat symphony\nEchoes of sweet destiny\nEvery beat a symphony\nYou and me, we're wild and free\n\n[Bridge]\nDance through stars and find our way\nIn your arms, I'll always stay\nLove so strong, it never bends\nTogether 'til the journey ends\n\n[Instrumental Break]\n\n[Outro]\nDance through stars and find our way\nIn your arms, I'll always stay\nLove so strong, it never bends\nTogether 'til the journey ends\nYou're my heartbeat symphony\nEchoes of sweet destiny\nEvery beat a symphony\nYou and me, we're wild and free\nDance through stars and find our way\nIn your arms, I'll always stay\nLove so strong, it never bends\nTogether 'til the journey ends",
+    "bpm": 100,
+    "duration": 228,
+    "keyscale": "A major",
+    "language": "en",
+    "timesignature": "4"
 }

examples/text2music/example_11.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "alternative rock, pop, rock",
-    "lyrics": "[verse]\nBright lights flashing in the city sky\nRunning fast and we don't know why\nElectric nights got our hearts on fire\nChasing dreams we'll never tire\n\n[verse]\nGrit in our eyes wind in our hair\nBreaking rules we don't even care\nShouting loud above the crowd\nLiving life like we're unbowed\n\n[chorus]\nRunning wild in the night so free\nFeel the beat pumping endlessly\nHearts collide in the midnight air\nWe belong we don't have a care\n\n[verse]\nPiercing through like a lightning strike\nEvery moment feels like a hike\nDaring bold never backing down\nKings and queens without a crown\n\n[chorus]\nRunning wild in the night so free\nFeel the beat pumping endlessly\nHearts collide in the midnight air\nWe belong we don't have a care\n\n[bridge]\nClose your eyes let your spirit soar\nWe are the ones who wanted more\nBreaking chains of the mundane\nIn this world we'll make our claim"
+    "caption": "An explosive, high-energy K-pop and EDM track driven by a relentless four-on-the-floor beat and a pulsing synth bassline. The arrangement opens with a bright piano melody and shimmering arpeggiated synths before slamming into the main groove. A powerful, clear male lead vocal delivers an anthemic melody, punctuated by energetic ad-libs and hype-man shouts. The production is dense with layered synthesizers, including soaring leads, atmospheric pads, and dynamic FX like risers and sweeps that build tension into the explosive choruses. A mid-song instrumental break features a melodic synth lead and vocal chops, leading into a final, climactic chorus and an abrupt ending with a tape-stop sound effect.",
+    "lyrics": "[Intro]\n[Synth arpeggio and pads]\n\n[Verse  1]\n[ko] hwangholhan i jilseo-e neon ppajyeo\n\n[en] (oh yeah)\n\n[ko] lideum wi-e sinhoneul jilleo\n\n[en] (oh, let's go)\n\n[ko] eodum sog-e bich-eul kkeul-eo\n[ko] modeun geos-i bultago\n\n\n[Pre-Chorus]\n[ko] bich-i nal mag-eulyeo\n[ko] moduga geudaelo bichna\n[ko] geo-ul sog-e bichwojwo\n[ko] naneun jasin-ui jibjunghae\n\n\n[Chorus]\n[ko] oelo-un i jilseo gip-eojyeo\n\n[en] (yeah, yeah)\n\n[ko] lideum wi-e him-eul jwi-yeo\n\n[en] (let's go)\n\n[ko] saelo-un sidaeleul yeol-eo\n\n[en] (oh yeah)\n\n[ko] gaseumsog-e bulkkoch-i pi-eo\n\n\n[Verse  2]\n[ko] geodaehan mudae sog-eseo\n[ko] nan tae-eonal geo-ya\n[ko] modeun geos-i choego-ui jilseo\n[ko] uli hamkke chumchwo\n\n\n[Bridge]\n[ko] bichgwa eodum-i salajigo\n[ko] ulin seololeul bichwo\n[ko] eunmilhan i sungan sog-eseo\n[ko] ulin hamkke mandeul-eoga\n\n\n[Chorus]\n[ko] oelo-un i jilseo gip-eojyeo\n\n[en] (yeah, yeah)\n\n[ko] lideum wi-e him-eul jwi-yeo\n\n[en] (let's go)\n\n[ko] saelo-un sidaeleul yeol-eo\n\n[en] (oh yeah)\n\n[ko] gaseumsog-e bulkkoch-i pi-eo\n\n\n[Instrumental Break]\n[Synth lead melody with vocal chops]\n\n[Outro]\n[Synth arpeggio fades out]\n[abrupt silence]",
+    "bpm": 40,
+    "duration": 210,
+    "keyscale": "E♭ minor",
+    "language": "ko",
+    "timesignature": "4"
 }

examples/text2music/example_12.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "rock, hip - hop, orchestral, bass, drums, electric guitar, piano, synthesizer, violin, viola, cello, fast, energetic, motivational, inspirational, empowering",
-    "lyrics": "### **[Intro – Spoken]**  \n*\"The streets whisper, their echoes never fade.  \nEvery step I take leaves a mark—this ain't just a game.\"*  \n\n### **[Hook/Chorus]**  \nBorn in the chaos, I weather the storm,  \nRising from ashes where warriors are born.  \nChains couldn't hold me, the system’s a maze,  \nI rewrite the rules, set the city ablaze!  \n\n### **[Verse 1]**  \nCold nights, empty pockets, dreams laced with fight,  \nEvery loss made me sharper, cut deep like a knife.  \nThey said I wouldn’t make it, now they watch in despair,  \nFrom the curb to the throne, took the pain, made it rare.  \nEvery siren’s a melody, every alley holds a tale,  \nRose from the shadows, left my name on the trail.  \nStreetlights flicker like warnings in the haze,  \nBut I move like a phantom, unfazed by the blaze.  \n\n### **[Hook/Chorus]**  \nBorn in the chaos, I weather the storm,  \nRising from ashes where warriors are born.  \nChains couldn't hold me, the system’s a maze,  \nI rewrite the rules, set the city ablaze!  \n\n### **[Verse 2]**  \nBarbed wire fences couldn't lock in my mind,  \nEvery cage they designed, I left broken behind.  \nThey want control, but I’m destined to roam,  \nWhere the lost find their voice, where the heart sets the tone.  \nSteel and concrete, where the lessons run deep,  \nEvery crack in the pavement tells a story of heat.  \nBut I rise, undefeated, like a king with no throne,  \nWriting scripts in the struggle, my legacy’s stone.  \n\n### **[Bridge]**  \nFeel the rhythm of the underground roar,  \nEvery wound tells a story of the battles before.  \nBlood, sweat, and echoes fill the cold midnight,  \nBut we move with the fire—unshaken, upright.  \n\n### **[Verse 3]**  \nNo regrets, no retreat, this game has no pause,  \nEvery step that I take is a win for the lost.  \nI took lessons from hustlers, wisdom from pain,  \nNow the echoes of struggle carve power in my name.  \nThey built walls, but I walk through the cracks,  \nTurned dirt into gold, never looked back.  \nThrough the struggle we rise, through the fire we claim,  \nThis is more than just music—it's life in the frame.  \n\n### **[Hook/Chorus – Reprise]**  \nBorn in the chaos, I weather the storm,  \nRising from ashes where warriors are born.  \nChains couldn't hold me, the system’s a maze,  \nI rewrite the rules, set the city ablaze!  \n\n### **[Outro – Spoken]**  \n*\"The scars, the struggle, the grind—it’s all part of the rhythm.  \nWe never break, we never fold. We rise.\"*"
+    "caption": "An energetic and anthemic pop-rock track driven by a clean, funky electric guitar riff and a punchy, tight drum and bass groove. The song opens with a catchy guitar hook before settling into a verse led by a clear, powerful male vocal. The arrangement builds dynamically into a soaring, uplifting chorus with layered vocals and driving instrumentation. A melodic guitar solo section features expressive bends and fills. The track includes a brief, atmospheric bridge with wordless vocalizations before launching into a final, powerful chorus and concluding with the initial guitar riff.",
+    "lyrics": "[Intro - Guitar Riff]\n\n[Verse 1]\n翻过山又像迷雾\n心也照亮每一步\n大地在脚下伸出手\n追逐梦想不要停留\n\n[Pre-Chorus]\n星光摇曳在夜幕\n未来就在眼前呼\n跳动脉搏像闪电火\n展开翅膀冲破束缚\n\n[Chorus]\n在梦的边缘跳跃光辉\n无数的光点汇聚成美\n燃烧的希望把世界敲碎\n一切再次新的节奏到位\n\n[Guitar Interlude]\n\n[Verse 2]\n音符交错夜色深处\n心的律动难以抗拒\n脚步合拍如同律步\n舞出一曲全新的篇幅\n\n[Pre-Chorus]\n银河展开狂奔翅膀\n你和我合奏不息光\n跨越万千热浪激荡\n前行不停冲破过往\n\n[Chorus]\n在梦的边缘跳跃光辉\n无数的光点汇聚成美\n燃烧的希望把世界敲碎\n一切再次新的节奏到位\n\n[Guitar Solo]\n\n[Bridge - Vocal Harmonies]\n(Hmm-mm-mm-mm)\n(Hmm-mm-mm-mm)\n(Hmm-mm-mm-mm)\n(Hmm-mm-mm-mm)\n\n[Chorus]\n在梦的边缘跳跃光辉\n无数的光点汇聚成美\n燃烧的希望把世界敲碎\n一切再次新的节奏到位\n\n[Outro - Guitar Riff]",
+    "bpm": 100,
+    "duration": 190,
+    "keyscale": "C# minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_13.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "tango finlandés, campanas, disco, dark pop, electro, guitarra clásica, corridos tumba",
-    "lyrics": "[inst]"
+    "caption": "A classic boom-bap hip-hop track built on a nostalgic foundation. A steady drum machine groove with a crisp snare drives the song forward beneath atmospheric synth pads that create an introspective mood. The male rapper delivers confident verses with a clear flow, while the chorus shifts to melodic, layered vocal harmonies soaked in reverb for an anthemic feel. The arrangement is punctuated by spoken ad-libs treated with echo effects and concludes with an extended outro featuring heavily filtered vocals over a stripped-back beat before fading into ambient textures.",
+    "lyrics": "[Intro]\nYo\nIt's Laðum\nYou already know\nMoney talks\nThey don't talk back\n\n[Verse 1]\nI stay with the heat, streets raised me\nNever folded under pressure, now I'm way crazy\nChains on my neck, still my grind daily\nBut my soul stay heavy, yeah the grind ain't lazy\nOld heads stackin', but I flip through the beat\nStackin' chips on chips, now they movin' in the street\nBut the streets ain't ready, now my path got heat\nUsed to dream to win, now I move with the beat\nCopped a cold brick, got heat in my heart\nNow my chain swing tight, I been playin' my part\nMoney ain't a thing, but it takes its part\nStill I carry my soul like it's built from the start\n\n[Chorus]\nSo tell me, why they hate when you live?\nGot the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\nWhy they hate when you live?\nWith the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\nWhy they hate when you live?\nWith the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\n\n[Verse 2]\nThey talk about the grind but I'm broke, I'm grindin'\nEvery dollar spent, now the people still whinin'\nBut I keep my head high, never fold, never lyin'\nAnd my heart's still cold even when the day's lyin'\nOne road left but I can't ride that chain\nI'ma keep it movin', yeah they all feel my pain\nThey say money talks, but I'm answerin' it's still\nSo I made it in silence, now I'm free from the deal\n\n[Chorus]\nTell me, why they hate when you live?\nGot the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\nWhy they hate when you live?\nWith the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\nWhy they hate when you live?\nWith the heart of a king, but the soul in the bills\nTell me, who the heck do they feel?\nWhen the world keep spinnin' but you lost it all still\n(Still, still, still)\n\n[Outro]\nThey can't cage the vibe, yeah\nYou gotta stay in\nIt's money and power, a universe, one rule\nMoney is the money\nAnd I'm not at the bottom\n[Instrumental fades out]",
+    "bpm": 100,
+    "duration": 210,
+    "keyscale": "F# major",
+    "language": "en",
+    "timesignature": "4"
 }

examples/text2music/example_14.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Nightclubs, dance parties, workout playlists, radio broadcasts",
-    "lyrics": "Burning in motion, set me alight!\nEvery heartbeat turns into a fight!\nCaged in rhythm, chained in time!\nLove’s a battle— You're Mine!  You're Mine!"
+    "caption": "An aggressive, high-energy electronic track driven by a relentless four-on-the-floor kick drum and a gritty, distorted synth bassline. The arrangement opens with a chiptune-style arpeggiated synth melody before slamming into the main groove. A forceful male vocal, delivered in a rhythmic, almost-rapped style, cuts through the dense mix of sharp synth stabs and atmospheric pads. The chorus is anthemic and powerful, with layered vocals creating a sense of defiance. The track features dynamic shifts, including a breakdown with filtered vocals and a build-up that reintroduces the intense beat, culminating in a powerful outro with vocal chops and a final, abrupt stop.",
+    "lyrics": "[Intro]\n[Synth arpeggio intro]\n[Drum build-up]\n\n[Verse 1]\nEfo, siren klocha' mi kocha\nDu mevi shemol al kol ha'bachar me'ha'ba\nTachat, kai la'pata\nNima bila mefila\n\n[Chorus]\nTishuvot lefater bach'ka\nBatecha nolecha im telev elor lachta\nTirot lanu, venashmeach\nKa'otzot pachta\n\n[Instrumental Drop]\n\n[Verse 2]\nHakavat et hasnan\nHakol bishvili al kol hachol rotza\nHesham tamash, tsurad lo\nHesham rishon, rosh, orad\n\n[Chorus]\nTishuvot lefater bach'ka\nBatecha nolecha im telev elor lachta\nTirot lanu, venashmeach\nKa'otzot pachta\n\n[Instrumental Drop]\n\n[Bridge]\nTishuvot, hetta, bach'ka\nBatecha nolecha im telev elor lachta\nTirot lanu, venashmeach\nKa'otzot pachta\n\n[Breakdown]\n[Filtered synth arpeggio]\n(Tishuvot lefater bach'ka) [pitched down, reverbed]\n\n[Chorus]\nTishuvot lefater bach'ka\nBatecha nolecha im telev elor lachta\nTirot lanu, venashmeach\nKa'otzot pachta\n\n[Outro]\n[Synth arpeggio fades out]",
+    "bpm": 100,
+    "duration": 254,
+    "keyscale": "E minor",
+    "language": "he",
+    "timesignature": "4"
 }

examples/text2music/example_15.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "melancholic, world, sad, medieval, soulful",
-    "lyrics": "[Verse]\nIn a world so grand he roams the skies alone\nHis heart a heavy stone a tale untold\nWhispers of his past echo through the night\nA lonely dragon searching for the light\n\n[Verse 2]\nOnce a mighty force now he drifts in pain\nHis scales once shimmered now they're dark with shame\nCast out by his kin in shadows he does hide\nA haunting sorrow burns deep inside\n\n[Chorus]\nRoaming endless fields with no friend in sight\nHis roar a mournful cry beneath the moon's pale light\nTears fall like stars as he flies on his way\nA lonely dragon yearning for the break of day\n\n[Bridge]\nThe world turns cold the nights grow long\nIn his heart he carries an ancient song\nOf battles fought and love long gone\nA legend now but his soul is torn\n\n[Verse 3]\nHoping for a day he'll find a kindred soul\nTo share his pain and make him whole\nTill then he drifts a shadow in the sky\nA lonely dragon with tears in his eye\n\n[Chorus]\nRoaming endless fields with no friend in sight\nHis roar a mournful cry beneath the moon's pale light\nTears fall like stars as he flies on his way\nA lonely dragon yearning for the break of day"
+    "caption": "An intense, high-energy electronic track driven by a relentless, arpeggiated synth melody and a hard-hitting trap beat with rapid-fire hi-hats and a deep sub-bass. A powerful female vocal, delivered in Mandarin, cuts through the mix with a clear, assertive tone. The arrangement builds through tense verses into an anthemic, soaring chorus where the vocals become more melodic and layered. The track features dynamic drops and builds, including a brief, atmospheric bridge with filtered vocals and pads before slamming back into a final, climactic chorus and an instrumental outro that deconstructs the main synth motif.",
+    "lyrics": "[Intro - Arpeggiated Synth Melody]\n\n[Verse 1]\n午夜城市灯光闪\n黑夜中你藏心弦\n一脚踏入深渊洞\n脚下跳动电波涌\n喧嚣夜深云雾涌\n孤单的你在夜空\n光与影交织梦境\n每一拍都在追忆\n\n[Chorus]\nYeah\n黑夜之上我是逃亡\n风暴之中寻找方向\n燃烧灵魂去疯狂\n这世界锁不住渴望\n\n[Instrumental Break - Synth Melody]\n\n[Verse 2]\n时间律动催人转\n繁星作画流星卷\n故事未完灵魂闪\n快将心跳连接遍\n\n[Instrumental Drop]\n\n[Bridge]\n冲破那无边黑暗\n光束驱散悲伤酸\n虚伪遮眼无需问\n黑夜是我唯一的吻\n\n[Chorus]\n黑夜之上我是逃亡\n风暴之中寻找方向\n燃烧灵魂去疯狂\n这世界锁不住渴望\n\n[Instrumental Break]\n\n[Chorus - Layered Vocals]\n冲破那无边黑暗\n光束驱散悲伤酸\n虚伪遮眼无需问\n黑夜是我唯一的吻\n\n[Outro - Synth Melody Fades Out]",
+    "bpm": 100,
+    "duration": 194,
+    "keyscale": "E minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_16.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "anime, cute female vocals, kawaii pop, j-pop, childish, piano, guitar, synthesizer, fast, happy, cheerful, lighthearted",
-    "lyrics": "[Chorus]\nねぇ、顔が赤いよ？\nどうしたの？ 熱があるの？\nそれとも怒ってるの？\nねぇ、言ってよ！\n\nどうしてそんな目で見るの？\n私、悪いことした？\n何か間違えたの？\nお願い、やめて… 怖いから…\nだから、やめてよ…\n\n[Bridge]\n目を閉じて、くるっと背を向けて、\n何も見なかったフリするから、\n怒らないで… 許してよ…\n\n[Chorus]\nねぇ、顔が赤いよ？\nどうしたの？ 熱があるの？\nそれとも怒ってるの？\nねぇ、言ってよ！\n\nどうしてそんな目で見るの？\n私、悪いことした？\n何か間違えたの？\nお願い、やめて… 怖いから…\nだから、やめてよ…\n\n[Bridge 2]\n待って、もし私が悪いなら、\nごめんなさいって言うから、\nアイスクリームあげるから、\nもう怒らないで？\n\nOoooh… 言ってよ！"
+    "caption": "An upbeat, retro-flavored synth-pop track with a distinct 80s city pop aesthetic. The song is built on a foundation of a punchy, gated-reverb drum machine beat and a funky, melodic synth bassline that drives the groove. Layers of bright, brassy synth stabs and shimmering pads create a vibrant, nostalgic soundscape. A clean, funky electric guitar adds rhythmic texture with muted strums and short licks. The lead vocal is a clear, smooth female voice, delivering a catchy melody with a touch of reverb. The track follows a classic pop structure, with an energetic chorus and a brief, more reflective bridge before a final instrumental fade-out.",
+    "lyrics": "[Intro - Synth & Drum Machine]\n\n[Verse 1]\n月亮挂在夜空轻轻摇\n城市霓虹在眼前飘\n你的脚步像慢动作漂\n心跳加速无法逃跑\n\n[Chorus]\n跟着节奏慢慢靠近\n梦里的旋律吸引不停\n你的眼神如星光的印\n我的世界因你而安静\n\n[Instrumental Break]\n\n[Verse 2]\n街头热气弥漫着热浪\n脚步和节拍越来越长\n空气中弥漫淡淡芬芳\n我们的距离瞬间在曝光\n\n[Chorus]\n跟着节奏慢慢靠近\n梦里的旋律吸引不停\n你的眼神如星光的印\n我的世界因你而安静\n\n[Bridge]\n午夜钟声在耳边回荡\n时间却静止无法隐藏\n你的轮廓在脑海点亮\n每一瞬间都让我疯狂\n\n[Instrumental Break]\n\n[Chorus]\n跟着节奏慢慢靠近\n梦里的旋律吸引不停\n你的眼神如星光的印\n我的世界因你而安静\n\n[Outro - Instrumental Fade Out]",
+    "bpm": 100,
+    "duration": 187,
+    "keyscale": "E minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_17.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "dark, death rock, metal, hardcore, electric guitar, powerful, bass, drums, 110 bpm, G major",
-    "lyrics": "[Verse]\nMy lovers betray me\nThe snake in my garden is hissing\nIn the air is the sweetness of roses\nAnd under my skin\nThere's a thorn\n\n[Verse 2]\nI should have known\nThat God sends his angel in shadows\nWith blood in his veins\nI watch the enemy\nGivin' me the hand of my savior\n\n[Chorus]\nAnd I can't love again\nWith the echo of your name in my head\nWith the demons in my bed\nWith the memories\nYour ghost\nI see it\n'Cause it comes to haunt me\nJust to taunt me\nIt comes to haunt me\nJust to taunt me\n\n[Verse 3]\nWith sugar and spice\nIt's hard to ignore the nostalgia\nWith the men on their knees\nAt the gates of my heart\nHow they beg me\n\n[Verse 4]\nThey say\n\"No one will ever love you\nThe way that I do\nNo one will ever touch you\nThe way that I do\"\n\n[Chorus]\nAnd I can't love again\nWith the echo of your name in my head\nWith the demons in my bed\nWith the memories\nYour ghost\nI see it\n'Cause it comes to haunt me\nJust to taunt me\nIt comes to haunt me\nJust to taunt me"
+    "caption": "A melancholic piano melody opens over atmospheric pads, setting an introspective tone before a crisp lo-fi hip-hop beat enters with a deep sub-bass foundation. The lead male vocal is smooth and emotive, delivered with significant reverb that enhances its spacious quality. The arrangement builds dynamically into a powerful chorus where layered vocals soar over swelling synth strings and intensified drums. A brief bridge offers a moment of reflection with more sparse instrumentation before launching back into a climactic final chorus featuring passionate ad-libs and harmonies. The track concludes with fading ambient textures and lingering piano notes.",
+    "lyrics": "[Intro: Piano Arpeggio]\n\n[Verse 1]\nFootsteps echo in the quiet night\nA lamp flicker, casting cold light\nShadows stretch long on the pavement below\nWhat they say, I don't know\n\n[Pre-Chorus]\nWhispers in the misty air\nA promise made, left somewhere\nWhoa\n\n[Chorus]\nI've got dreams that fade, but they stay\nLost in the glow of yesterday\nWhoa\nI'm chasing shadows in the rain\nBut every step feels the same\n\n[Verse 2]\nNeon signs hum their lonesome tune\nThe sky painted dark, a fading moon\nWindows fogged with stories untold\nThis city's heart is icy cold\n(Ooooh)\n\n[Pre-Chorus]\nEvery corner hides a piece of me\nA ghost of who I used to be\nOh\n\n[Chorus]\nI've got dreams that fade, but they stay\nLost in the glow of yesterday\nWhoa\nI'm chasing shadows in the rain\nBut every step feels the same\n\n[Outro]\n[Piano arpeggio fades out]",
+    "bpm": 100,
+    "duration": 185,
+    "keyscale": "B♭ minor",
+    "language": "en",
+    "timesignature": "4"
 }

examples/text2music/example_18.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "aggressive, Heavy Riffs, Blast Beats, Satanic Black Metal",
-    "lyrics": "[verse]\nFloating through the galaxy on a midnight ride\nStars are dancing all around in cosmic tides\nFeel the pulse of space and time beneath our feet\nEvery beat a heartbeat in this endless suite\n\n[chorus]\nGalactic dreams under neon lights\nSailing through the velvet nights\nWe are echoes in a cosmic sea\nIn a universe where we are free\n\n[verse]\nPlanetary whispers in the sky tonight\nEvery constellation's got a secret sight\nDistant worlds and moons we have yet to see\nIn the void of space where we can just be\n\n[bridge]\nAsteroids and comets in a ballet they spin\nLost in the rhythm of where our dreams begin\nClose your eyes and let the synths take flight\nWe're voyagers on an electric night\n\n[verse]\nLet the piano keys unlock the stars above\nEvery chord a memory every note is love\nIn this synth symphony we find our grace\nDrifting forever in this boundless space\n\n[chorus]\nGalactic dreams under neon lights\nSailing through the velvet nights\nWe are echoes in a cosmic sea\nIn a universe where we are free"
+    "caption": "A clean, melodic electric guitar riff with a touch of chorus effect opens the track, establishing a catchy, looping motif. A male vocalist enters with a confident, rhythmic rap flow over a steady, mid-tempo hip-hop beat and a smooth bassline. The chorus elevates the energy with a more powerful, sung vocal delivery, punctuated by echoed, soaring ad-libs. The arrangement is built around the interplay between the crisp guitar and the dynamic vocal performance, which shifts between rapping and melodic singing. The track concludes with the initial guitar riff returning, accompanied by a whispered vocal, before fading to silence.",
+    "lyrics": "[Intro: Electric Guitar Melody]\n\n[Verse 1]\nJCC的爱\n是永不凋零的rose\nJCC的爱\n是永不凋零的rose\nJCC的爱\n是永不凋零的rose\nJCC的爱\n是永不凋零的rose\n\n[Pre-Chorus]\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\n\n[Chorus]\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\n\n[Verse 2]\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\n\n[Chorus]\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\nJCC的爱是永不凋零的rose\n\n[Outro: Electric Guitar Melody]\nJCC的rose\nJCC的rose",
+    "bpm": 100,
+    "duration": 175,
+    "keyscale": "E minor",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_19.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "r&b, soul, funk/soul",
-    "lyrics": "[verse]\nDancing through electric fires\nHeart is buzzing like live wires\nIn your arms I find desire\nFeel the beat as we get higher\n\n[chorus]\nElectric love in the night sky\nWe’re gonna soar baby you and I\nDrop the bass let the rhythm fly\nFeel the heat and don't ask why\n\n[verse]\nWhisper secrets that make me blush\nUnder the neon city hush\nYour touch gives me such a rush\nTurn it up we're feeling lush\n\n[chorus]\nElectric love in the night sky\nWe’re gonna soar baby you and I\nDrop the bass let the rhythm fly\nFeel the heat and don't ask why\n\n[bridge]\nThrough the lights and the smoky haze\nI see you in a thousand ways\nLove's a script and we’re the play\nTurn the page stay till we sway\n\n[chorus]\nElectric love in the night sky\nWe’re gonna soar baby you and I\nDrop the bass let the rhythm fly\nFeel the heat and don't ask why"
+    "caption": "A classic boom-bap hip-hop track built on a steady drum machine groove and a prominent, funky bassline. A clean electric guitar plays a recurring melodic riff and chords, giving the song a slightly jazzy, contemplative feel. The lead vocal is a male rapper delivering a clear, narrative-driven story in French. The choruses introduce layered, harmonized male vocals that add a melodic and slightly melancholic dimension. The track concludes with a spoken-word outro over the fading instrumental, reinforcing the song's humorous and relatable theme.",
+    "lyrics": "[Intro]\nYo, écoute l'histoire de Cécile 8\nLa femme du crocodile\nElle achète pas du pain mais du vernis sans malice\nQuand elle passe, ça devient trop pire, trop radical\nElle préfère des roupettes, c'est pas méchant, c'est capital\n\n[Verse 1]\nSa femme se fait des coups par trois équipes sévères\nMais quand elle sort, c'est pas des coups mais des bières\nElle a montré un CV, pas un collègue à l'école\nElle a frappé le premier, dit que c'était du mollo sans bol\n\n[Chorus]\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\n\n[Verse 2]\nUn soir elle s'est fait pousser, trois pierres de regards\nElle l'a pris pour une coquée, pas le genre de fille qui s'égare\nMais dans la cour, elle court, se balade sur le trottoir\nElle traque ses fichiers bidons comme un dandy sans espoir\nElle a vu l'argent du renouveau, pas la galère\nMais dans la cour elle plane comme un résidu dans la mer\nCertains s'excusent du bien avec leurs valises\nMais moi ce que je veux, c'est des billets frits en vitrine\n\n[Chorus]\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\n\n[Bridge]\nL'argent monte à l'eau mais c'est du cocaïne\nElle veut des femmes en trop mais c'est pas l'amour qui gagne\nLes romans partent plus tard, elle compte des milliers\nUn service de boulot, c'est pas ce qu'elle veut pour les oubliés\nLe seul travail qui paye, c'est la richesse du décor\nLe seul job qu'elle trouve, c'est lui même en cas de mort\nElle arrive à la porte, amour que reine sa libido\nCe mec, t'as capté ? C'est du vin dans ton vino, mon amigo\n\n[Verse 3]\nSon bar tabou, Sécile 8 au premier virage\nElle s'installe dans le théâtre, love dans le carnage\nMais la seule vérité, c'est qu'elle fait pas le bien\nElle veut voler, elle veut ça, sans jamais aller loin\n\n[Chorus]\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\nCécile 8 est ma muse, elle a le front censé clair\nMais derrière ses lunettes, y'a que des bonnes affaires\n\n[Outro]\nYo, et dans le chill, y'a des flammes à manger\nMais Cécile 8, elle peut pas les oublier\nElle rêve de biff ou d'un voyage loin d'ici\nUn jour, je dirai ce que j'ai déjà pris\nC'est pas l'amour, c'est un carnaval\nUn jour, elle verra, je suis pas ton carnaval\nCecile 8, la fille du coin dans la vie\nTu dois courir après l'argent ou mourir pour la justice\nCécile 8, la femme du crocodile\nElle aime, elle craint, mais elle veut son style\nCécile 8, la femme du crocodile\nElle aime, elle craint, mais elle veut son style\nCécile 8, la femme du crocodile\nElle aime, elle craint, mais elle veut son style\nCecile 8, la femme du crocodile\nElle aime, elle craint, mais elle veut son style",
+    "bpm": 200,
+    "duration": 213,
+    "keyscale": "D minor",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_20.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Rap, adult, male, spoken word, singing, bright, energetic, clear",
-    "lyrics": "[Intro]\n他们说我来自阴影里\n说我的肤色是原罪的印记\n\n[Verse]\n眼神像刀子刮过 穿透我的皮肤\n带着审判和偏见 让我无处可逃处\n你没听过我的故事 没走过我的路\n凭什么就下一个判决 把我划出你的版图\n你说我威胁到你 抢走了你的机会\n可你可知我付出的 是你不敢想象的血泪\n被贴上标签 被区别对待\n呼吸都是错的 只因我生来就不一样态\n\n[Chorus]\n看不见的墙 把我阻隔在外面\n听不见的声音 屏蔽了我的呼唤\n他们制造偏见 他们散播谎言\n只因为我的存在 让他们觉得不安\n\n[Verse]\n每一次努力争取 都会被审视被放大\n每一个细微的错误 都变成攻击的靶\n他们选择性失明 看不见我的汗水\n只看见他们想看的 带着恶意的定位\n系统性的歧视 像一张无形的网\n把我困在原地 无法自由地翱翔\n他们在享受特权 却指责我的贫困\n嘲笑我的口音 我的名字 我的出身\n\n[Chorus]\n看不见的墙 把我阻隔在外面\n听不见的声音 屏蔽了我的呼唤\n他们制造偏见 他们散播谎言\n只因为我的存在 让他们觉得不安\n\n[Bridge]\n我不想寻求同情 只想被公平对待\n不想被定义被束缚 有选择自己未来的权利\n什么时候 才能放下心中的成见\n看到真正的我 而不是你脑海里的画面\n\n[Outro]\n画面... 不安...\n偏见... 歧视...\n什么时候能停止..."
+    "caption": "A classic, gritty hip-hop track built on a steady boom-bap drum machine groove and a deep, foundational sub-bass line. A male vocalist delivers a confident, narrative rap in Spanish, detailing the life of a tough-guy truck driver with a clear, rhythmic flow. The production is clean yet raw, with subtle atmospheric pads in the background and standard vocal reverb. The track transitions into a more reflective, spoken-word style bridge before concluding with an instrumental outro featuring a melodic synth line over the persistent beat.",
+    "lyrics": "[Intro]\nHey\nEste es pa' la güera\nEn San Antonio\nEl corazón frío\nPa' que quede frío\nAsí\nYeah\n\n[Verse 1]\nPa' la güera en el hospital\nDonde el 11 de 2001 suena criminal\nYa se metió pa' pegarle a la lejanía\nDonde el calor es de verdad, ¿dónde empieza el día?\nYa no es la abuela, la sonrisa en los zapatos\nAquí el café en la mesa ya tiene los cuajazos\nAquí el taller de ley ya viste el tajío\nY el viejo sin lluvia llega, qué desafío\nConecta, te conecta, siente el mensaje\nDale que en la calle no hay sabotaje\nLa calle es caldo, el asfalto pura vida\nAquí se escribe pa' cada chiva la movida\nYa es tradición, la historia bien escrita\nEn cada puerta que toca hay mil citas\nEl tío Javier ya se jodió el lío\nY ahora en el punk ya está su artilgo\nDesde Jalisco hasta la 115 en la piel\nYa está el vecindario y el volante a romper\nYa es borrar pendejos, el arte respeto\nPonte trucha en lo que te pillo, es nuestro decreto\n\n[Verse 2]\nUnos en ayuda, otros en parque\nOtros tienen GPS, otros cruzan barcos\nEn caminos que pasan sin conocer\nPero, ¿quién te pone el estilo, quién?\nEn 4L, el gas bajo, pa' allá nadie\nSomos tronos, tú ya cayó que esto no es botarla\nAsí que chinguen a su madre\nCon buenas frases las cantas, bro\n\n[Outro]\nY te voy pa' allá\nPa' la ruta\nAh\nHey\n[Instrumental fades out]",
+    "bpm": 100,
+    "duration": 124,
+    "keyscale": "F# minor",
+    "language": "es",
+    "timesignature": "4"
 }

examples/text2music/example_21.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Chorus Hook, Melodic Rap, Ambient Synth Pads, adult, rap, Very Fast, Storytelling, Chinese Rap, male, spoken word, bright, energetic, Melodic Flow, clear, clarity, 130 bpm",
-    "lyrics": "[Intro]\n舌 头 打 结 了... 快 念 快 念...\n\n[Verse 1]\n这 个 赌 鬼 蹲 在 柜 台 啃 着 苦 瓜 干 快 很 干\n赌 桌 堆 满 骨 牌 古 怪 股 票 和 五 块 钢 镚 儿 钢 镚\n他 甩 出 扑 克 牌 啪 啪 啪 拍 扁 螃 蟹 壳 哦 壳 扁\n又 摸 摸 麻 将 摸 出 幺 鸡 摸 出 发 财 摸 出 一 条 蛇 蛇 蛇\n庄 家 咳 嗽 咳 破 锣 嗓 子 喊 开 开 开 快 开 开\n赌 鬼 咕 嘟 咕 嘟 灌 咖 啡 灌 到 筷 子 戳 穿 碗 快 戳 穿\n空 气 里 飘 着 锅 巴 味 混 合 隔 夜 的 酸 奶 罐 哦 酸\n输 光 裤 带 还 想 翻 盘 翻 成 煎 饼 摊 老 板 快 翻 盘\n\n[Chorus]\n赌 鬼 赌 鬼 哦 赌 鬼 赌 鬼 快 很 快\n舌 头 打 结 着 念 这 段 哦 这 段 绕 口 令 牌\n若 念 错 一 字 就 罚 你 哦 罚 你 吞 十 斤 海 带\n赌 场 规 矩 就 是 绕 晕 你 哦 绕 晕 你 快 很 快\n\n[Verse 2]\n他 掏 出 铜 板 抠 出 口 袋 最 后 一 颗 快 很 颗\n庄 家 哗 啦 哗 啦 摇 骰 子 摇 出 三 点 又 三 点 哦 三 点\n赌 鬼 急 得 咬 牙 切 齿 咬 到 舌 头 打 蝴 蝶 结 快 打 结\n还 想 押 上 祖 传 的 拖 鞋 拖 把 铁 锅 和 半 包 盐 盐 盐\n突 然 警 笛 嘀 嘟 嘀 嘟 吓 得 他 钻 进 垃 圾 罐 哦 垃 圾\n警 察 咔 嚓 咔 嚓 拍 照 拍 到 他 头 顶 菠 菜 叶 快 拍 照\n最 后 赌 鬼 蹲 监 狱 天 天 背 这 首 绕 口 令 哦 背 不 完\n若 背 错 一 句 就 加 刑 十 年 再 加 十 年 快 加 刑\n\n[Outro]\n舌 头 打 结 了... 赌 鬼 哭 了 哦...\n这 首 歌... 绕 死 人 了 哦..."
+    "caption": "A smooth, jazzy lo-fi hip-hop track built on a foundation of a gentle piano melody and a relaxed, steady drum machine groove. A warm, round bassline provides a solid harmonic base. The song features a duet between a clear, melodic female vocalist and a smooth, conversational male vocalist who trade verses and harmonize beautifully in the choruses. The arrangement is punctuated by tasteful, melodic saxophone fills that enhance the jazzy, late-night atmosphere. The track concludes with an extended instrumental outro where the saxophone takes center stage with an expressive, improvisational solo over the core piano and rhythm section, before fading out with a final, lingering piano chord and a soft whoosh effect.",
+    "lyrics": "[Intro: Piano & Saxophone]\n\n[Verse 1: Female Vocal]\n月光落进窗\n像是起色的那杯\n你还在我的心底旋转\n抛开所有烦忧\n这爱像流星坠落\n在夜空闪耀\n无需解答 心动是唯一的解药\n\n[Verse 2: Male Vocal]\n你的眼神像迷宫\n我徘徊其中\n每个音符都带了心脏的轰鸣\n触碰你的温度就像电流穿心\n无法抗拒你的爱把我拉入梦境\n\n[Chorus: Duet]\n心跳随着你狂奔\n如海潮汹涌\n爱的节奏翻滚着\n像梦境中梦\n你的名字是旋律\n旋律中成风\n每一拍都诉说爱无声的悸动\n\n[Instrumental Break: Saxophone Solo]\n\n[Verse 3: Male Vocal]\n黑夜像个舞台\n我们点了彩灯\n节拍跟随心跳\n默契融成一吻\n别让夜晚停下来\n爱永远在认真\n每一秒都像歌\n开启奇妙人生\n\n[Chorus: Duet]\n心跳随着你狂奔\n如海潮汹涌\n爱的节奏翻滚着\n像梦境中梦\n你的名字是旋律\n旋律中成风\n每一拍都诉说爱无声的悸动\n\n[Bridge: Duet]\n当黎明来临时\n我们点燃光芒\n你的爱是我心中最暖的信仰\n再多风雨兼程\n也不愿停下\n因为你的存在是我唯一牵挂\n\n[Instrumental Break: Saxophone Solo]\n\n[Chorus: Duet]\n心跳随着你狂奔\n如海潮汹涌\n爱的节奏翻滚着\n像梦境中梦\n你的名字是旋律\n旋律中成风\n每一拍都诉说爱无声的悸动\n\n[Outro: Duet & Saxophone]\n当黎明来临时\n我们点燃光芒\n你的爱是我心中最暖的信仰\n再多风雨兼程\n也不愿停下\n因为你的存在是我唯一牵挂\n[Saxophone melody fades out]",
+    "bpm": 100,
+    "duration": 258,
+    "keyscale": "B♭ major",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_22.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "electronic, hip-hop, rap, synthesizer, drums, vocals, fast, energetic, modern, uplifting, young adult, male, spoken word, singing, bright, energetic, clear, 140 bpm, female",
-    "lyrics": "[Verse 1]\n红鲤鱼绿鲤鱼，驴在河里追鲤鱼，\n驴追鲤鱼鱼躲驴，气得驴子直喘气。\n扁担长板凳宽，扁担绑在板凳边，\n扁担要绑板凳不让绑，扁担偏要绑上板凳面！\n\n[Chorus]\n绕口令，练嘴皮，\n说快说慢别迟疑，\n红鲤鱼驴扁担板凳，\n一口气念完算你赢！\n\n[Verse 2]\n四是四十是十，十四是十四四十是四十，\n谁说四十是十四，舌头打结别放肆。\n黑化肥会挥发，灰化肥也发黑，\n化肥混一起，黑灰不分嘴发废！\n\n[Chorus]\n绕口令，练嘴皮，\n说快说慢别迟疑，\n四十十四化肥灰，\n念错罚你唱十回！\n\n[Bridge]\n坡上立着一只鹅，坡下流着一条河，\n鹅要过河河渡鹅，河要渡鹅鹅笑河——\n到底谁更啰嗦？！\n\n[Outro]\n嘴皮子功夫别小瞧，\n绕口令rap我最飙，\n下次挑战准备好，\n舌头打结别求饶！"
+    "caption": "An upbeat, celebratory synth-pop track driven by a punchy four-on-the-floor drum machine beat and a vibrant synth bassline. The song opens with a catchy, wordless vocal chop melody that serves as a recurring hook. A clear, enthusiastic male lead vocal carries the verses and soars into an anthemic, layered chorus. The production is polished and modern, featuring bright synth pads, melodic synth leads, and subtle atmospheric effects that create a sense of scale and joy. The track builds dynamically into each chorus, followed by an instrumental break that reprises the main vocal chop hook, before culminating in a powerful final chorus and a clean, decisive ending.",
+    "lyrics": "[Intro - Vocal Chop Melody]\n\n[Verse 1]\nÉbred a természet, vigan\nA végtelen a tájba\nAhol a nyár áll az iskola\nA szívünkben tombol a zaj\n\n[Pre-Chorus]\nFel is encontrató mámor\nEgy szolidi táncparkett\nEz a miénk tű early\nEnrég a tűz, enyhék a fák közt\n\n[Chorus]\nEz a mi időnk, rád gondolok\nTáncolok a holdvilág\nA szívünk dobban, de mégis újra élek\nMost és újra\nÉbred a természet, mindkötöm nevét most az élet\nSzállj velem, csak értünk lennem\nA város szívében élj egy szebb világban\n\n[Instrumental Break - Vocal Chop Melody]\n\n[Verse 2]\nA színek és áram szövedezik\nA parkett üteme hallom\nNézd, ahogy a fény táncot jár\nA világ velünk egy táncot vonva\n\n[Pre-Chorus]\nCsillagok mutatója\nEgyütt lélegzünk\nA tenger színeivel esküszünk\n\n[Chorus]\nEz a mi időnk, rád gondolok\nTáncolok a holdvilág\nA szívünk dobban, de mégis újra élek\nMost és újra\nÉbred a természet, mindkötöm nevét most az élet\nSzállj velem, csak értünk lennem\nA város szívében élj egy szebb világban\n\n[Bridge]\nA naplemente vezető ma\nCsak miénk az idő és a pillanat\nA természetünk vivan már\nMost és most húz a szívbe, hopp!\n\n[Instrumental Break - Vocal Chop Melody]\n\n[Outro]\nÉbredj már, együtt repülünk\nAhol a nyár a tengernek ad\nÚjra itt leszünk\nA természet vivan most és mindkötöm nevem",
+    "bpm": 201,
+    "duration": 184,
+    "keyscale": "F# minor",
+    "language": "hu",
+    "timesignature": "4"
 }

examples/text2music/example_23.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "singing, bright, slightly nasal, energetic, spoken word, young adult, male, rap music",
-    "lyrics": "[Intro]\nYo, check it—speed demon, lyrical heat, uh!\nRatatat like a drum when the beat bumps, uh!\n\n[Verse 1]\nRapatapa tap tap, flash like a snap,\nRap tap tap, I don’t chat, I clap clap clap!\nFingers snap, flow don’t slack, rapataptaptap,\nSpit it fast, hit the gas, rap tap tap rap!\n\n[Pre-Chorus]\nBoom-bap, zoom past, leave ’em flat,\nRap taptaprapataptaptap—where ya at?\n\n[Chorus]\nRapatapa tap tap, yeah, I go brrrr,\nRap tap tap, make the crowd stir!\nRapataptaptap, no lag, just spit,\nRap taptaprapataptaptap—I’m lit!\n\n[Verse 2]\nTongue-twist, quick wrist, rapatapa boom,\nTap tap rap, leave ya stuck like glue-gum!\nNo slow-mo, turbo, rapataptaptap,\nRap tap rap, yeah, I clap clap clap!\n\n[Outro]\nRapatapa—TAP! Mic drop—that’s that."
+    "caption": "A heartfelt Latin pop-rock ballad driven by a clean, melodic electric guitar that weaves intricate fills and arpeggios throughout the arrangement. The track opens with a distinctive guitar hook before settling into a steady groove laid down by an acoustic drum kit and a solid bassline. The lead male vocal is clear and emotional, delivering a nostalgic and melancholic narrative with increasing passion in the soaring choruses. The song's structure includes a dynamic bridge that shifts the mood towards hopefulness, followed by an expressive and melodic guitar solo that carries the emotional weight of the track before a final, powerful chorus and a gentle, reflective outro.",
+    "lyrics": "[Intro - Electric Guitar Melody]\n\n[Verse 1]\nTe extraño demasiado\nSin ti ya no es igual\nEstoy viendo tus fotos\nY el mundo se me va\nEspero que un día\nMe hayas vuelto a ver\nQue yo por lo menos\nPueda hacerte saber\n\n[Pre-Chorus]\nQue aunque ya no estés aquí\nMi corazón te extraña, sí\nY eso es todo lo que soy\nSiempre te recordaré\n\n[Chorus]\nY aunque ya no estés aquí\nMi corazón te extraña, sí\nY eso es todo lo que soy\nSiempre te recordaré\n\n[Instrumental Break - Guitar Melody]\n\n[Verse 2]\nSi acaso he pensado\nPero jamás te olvidé\nEres parte de mi vida\nUn regalo que soñé\nMe sobran tus recuerdos\nPero falta tu calor\nMe falta la esperanza\nDe que vuelvas, por favor\n\n[Pre-Chorus]\nY aunque ya no estés aquí\nMi corazón te extraña, sí\nY eso es todo lo que soy\nSiempre te recordaré\n\n[Chorus]\nY aunque ya no estés aquí\nMi corazón te extraña, sí\nY eso es todo lo que soy\nSiempre te recordaré\n\n[Bridge]\nTal vez el tiempo me traiga de vuelta a tu vida\nTal vez en tus días\nYa no quede cicatriz\nPero yo quiero contarte\nQue aún vives en mí\nQue te quiero, lo sabes\nY te extraño, lo sé\n\n[Chorus]\nY aunque ya no estés aquí\nMi corazón te extraña, sí\nY eso es todo lo que soy\nSiempre te recordaré\n\n[Outro]\nCariño mío\nLo sentiste tan real\nTe amo tanto\nQue ni lo puedo olvidar\nEres mi razón\nY mi eterno despertar\n\n[Guitar Solo]\n\n[Final Outro]\nCariño mío\nLo sentiste tan real\nTe amo tanto\nQue ni lo puedo olvidar\nEres mi razón\nY mi eterno despertar\n[Song fades out]",
+    "bpm": 100,
+    "duration": 245,
+    "keyscale": "A major",
+    "language": "es",
+    "timesignature": "4"
 }

examples/text2music/example_24.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "G-Funk, Hip Hop, Rap, Female Vocals, Melodic Rap, Summer, Laid-back Groove, Smooth Rhythm, Synthesizer Lead, Heavy Bassline, Groovy, West Coast Hip Hop",
-    "lyrics": "(Intro)\nOh yeah... \n\n(Verse 1)\n阳光下，沙滩排球场，一个身影跳跃\n小麦色，运动背心，闪耀活力四射\n她跳起扣杀，动作利落又巧妙\n汗水浸湿发梢，笑容比阳光更美好\n摇摆的节奏，是她的背景配乐\n每一次移动，都踩在鼓点上那么和谐\n我不由自主地停下脚步\n目光被她紧紧锁住\n\n(Chorus)\n沙滩排球女孩， 摇摆节拍下的身材\n无忧无虑的笑容，把我的心都填满\n想走上前去搭讪，嫌自己笨拙呆板\n这青春的气息，耀眼，灿烂！\n\n(Verse 3)\n她和队友击掌庆祝，笑声清脆悦耳\n拿起毛巾擦汗，不经意间瞥我一眼\n鼓起勇气走上前，假装问问时间\n她友好地回答，笑容灿烂没有敷衍\n聊了几句，发现彼此爱这摇摆音乐\n她眼中也闪过惊喜和亲切\n这共同点，让气氛变得融洽又热烈！\n夏天的故事，就这样开始了感觉真切！\n\n(Chorus)\n沙滩排球女孩， 摇摆节拍下的身材\n无忧无虑的笑容，把我的心都填满\n不再犹豫和等待，勇敢把脚步迈开\n这夏天的感觉，心跳，不断！"
+    "caption": "An energetic J-rock or anime-style rock track driven by powerful, distorted electric guitars and a driving drum beat. The song opens with a clean, arpeggiated guitar figure before exploding into a full-band arrangement with a soaring lead guitar melody. A clear, powerful female vocal delivers an emotional performance in Japanese, building from the verses into an anthemic, belted chorus. The arrangement features a dynamic bridge with atmospheric vocalizations and a melodic, technically proficient guitar solo complete with expressive bends and fast runs. The track concludes with a powerful final chorus and a climactic guitar flourish.",
+    "lyrics": "[Intro - Arpeggiated Electric Guitar]\n\n[Verse 1]\n晨光吻醒陌生的街道\n耳边风轻拂昨日的欢笑\n时针转为脚边的影子奔跑\n梦醒之后谁还在拥抱\n\n[Pre-Chorus]\n轻快的节拍 穿越了时间\n青春像火在燃烧着热血\n遗失的时间 留不住一切\n青春像光它忽明又忽灭\n\n[Chorus]\n黄昏的余音在耳边倾诉\n追逐的脚印已模糊模糊\n翻滚的青春是烈焰的舞\n我们在光阴里漫步反复\n\n[Instrumental Break - Guitar Melody]\n\n[Verse 2]\n汗水挥洒出闪亮的汗珠\n双手扣紧那未完的赌注\n岁月灼热月却描摹着温度\n青春绽放刺破天空深处\n\n[Pre-Chorus]\n轻快的节拍 穿越了时间\n青春像火在燃烧着热血\n遗失的时间 留不住一切\n青春像光它忽明又忽灭\n\n[Bridge - Vocalizations]\n(Ooooh)\n(Ooooh)\n(Ooooh)\n(Ooooh)\n\n[Guitar Solo]\n\n[Pre-Chorus]\n轻快的节拍 穿越了时间\n青春像火在燃烧着热血\n遗失的时间 留不住一切\n青春像光它忽明又忽灭\n\n[Chorus]\n黄昏的余音在耳边倾诉\n追逐的脚印已模糊模糊\n翻滚的青春是烈焰的舞\n我们在光阴里漫步反复\n\n[Outro - Guitar Solo over Chorus Chords]\n[Song ends abruptly]",
+    "bpm": 100,
+    "duration": 213,
+    "keyscale": "F minor",
+    "language": "zh",
+    "timesignature": "2"
 }

examples/text2music/example_25.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "lyrical rap, young adult, female, rap flow, spoken word, ad-libs, bright, energetic, eat, Fast, Engaging, Energetic",
-    "lyrics": "[Intro]\n扁擔寬 板凳長 扁擔想綁在板凳上\n扁擔寬 板凳長 扁擔想綁在板凳上\n\n[Verse]\n倫敦 瑪莉蓮 買了 件 旗袍 送 媽媽\n莫斯科 的 夫司基 愛上 牛肉 麵 疙瘩\n各種 顏色 的 皮膚 各種 顏色 的 頭髮\n嘴裡念的 說的 開始 流行 中國話 (中國話)\n\n[Bridge]\n多少年 我們 苦練 英文 發音 和 文法 (yeah)\n這幾年 換他們 捲著 舌頭 學 平上去入 的 變化\n平平 仄仄 平平 仄\n好聰明 的 中國人 好優美 的 中國話\n\n[Verse]\n扁擔寬 板凳長 扁擔想綁在板凳上\n板凳不讓扁擔綁在板凳上 扁擔偏要綁在板凳上\n板凳偏偏不讓扁擔綁在那板凳上\n到底扁擔寬 還是板凳長？\n\n[Verse]\n哥哥弟弟坡前坐\n坡上臥著一隻鵝 坡下流著一條河\n哥哥說 寬寬的河 弟弟說 白白的鵝\n鵝要過河 河要渡鵝\n不知是那鵝過河 還是河渡鵝\n\n[Chorus]\n全世界都在學中國話\n孔夫子的話 越來越國際化\n全世界都在講中國話\n我們說的話 讓世界都認真聽話\n\n[Verse]\n紐約蘇珊娜開了間禪風 lounge bar\n柏林來的沃夫岡拿胡琴配著電吉他\n各種顏色的皮膚 各種顏色的頭髮\n嘴裡念的 說的 開始流行中國話 (中國話)\n\n[Bridge]\n多少年我們苦練英文發音和文法 (yeah)\n這幾年換他們捲著舌頭學平上去入的變化\n仄仄平平仄仄平\n好聰明的中國人 好優美的中國話\n\n[Verse]\n有個小孩叫小杜 上街打醋又買布\n買了布 打了醋 回頭看見鷹抓兔\n放下布 擱下醋 上前去追鷹和兔\n飛了鷹 跑了兔 灑了醋 濕了布\n\n[Verse]\n嘴說腿 腿說嘴\n嘴說腿 愛跑腿\n腿說嘴 愛賣嘴\n光動嘴 不動腿\n光動腿 不動嘴\n不如不長腿和嘴\n到底是那嘴說腿 還是腿說嘴？\n\n[Chorus]\n全世界都在學中國話\n孔夫子的話 越來越國際化\n全世界都在講中國話\n我們說的話 讓世界都認真聽話\n\n[outro]\n全世界都在學中國話 (在學中國話)\n孔夫子的話 越來越國際化\n全世界都在講中國話\n我們說的話 (讓他) 讓世界 (認真) 都認真聽話"
+    "caption": "An explosive, high-energy anime rock anthem driven by a powerful male tenor vocal performance in Mandarin. The track kicks off with an anthemic 'whoa-oh' vocal hook over a driving four-on-the-floor beat, immediately establishing a J-rock or C-pop rock feel. The production is dense and polished, layering crunchy, distorted power-chord guitars with a punchy synth bass and a tight, modern drum kit. The choruses erupt with soaring, emotional vocals and layered harmonies, creating a massive wall of sound. A brief instrumental bridge features a chiptune-style synth arpeggio before a final, powerful chorus and an outro that fades out with the signature vocal hook.",
+    "lyrics": "[Intro]\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\n\n[Verse 1]\n满天星辰都沉默\n谁在黑暗中闪烁\n脚步轻踏入枷锁\n世界若你如此执着\n\n[Pre-Chorus]\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\n\n[Verse 2]\n你听风声在飘散\n梦的影子太凌乱\n无边黑夜探不安\n你的光芒怎能熄灭\n\n[Chorus]\n燃烧吧烈火之光\n黑暗中我们敢闯\n即使跌倒心不慌\n听那世界心跳响\n\n[Post-Chorus]\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\n\n[Verse 3]\n冷雨敲窗滴作响\n谁的伤痛扎中央\n双拳紧握去前方\n哪怕破碎也不会慌\n\n[Bridge]\n与爱相逢也能追\n日出时分睁开眼\n点起它如火蔓延\n这一切都有它的未来\n\n[Chorus]\n燃烧吧烈火之光\n黑暗中我们敢闯\n即使跌倒心不慌\n听那世界心跳响\n\n[Post-Chorus]\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\n\n[Instrumental Break]\n\n[Bridge]\n与爱相逢也能追\n日出时分睁开眼\n点起它如火蔓延\n这一切都有它的未来\n\n[Chorus]\n燃烧吧烈火之光\n黑暗中我们敢闯\n即使跌倒心不慌\n听那世界心跳响\n\n[Outro]\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\nWoah-oh-oh-oh-oh-oh\n[abrupt silence]",
+    "bpm": 200,
+    "duration": 190,
+    "keyscale": "E♭ major",
+    "language": "zh",
+    "timesignature": "2"
 }

examples/text2music/example_26.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "articulate, spoken word, young adult, warm, rap music, male, clear, street, dark, rap flow, hardcore rap",
-    "lyrics": "[verse]\n球场 的 橡胶味 弥漫 隔壁 是 健身房\n场 边上 的 老教练 战术 有 三套\n教 交叉 运球 的 大叔 会 欧洲步 耍 背后 传\n硬 身板 对抗 最 擅长 还 会 急停跳 后仰 投\n他们 徒弟 我 习惯 从小 就 耳濡目染\n什么 胯下 跟 变向 我 都 玩 的 有模有样\n什么 招式 最 喜欢 转身 过 人 柔中 带 刚\n想要 去 纽约 街头 斗 洛克 公园 场\n\n[chorus]\n看什么 看什么\n变速 突破 心 自在\n看什么 看什么\n假动作 晃 开 防守 来\n看什么 看什么\n每日 训练 绑 沙袋\n空中拉杆 莫 奇怪\n唰唰 入袋\n\n[verse]\n一个 试探 步后 一记 左 变向 右 变向\n一句 挑衅 我 的 人 别 嚣张\n一再 重演 一颗 我 不 投 的 球\n悬在 篮筐 上 它 一直 在 摇晃\n\n[chorus]\n看什么 看什么\n我 激活 小宇宙 来\n看什么 看什么\n菜鸟 新人 的 名号\n看什么 看什么\n已 被 我 一球 击倒\n\n[chorus]\n快 秀出 指尖 转球 砰砰 啪嗒\n快 秀出 指尖 转球 砰砰 啪嗒\n篮球 之 人 切记 勇者 无惧\n是 谁 在 玩 花式 引爆 空气\n快 秀出 指尖 转球 砰砰 啪嗒\n快 秀出 指尖 转球 砰砰 啪嗒\n如果 我 有 滞空 逆天 补扣\n为人 热血 不怂 一生 傲骨 吼\n\n[verse]\n他们 徒弟 我 习惯 从小 就 耳濡目染\n什么 胯下 跟 变向 我 都 玩 的 有模有样\n什么 招式 最 喜欢 转身 过 人 柔中 带 刚\n想要 去 纽约 街头 斗 洛克 公园 场\n\n[outro]\n快 秀出 指尖 转球 砰\n快 秀出 指尖 转球 砰\n如果 我 有 滞空 吼\n为人 热血 不怂 一生 傲骨 吼\n快 秀出 指尖 转球 砰\n我 用 背传 助攻 吼\n压哨 的 三分 球"
+    "caption": "A modern reggaeton track with a confident, defiant energy, built on a classic dembow drum machine beat and a deep sub-bass line. The song opens with a catchy, pitched-up vocal chop melody that serves as a recurring hook. The lead female vocal is clear and assertive, delivering lyrics with a rhythmic, almost confrontational flow. The chorus expands with layered vocal harmonies, adding power and emphasis. A bridge section introduces a filtered, telephone-like vocal effect for contrast before the track builds back into a final, powerful chorus and fades out with the initial vocal chop melody and atmospheric synth pads.",
+    "lyrics": "[Intro - Vocal Chop Melody]\n\n[Verse 1]\nMe decían que no iba a llegar\nPero este paso me hizo dudar\nTodo lo que quise, todo lo que fui\nHoy vuelven y quieren venir hacia mí\nDicen que soy fría, que no valgo na'\nPero lo que ves es lo que gané ya\nNo me digas pobre, yo ya me cansé\nConozco lo duro que vengo y lo sé\n\n[Chorus]\nDicen que no valgo, que yo soy raro\nPero si me ven raro, mejor ni me hablen claro\nLa envidia y la envidia, eso no se nota\nNo me vendo, no me borran, no me joden en la boca\n\n[Verse 2]\nTengo el estilo que quieren copiar\nPero yo prefiero a quien me quiera mirar\nDicen que me odian, que quieren juzgar\nPero no me importa, ya aprendí a luchar\nTengo el estilo que quieren copiar\nPero yo prefiero a quien me quiera mirar\nDicen que me odian, que quieren juzgar\nPero no me importa, ya aprendí a luchar\n\n[Chorus]\nDicen que no valgo, que yo soy raro\nPero si me ven raro, mejor ni me hablen claro\nLa envidia y la envidia, eso no se nota\nNo me vendo, no me borran, no me joden en la boca\n\n[Bridge - Filtered Vocals]\nNo soy de los que suelen juzgar\nNo soy de los que saben tratar\nDicen que he cambiado, que ya no soy igual\nPero no me quejo, no voy a parar\n\n[Chorus - Layered Vocals]\nDicen que no valgo, que yo soy raro\nPero si me ven raro, mejor ni me hablen claro\nLa envidia y la envidia, eso no se nota\nNo me vendo, no me borran, no me joden en la boca\n\n[Outro]\nNo me vendo, no me borran, no me joden en la boca\nNo me vendo, no me borran, no me joden en la boca\n[Vocal chop melody fades out]",
+    "bpm": 100,
+    "duration": 140,
+    "keyscale": "E♭ minor",
+    "language": "es",
+    "timesignature": "4"
 }

examples/text2music/example_27.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "articulate, spoken word, young adult, warm, rap music, male, clear, street, dark, rap flow, hardcore rap, fast",
-    "lyrics": "[verse]\n球场 的 橡胶味 弥漫 隔壁 是 健身房\n场 边上 的 老教练 战术 有 三套\n教 交叉 运球 的 大叔 会 欧洲步 耍 背后 传\n硬 身板 对抗 最 擅长 还 会 急停跳 后仰 投\n他们 徒弟 我 习惯 从小 就 耳濡目染\n什么 胯下 跟 变向 我 都 玩 的 有模有样\n什么 招式 最 喜欢 转身 过 人 柔中 带 刚\n想要 去 纽约 街头 斗 洛克 公园 场\n\n[chorus]\n看什么 看什么\n变速 突破 心 自在\n看什么 看什么\n假动作 晃 开 防守 来\n看什么 看什么\n每日 训练 绑 沙袋\n空中拉杆 莫 奇怪\n唰唰 入袋\n\n[verse]\n一个 试探 步后 一记 左 变向 右 变向\n一句 挑衅 我 的 人 别 嚣张\n一再 重演 一颗 我 不 投 的 球\n悬在 篮筐 上 它 一直 在 摇晃\n\n[chorus]\n看什么 看什么\n我 激活 小宇宙 来\n看什么 看什么\n菜鸟 新人 的 名号\n看什么 看什么\n已 被 我 一球 击倒\n\n[chorus]\n快 秀出 指尖 转球 砰砰 啪嗒\n快 秀出 指尖 转球 砰砰 啪嗒\n篮球 之 人 切记 勇者 无惧\n是 谁 在 玩 花式 引爆 空气\n快 秀出 指尖 转球 砰砰 啪嗒\n快 秀出 指尖 转球 砰砰 啪嗒\n如果 我 有 滞空 逆天 补扣\n为人 热血 不怂 一生 傲骨 吼\n\n[verse]\n他们 徒弟 我 习惯 从小 就 耳濡目染\n什么 胯下 跟 变向 我 都 玩 的 有模有样\n什么 招式 最 喜欢 转身 过 人 柔中 带 刚\n想要 去 纽约 街头 斗 洛克 公园 场\n\n[outro]\n快 秀出 指尖 转球 砰\n快 秀出 指尖 转球 砰\n如果 我 有 滞空 吼\n为人 热血 不怂 一生 傲骨 吼\n快 秀出 指尖 转球 砰\n我 用 背传 助攻 吼\n压哨 的 三分 球"
+    "caption": "A moody, atmospheric trap track opens with a clean, arpeggiated synth melody and a deep sub-bass foundation. A crisp trap beat with punchy kicks and rattling hi-hats drives the rhythm. The male lead vocal, delivered in Polish, shifts between melodic rapping and singing, conveying a sense of urban melancholy and introspection. The choruses are marked by a prominent, expressive saxophone melody that weaves through the mix, adding a layer of jazzy, melancholic character. The track progresses through verses and choruses, culminating in an extended instrumental outro where the beat deconstructs, leaving behind atmospheric synth pads and a final, lingering saxophone motif before fading into a music box-like synth melody.",
+    "lyrics": "[Intro]\nYeah\nYeah\n\n[Verse 1]\nDrzwi otwarte na zakrętach, suki uderzały w lód\nEj, ej\nW białych obłokach mrok unosi, to sen dla nas jest blues\nBłyskawice na kolanach, a serca na dłoni\nŻycie bez reguł, jedno ja i ona, patrząc w prawdę, ale bez słów\nŻycie bez reguł, jedno ja i ona, patrząc w prawdę, ale bez słów\n\n[Chorus - Instrumental Drop]\n[Synth lead melody with vocal chops]\n\n[Verse 2]\nZnów nie widać dla nich, patrz, szkło zasłania się światłem\nEj, ej\nKażda zła chwila to ciężar, by zgubić ślad choćby jasne\nBłyskawice na kolanach, a serca na dłoni\nŻycie bez reguł, jedno ja i ona, patrząc w prawdę, ale bez słów\nŻycie bez reguł, jedno ja i ona, patrząc w prawdę, ale bez słów\n\n[Chorus - Instrumental Drop]\n[Synth lead melody with vocal chops]\n\n[Bridge]\nZnów nie widać dla nich, patrz, szkło zasłania się światłem\nKażda zła chwila to ciężar, by zgubić ślad choćby jasne\n[Synth arpeggios and beat]\n\n[Chorus - Instrumental Drop]\n[Synth lead melody with vocal chops]\n\n[Outro]\n[Synth lead melody with vocal chops and beat]\n[Beat fades out, synth melody continues and fades]",
+    "bpm": 100,
+    "duration": 177,
+    "keyscale": "F minor",
+    "language": "pl",
+    "timesignature": "4"
 }

examples/text2music/example_28.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "articulate, spoken word, young adult, rap music, female, clear, energetic, warm",
-    "lyrics": "[Intro]\n\"System booting... 语言 模型 loading...\"\n\n[Verse 1]\n硅谷 那个 coder 调试 neural network\n北京 的 极客 训练 A I 写 report\n不同 架构 的 chip 不同 算法 的 war\n屏幕上 跑的 全是 machine learning (learning)\n\n[Bridge]\n多少年 我们 chase 摩尔 定律 的 trend (yeah)\n这两年 换他们 study 中文 N L P\nConvolution L S T M\n好烧脑 的 backprop 好暴力 的 big data\n\n[Verse 2]\nPython 强 say加加 刚 Python 调用 C++ 的 A P I\nsay加加 嫌 Python 太 slow Python 笑 C++ 太 hardcore\nL L V M 默默 generate 中间 code\n到底 interpreter 还是 compiler 屌？\n\n[Verse 3]\nP M 和 engineer\n白板 画满 flow chart 服务器 闪着 red light\nP M 说 add feature engineer 说 no way\n需求 变更 code 重构\n不知 是 P M 太 fly 还是 deadline 太 high\n\n[Chorus]\n全世界 都在 train neural network\nTransformer 的 paper 越来越 难 go through\n全世界 都在 tune 超参数\n我们 写的 bug 让 G P U 都 say no\n\n[Verse 4]\n柏林 hackathon demo blockchain contract\n上海 的 dev 用 federated learning 破 data wall\n各种 语言 的 error 各种 框架 的 doc\nterminal 里 滚的 全是 dependency 冲突\n\n[Bridge]\n曾以为 English 才是 coding 的 language (yeah)\n直到见 G P T 用 文言文 generate 正则 expression\nGradient explode\n好硬核 的 prompt 好头秃 的 debug road\n\n[Verse 5]\n有个 bug 叫 quantum\n测试 环境 run perfect 上线 立即就 crash\n查 log 看 monitor 发现是 thread 不同步\n改 sync 加 lock 慢 deadlock 更难办\n量子 computer 也解不开 这 chaos chain\n\n[Verse 6]\n你说 996 我说 007\n你说 福报 我说 burnout\nProduct 要 agile Boss 要 KPI\nCode 要 elegant deadline 是 tomorrow\n不如 直接 script 自动 submit 离职信\n\n[Outro]\n\"Warning: 内存 leak...core dumping...\"\n全世界 都在 train neural network (neural network)\nLoss 还没 converge 天已经亮\n全世界 都在 tune 超参数\n我们 写的 code (让它) 让 world (reboot) 都 reboot 无效"
+    "caption": "An energetic, high-tempo J-pop and chiptune fusion track driven by a relentless four-on-the-floor drum machine beat and a pulsing synth bassline. The soundscape is dominated by bright, arpeggiated 8-bit style synthesizers that create a vibrant, video-game-like atmosphere. A clear, high-pitched female vocal delivers the melody with an upbeat, almost frantic energy, soaring over the dense electronic arrangement. The structure includes dynamic builds and drops, with a brief, atmospheric bridge featuring a whispered English vocal line before launching back into the powerful, synth-laden chorus and an instrumental synth-heavy outro.",
+    "lyrics": "[Intro - Synth Arpeggio]\n\n[Verse 1]\n森の奥で育ちの音\n素衣ざなたてが木が笑う\n流れてゆく時の中で\nそれたちの声が踊り出す\n\n[Pre-Chorus]\n鳥が歌えば風がふわりと\n母の愛を運んできたよ\n子供たちの背中を追って\n深い森の上で生きる\n\n[Chorus]\n森の鳥語るよ\n好き君を想うその時に\n月夜の影忍び寄る\n静けさの中君を呼ぶ\n\n[Instrumental Break]\n\n[Verse 2]\n木の下で続く坂道\n鳥たちの声は夢を誘う\n絵を描いた木の夢\n木が叫ぶ運命の声\n\n[Pre-Chorus]\n時は流れて僕は行く\n名もない木は今もここに\n君がくれたそのぬくもり\n風がそっと背中を押す\n\n[Chorus]\n森の鳥語るよ\n好き君を想うその時に\n月夜の影忍び寄る\n静けさの中君を呼ぶ\n\n[Bridge]\n月の影に隠れた船\n季節巡り風が歌う\n風が呼ぶ声の先に\n静かにささやく君の名を\n\n[Breakdown]\n[whispered]\n響けその歌で\n静かに響く木の音\n\n[Instrumental Break]\n\n[Outro]\n森の奥の果てまで\n[Synth arpeggio fades out]",
+    "bpm": 100,
+    "duration": 218,
+    "keyscale": "F minor",
+    "language": "ja",
+    "timesignature": "4"
 }

examples/text2music/example_29.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "articulate, spoken word, young adult, rap music, male, clear, energetic, warm, relaxed, breathy, night club, auto-tune, mumble rap, trap",
-    "lyrics": "[verse]\n这 这 谁 又 在 派 对 喝 多\n我 的 脑 袋\n像 被 驴 踢 过\n不 对 劲\n舌 头 打 结 不 会 说\n你 来 挑 战 我 就 跪\n开 局 直 接 崩 溃\n\n[chorus]\n就 咪 乱 咪 念 咪 错 咪\n嘴 咪 瓢 咪 成 咪 狗 咪\n脑 咪 袋 咪 像 咪 浆 咪 糊 咪\n跟 咪 着 咪 节 咪 奏 咪\n把 咪 歌 咪 词 咪 全 咪 忘 咪\n一 咪 张 咪 嘴 咪 就 咪 废 咪\n只 咪 剩 咪 下 咪 尴 咪 尬 咪 回 咪 忆\n草！\n\n[verse]\n错 错 错 错 了\n一 口 气 全 念 错\n错 错 错 错 了\n舌 头 打 结 甩 锅\n甩 甩 甩 甩 锅\n甩 锅 甩 锅\n拍 子 全 部 乱 套\n观 众 笑 到 吐 血\n\n[verse]\n你 的 歌 词 我 的 噩 梦\n唱 完 直 接 社 死\n调 跑 到 外 太 空\n观 众 表 情 裂 开\n你 笑 我 菜\n我 笑 你 不 懂\n这 叫 艺 术 表 演\n不 服 你 来！\n\n[verse]\n这 这 谁 又 在 派 对 丢 人\n我 的 世 界\n已 经 彻 底 崩 溃\n没 有 完 美\n只 有 翻 车 现 场\n以 及 观 众 的 嘲 讽\n\n[chorus]\n就 咪 乱 咪 念 咪 错 咪\n嘴 咪 瓢 咪 成 咪 狗 咪\n脑 咪 袋 咪 像 咪 浆 咪 糊 咪\n跟 咪 着 咪 节 咪 奏 咪\n把 咪 歌 咪 词 咪 全 咪 忘 咪\n一 咪 张 咪 嘴 咪 就 咪 废 咪\n只 咪 剩 咪 下 咪 尴 咪 尬 咪 回 咪 忆\n草！\n\n[verse]\n错 错 错 错 了\n一 口 气 全 念 错\n错 错 错 错 了\n舌 头 打 结 甩 锅\n甩 甩 甩 甩 锅\n甩 锅 甩 锅\n拍 子 全 部 乱 套\n观 众 笑 到 吐 血\n\n[verse]\n你 的 歌 词 我 的 噩 梦\n唱 完 直 接 社 死\n调 跑 到 外 太 空\n观 众 表 情 裂 开\n你 笑 我 菜\n我 笑 你 不 懂\n这 叫 艺 术 表 演\n不 服 你 来！"
+    "caption": "An energetic chiptune-pop track driven by a punchy electronic drum machine and a bright, catchy 8-bit synth lead that plays a memorable riff. A clear, powerful male vocal delivers an anthemic melody in Mandarin, soaring over the vibrant digital arrangement. The song follows a classic pop structure with verses building into an explosive, uplifting chorus. A brief, more atmospheric bridge provides a moment of reflection before launching back into the final chorus and an instrumental outro that fades out on the iconic synth melody.",
+    "lyrics": "[Intro - Chiptune Synth Melody]\nOh!\n\n[Verse 1]\n翻过山峰\n找寻光\n脚步轻快像飞翔\n天空洒满\n梦的光\n未来在呼唤我方向\n\n[Chorus]\n一起飞越那片天\n勇敢拥抱每个瞬间\n冒险是我们的语言\n心跳同步无边无边\n\n[Instrumental Break - Chiptune Synth Melody]\n\n[Verse 2]\n风吹过\n沙漠青烟\n勇气化作星的线\n脚步带着\n希望远\n未知旅途从不停歇\n\n[Chorus]\n一起飞越那片天\n勇敢拥抱每个瞬间\n冒险是我们的语言\n心跳同步无边无边\n\n[Instrumental Break - Chiptune Synth Melody]\n\n[Bridge]\n脚印落在时间线\n画下每一个明天\n星辰为我们兑现\n爱是唯一的心愿\n\n[Chorus]\n一起飞越那片天\n勇敢拥抱每个瞬间\n冒险是我们的语言\n心跳同步无边无边\n\n[Outro - Chiptune Synth Melody & Beat]\n[Beat fades out, synth melody continues and fades]",
+    "bpm": 100,
+    "duration": 163,
+    "keyscale": "B♭ major",
+    "language": "zh",
+    "timesignature": "4"
 }

examples/text2music/example_30.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "四川话, spoken word, male, Tempo - Fast, Elements - Chorus Hook, Subgenre-Satirical Hip Hop, Rap, Chinese-language music, energetic, slightly nasal, Instrument - Live Bass Guitar, adult, Vocals - Syncopated Flow, Genre - Hip-Hop, rapping, bright",
-    "lyrics": "[chorus]\n黑 墨镜 金 链子 越 低调 越 霸气\n玩 街机 泡 吧里 再 野的 场子 都 不 怯气\n上海 滩 老 江湖 外滩 钟声 敲 胜负\n陆家嘴 黄浦江 财路 宽 给 你 开 扇窗\n\n[verse]\n老子 在 弄堂 斜起 走 想 拦路 的 先 报 名号\n我 早看透 你们 手抖 脚软\n只敢 网上 吠 现实 怂成 猫\n看 你们 混的 真 可怜 整天 蹲在 网吧 蹭 烟\n钱 赚不到 架 不敢打 还 学人 摆 大哥 脸\n\n[verse]\n叫 我 沪上 老 克勒 不是 拉菲 我 不 碰杯\n规矩 我 懒得 讲 太多 钞票 直接 拍 你 脸上 飞\n老子 耐心 差 门槛 高 你 找茬 等于 自 寻 烦恼\n要么 跪 要么 爬 最后 警告 只 说 一 遭\n\n[chorus]\n黑 墨镜 金 链子 越 低调 越 霸气\n玩 街机 泡 吧里 再 野的 场子 都 不 怯气\n上海 滩 老 江湖 外滩 钟声 敲 胜负\n陆家嘴 黄浦江 财路 宽 给 你 开 扇窗\n\n[verse]\n古巴 雪茄 在 指间 绕 代表 魔都 格调 必须 顶\nOG 在 你 够不到 的 高度 My bro 永远 在 顶层 盯\nCheck my vibe 不靠 大 金劳 留声机 放 周璇 和 白光\n爹妈 太 宠你 养出 巨婴 症 早晚 社会 教你 做人 经\n\n[verse]\n玩 说唱 小囡 太 年轻 要 比 flow 先去 练 气功\n廿年 磨 枪 才 亮 锋芒 我 三十六 招 收 你 入 瓮\n老子 存在 就是 打假 标\n多少 人 眼红 又 不敢 挑\n键盘 侠 的 狠话 像 棉花 糖\n见 真人 秒变 Hello Kitty 叫\n\n[chorus]\n黑 墨镜 金 链子 越 低调 越 霸气\n玩 街机 泡 吧里 再 野的 场子 都 不 怯气\n上海 滩 老 江湖 外滩 钟声 敲 胜负\n陆家嘴 黄浦江 财路 宽 给 你 开 扇窗\n\n[chorus]\n黑 墨镜 金 链子 越 低调 越 霸气\n玩 街机 泡 吧里 再 野的 场子 都 不 怯气\n上海 滩 老 江湖 外滩 钟声 敲 胜负\n陆家嘴 黄浦江 财路 宽 给 你 开 扇窗"
+    "caption": "An upbeat French pop-rap track built on a clean, catchy electric guitar riff that loops throughout. The production is modern and crisp, featuring a punchy drum machine beat with sharp claps and steady hi-hats, anchored by a smooth synth bassline. A confident male lead vocal delivers rhythmic verses and an anthemic chorus filled with layered harmonies for emphasis. The song includes sections where the vocals become more processed and filtered, adding texture before transitioning into a brief instrumental break highlighting the main guitar motif alongside melodic humming ad-libs.",
+    "lyrics": "[Intro]\n[Clean electric guitar riff]\nYeah\nYeah\n\n[Verse 1]\nDans le game, on trace, on trace\nDès le matin, on ne se relâche\nRêve en couleur, rêve en or\nTous les autres, quelqu'un ignore\n\n[Chorus]\nMains dans les poches\nLèvres dans les nuits\nLis-moi pas, moi\nC'est quoi? C'est qu'une vraie vie\n\n[Verse 2]\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\n\n[Pre-Chorus]\nMains dans les poches, lèvres dans les nuits\nLis-moi pas, moi, c'est quoi? C'est qu'une vraie vie\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\n\n[Chorus]\nMains dans les poches\nLèvres dans les nuits\nLis-moi pas, moi\nC'est quoi? C'est qu'une vraie vie\n\n[Bridge]\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\n\n[Pre-Chorus]\nMains dans les poches, lèvres dans les nuits\nLis-moi pas, moi, c'est quoi? C'est qu'une vraie vie\n\n[Chorus]\nMains dans les poches, lèvres dans les nuits\nLis-moi pas, moi, c'est quoi? C'est qu'une vraie vie\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\n\n[Outro]\nMains dans les poches\nLèvres dans les nuits\nLis-moi pas, moi\nC'est quoi? C'est qu'une vraie vie\n[Instrumental break with guitar solo]\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\nMains dans les poches, lèvres dans les nuits\nLis-moi pas, moi, c'est quoi? C'est qu'une vraie vie\nJe pars en bataille, sourire au coin\nNouvelle vision, non, j'en ai rien\nToujours au top, jamais en panne\nLa révolution vient d'une autre âge, c'est l'heure du barrage\n[Music fades out]",
+    "bpm": 100,
+    "duration": 220,
+    "keyscale": "A major",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_31.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Rap, Chinese Rap, J-Pop, Anime, kawaii pop, EDM, Aggressive, Intense, Crisp Snare, Super Fast, Clear",
-    "lyrics": "(Intro)\nLet's drift away...\n\n(Verse 1)\n现实是灰色的格子间，重复的工作，枯燥的报表 \n敲打着键盘，眼神却放空，意识早已挣脱了肉体的镣铐\n飘向窗外，飞过拥挤的街道，穿过云层，到达想象的群岛\n那里色彩斑斓，形状奇异，逻辑失效，一切都随心所欲地飘摇\n迷幻的鼓点，像心跳的变奏，忽快忽慢，难以预料\n抽象的采样，扭曲的人声，构建一个超现实的音景环绕\n我变成一只鸟，一条鱼，一束光，自由地变换形态和奔跑\n在这白日梦里，我无所不能，摆脱了所有现实的烦恼， feeling the afterglow\n\n(Chorus)\n意识漫游，逃离乏味的轨道 \n迷幻嘻哈的节拍，是白日梦的引导 \n抽象的世界，逻辑被重新构造\nMind wandering free, where reality starts to fade slow\n\n(Verse 2)\n会议室里老板在讲话，声音模糊，像隔着水听不清道\n我的思绪，早已潜入深海，与发光的水母一起舞蹈\n或者飞向外太空，在星云间穿梭，探索未知的星球和轨道\n现实的规则，在这里被打破，物理定律也失去效劳\n白日梦是我的避难所，是精神的氧气罩\n在乏味的现实里，为我注入一点色彩和奇妙\n虽然短暂，虽然虚幻，但它让我能够喘息，重新把能量找到\n然后回到现实，继续扮演那个，循规蹈矩的角色，把梦藏好， keep the dream aglow\n\n(Chorus)\n意识漫游，逃离乏味的轨道\n迷幻嘻哈的节拍，是白日梦的引导\n抽象的世界，逻辑被重新构造\nMind wandering free, where reality starts to fade slow\n"
+    "caption": "A classic country-folk ballad driven by a steady acoustic guitar rhythm and a warm, earnest male vocal with a gentle twang. A simple bassline and a straightforward drum beat establish a heartfelt, mid-tempo groove. A plaintive harmonica weaves in and out, providing melodic fills and taking the lead during the instrumental breaks and the extended outro. The production is clean and direct, focusing on the storytelling and the sincere vocal performance, creating a nostalgic and romantic atmosphere.",
+    "lyrics": "[Intro - Acoustic Guitar]\n\n[Verse 1]\nFound you standing in the river's roar\nHeartbeats racing like never before\nEyes met mine and time stood still\nDreams were spinning 'round this hill\nWalking barefoot on a dusty road\nFingers tracing where the pelo grow\nHear your laugh, it's a sweet melody\nLove a tune that sets me free\n\n[Chorus]\nI'll ride these heartstrings all night long\nSing our song where we belong\nIn the quiet moments where we're strong\nYour heartbeat's my favorite song\n\n[Instrumental Break - Harmonica Solo]\n\n[Verse 2]\nStars above us start to glow\nEvening whispers, soft and low\nUnderneath the moonlit sky\nWe're just two souls, you and I\nIn the silence, hearts will find their tune\nUnderneath the prairie moon\n\n[Instrumental Break - Acoustic Guitar]\n\n[Bridge]\nI'll ride those heartstrings 'til they stray\nWe'll find our way back someday\n\n[Chorus]\nI'll ride these heartstrings all night long\nSing our song where we belong\nIn the quiet moments where we're strong\nYour heartbeat's my favorite song\n\n[Outro - Harmonica and Guitar]\n[Final guitar strum]",
+    "bpm": 200,
+    "duration": 202,
+    "keyscale": "A major",
+    "language": "en",
+    "timesignature": "4"
 }

examples/text2music/example_32.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "J-Pop, Anime, kawaii future bass, Femal vocals, EDM, Boombap, Aggressive, Intense, Crisp Snare, Super Fast, Rap",
-    "lyrics": "[Intro]\nYo, 这是来自深渊的怒吼\n\n[Verse]\n指尖飞快刷新，屏幕又亮起\n渴望那点赞，像致命的氧气\n精心修饰的脸庞，完美到诡异\n背后隐藏的疲惫，谁又会在意\n光鲜亮丽的橱窗，贩卖着焦虑\n每个人都在表演，戴着虚伪面具\n比较的游戏，让人逐渐窒息\n迷失在数据洪流，找不到自己\n\n[Chorus]\n这流量的时代，真假早已分不清\n盲目追随潮流，丢掉了初心\n为了那点虚荣，灵魂在沉沦\n看不见的锁链，捆绑每个灵魂\n\n[Verse]\n滤镜下的生活，美得不切实际\n营造虚假繁荣，掩盖内心空虚\n他人的光环下，显得自己多余\n嫉妒和自卑，交织成悲剧\n\n[Chorus]\n朋友圈里炫耀，现实中却叹气\n刷着别人的故事，忘记了呼吸\n算法推荐着你，想看的一切东西\n不知不觉间，你已不再是你\n他们说这是进步，我看是种病\n精神鸦片侵蚀，慢慢要了你的命\n\n[Bridge]\n屏幕亮了又暗，一天又过去\n究竟得到了什么，还是失去了自己\n那真实的连接，在何处寻觅\n困在这迷宫里，找不到出口的轨迹\n\n[Outro]\n我想挣脱，我想呼吸\n这虚拟的繁华，让我喘不过气\n谁能告诉我，这到底有什么意义\n一切都像泡沫，一触就破裂没余地"
+    "caption": "An energetic, bilingual J-pop and French pop vocal performance over a driving electronic beat. The track is built on a foundation of punchy synth bass, crisp drum machine rhythms, and layered synthesizers. A powerful female lead vocal delivers catchy melodies in Japanese, contrasted by a male rapper who performs a verse in French and English. The production is polished and modern, featuring vocal chops, synth arpeggios, and dynamic builds that lead into an anthemic, soaring chorus with layered harmonies and expressive vocal ad-libs. The song concludes with a brief, reflective spoken-word section in Japanese before fading out.",
+    "lyrics": "[Intro]\nKaze ni notte\nAtarashii\nDoboji to omoide\n\n[Verse  1]\nJono wa hirogaru  100% yorokobi\nKamishime dori mo imi ni naru shi\nMinna ore o yurugasu\nTaifuu ga shinai no kamo\n\n[Pre-Chorus]\nJinchou onna no you ni ikki\nKoukou de kyou mo ikitsuku\nJiyuu no kaze ni notte yotte\nDoboji to omoide\n\n[Chorus]\nChoushi sawarazu\nHidari ni wo terashite\nKyou mo tokubetsu sugite\nOre no jinsei wo kakeru\n\n[Post-Chorus]\nSubete wo yakareta ore no machi\nIchido ikiru nara imi ga wakaru\nOmae wa sou yatte karada wo katte\nKuzusenai nai\nKoko de doko made mo yukou ka\n\n[Instrumental Break]\n\n[Verse  2 -  Male Rap]\nCulte de fiesta\nHousou shite\nKyou mo tokubetsu sugite\nOre no jinsei wo kakeru\n\n[Bridge]\nSubete wo yakareta ore no machi\nIchido ikiru nara imi ga wakaru\nOmae wa sou yatte karada wo katte\nKuzusenai nai\nKoko de doko made mo yukou ka\n\n[Chorus]\nSubete wo yakareta ore no machi\nIchido ikiru nara imi ga wakaru\nOmae wa sou yatte karada wo katte\nKuzusenai nai\nKoko de doko made mo yukou ka\n\n[Outro]\nSubete wo yakareta ore no machi\nIchido ikiru nara imi ga wakaru\nOmae wa sou yatte karada wo katte\nKuzusenai nai\nKoko de doko made mo yukou ka\nSaa saa doko made mo oikakero\nJiyuu no kaze ni notte yotte\nDoboji to omoide\n[Synth arpeggio fades out]",
+    "bpm": 100,
+    "duration": 197,
+    "keyscale": "F# major",
+    "language": "ja",
+    "timesignature": "4"
 }

examples/text2music/example_33.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Hip Hop, Hi-hat Rolls, spoken word, Melodic Flow, articulate, Female Rap, 120 BPM, clear, warm, female, melodic Rap, adult, super fast",
-    "lyrics": "[Verse 1]\n打南边来了个喇嘛，手里提拉着五斤鳎目，\n打北边来了个哑巴，腰里别着个喇叭。\n喇嘛想换哑巴的喇叭，哑巴摇头不说话，\n鳎目一甩像道闪电，喇叭一响震天涯！\n\n[Chorus]\n丁丁当当，乒乓乓乓，\n话赶话，舌绕梁，\n东边的钉，西边的墙，\n绕不完的弯，唱不完的慌！\n\n[Verse 2]\n墙上一根钉，钉下绳摇晃，\n绳吊着瓶，瓶碰碎了光。\n灯骂瓶，瓶怪绳，绳怨钉，\n稀里哗啦，一场荒唐！\n\n[Chorus]\n丁丁当当，乒乓乓乓，\n话赶话，舌绕梁，\n东边的钉，西边的墙，\n绕不完的弯，唱不完的慌！\n\n[Verse 3]\n板凳宽，扁担长，\n一个偏要绑，一个偏不让。\n青龙洞里龙翻身，\n千年大梦变稻香！\n\n[Bridge]\n麻婆婆的狗，咬破麻叉口，\n麻线穿针眼，补丁也风流。\n左一句，右一句，\n舌头打结心自由！\n\n[Chorus]\n丁丁当当，乒乓乓乓，\n话赶话，舌绕梁，\n东边的钉，西边的墙，\n绕不完的弯，唱不完的慌！"
+    "caption": "A romantic Latin pop track opens with a clean, arpeggiated piano melody that forms the song's harmonic core. A heartfelt male vocal enters, singing in Spanish with an earnest and emotional delivery. The arrangement builds as a steady, mid-tempo reggaeton-style drum machine beat and a subtle synth bassline kick in, transforming the ballad into a danceable yet tender piece. The chorus is memorable and features layered vocal harmonies for emphasis. The track includes an instrumental break highlighting the piano and synth melody before concluding with a final chorus and a brief, atmospheric piano-led outro.",
+    "lyrics": "[Intro: Piano Melody]\n\n[Verse 1]\nMe enamoré de ti\nPorque tú eres la princesa\nY yo un príncipe soñando con tu belleza\nY ya no sé cómo puedo explicarlo\nPero ya la vida me trajo\nA tu lado, me siento orgulloso de ser el caballo más enamorado\n\n[Pre-Chorus]\nPero si yo pudiera retroceder el tiempo\nMe acercaría más a ti\n\n[Chorus]\nPero, oh-oh-oh\nPero, oh-oh-oh\nSi yo pudiera retroceder el tiempo\nMe acercaría más a ti\nPero, oh-oh-oh\nPero, oh-oh-oh\n\n[Verse 2]\nTú te quedaste para siempre en mi mente\nSi te miro a los ojos, siento todo lo diferente\nPorque contigo se detiene el tiempo\nSolo tú y yo sabemos el momento\nMe gustaría detener el tiempo\nAbrazarte, besarte, bailarte lento\nTu mirada me atrapa, es un cuento\nQue aún no termina, no, no\n\n[Bridge]\nEs imposible\nQue yo te quiera más\nY es increíble\nQue yo te quiera más\nY es increíble\nQue yo te quiera más\nMe haces falta\nTe amo y duele tanto, mi vida\n\n[Chorus]\nPero si yo pudiera retroceder el tiempo\nMe acercaría más a ti\nPero, oh-oh-oh\nPero, oh-oh-oh\nPero si yo pudiera retroceder el tiempo\nMe acercaría más a ti\nPero, oh-oh-oh\nPero, oh-oh-oh\n\n[Outro: Instrumental with Piano and Synth Pads]\n[Song fades out]",
+    "bpm": 100,
+    "duration": 150,
+    "keyscale": "G major",
+    "language": "es",
+    "timesignature": "4"
 }

examples/text2music/example_34.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "东北话, spoken word, male, Tempo - Fast, Elements - Chorus Hook, Subgenre-Satirical Hip Hop, Rap, Chinese-language music, energetic, slightly nasal, Instrument - Live Bass Guitar, adult, Vocals - Syncopated Flow, Genre - Hip-Hop, rapping, bright",
-    "lyrics": "[verse]\n挣着 憋屈的 工资 还得 装乐呵\n猫着 怂样儿 还搁 朋友圈 嘚瑟\n扛着 傻逼的 指标 没人 搭把手\n这儿 不是 托儿所 少整 那出儿 哭唧尿嚎\n\n俺们 就像 一条条 老板的 裤衩子\n陪着 笑脸 接他 每一回 突突\n哎呦 老板 今儿个 穿我呗\n他 撅个腚 眼角 瞟你 那熊样\n\n[chorus]\n他们 骂我 打工仔 太多人 没睡醒\n寻思 抠搜 老板 一天天 穷折腾\n不想 俺的 人生 烂在 这嘎达\n不想 俺的 将来 折在 这破棚\n\n老子 不想 上班 老子 是外星人\n你都 把俺 骂急眼了 俺还 这么淡定\n现实 才是 梦 啥时候 能醒啊\n那 糟践人的 答案 在西北风 里飘\n\n[verse]\n瞅见 二愣子 同事 给老板 舔腚沟子\n瞅见 浪蹄子 女同事 在老板 胯骨轴 扭搭\n瞅见 白瞎的 光阴 耗在 没亮儿的 道儿\n瞅见 公交车上 一帮 僵尸 吐酸水\n\n瞅见 俺的 命 定在 苦逼的 坑里\n瞅见 俺的 爱情 被轮了 成了 老处女\n瞅见 好事儿 全归 高富帅\n还有 那些 臭不要脸 扭腚的 货色\n\n[chorus]（重复）\n他们 骂我 打工仔 太多人 没睡醒...\n\n[bridge]\n加班 没补助 俺认了\n欠薪 揍员工 把俺 当牲口\n去你妈 的小姘头\n\n[verse]\n破逼 管理制度 净整 娱乐八卦\n撸管式 管理 也就 你自己 嗨\n出点儿 屁事儿 就往 下属 脑瓜子 扣\n挣俩 钢镚儿 立马 牛逼 不分 公母\n\n你挖个 大坑 把俺们 往里 踹\n说这 叫梦想 你当年 多能耐\n俺们 就当 听传销 洗脑课\n可怜 连骗人 你都 就会 这一套\n\n[outro]\n老子 不想 上班\n老子 不想 上班\n老子 不想 上班"
+    "caption": "An explosive modern metal track driven by high-gain, palm-muted guitar riffs and relentless double-bass drumming. The song kicks off with an aggressive, chugging guitar intro that sets a powerful, energetic tone. A clean, forceful male vocal enters, delivering an anthemic melody over the driving rhythm section. The arrangement features dynamic shifts, including a melodic and technical guitar solo section with harmonized leads. The production is polished and punchy, emphasizing the tight, powerful interplay between the dual guitars and the rhythm section, creating a sound reminiscent of high-energy anime theme music or J-rock fusedic metal.",
+    "lyrics": "[Intro - Guitar Riff]\n\n[Verse 1]\nMain selfie, bukan main kejar\nKata-kata manis jadi gelombang besar\nInsta buat teman, semua terkata\nTapi Wi-Fi raya tarik ke mana?\n\n[Chorus]\nMain dengan lowkey, buat hidup terasa\nHarga naik, story mahal punya\nBicara macam maya, suka duka\nTapi cuma idea buat dunia\n\n[Guitar Solo]\n\n[Verse 2]\nScroll TikTok semua berisi hiburan\nCaption kejar hilang, semua tumpang keban\nTapi lapar jalan gelak ketawa\nIngat kenapa setiap malam termenung rasa\n\n[Chorus]\nMain dengan lowkey, buat hidup terasa\nHarga naik, story mahal punya\nBicara macam maya, suka duka\nTapi cuma idea buat dunia\n\n[Bridge]\nPandai dengan semua orang\nTapi tahu di mana damaikan\nRekam-mu belaka, sini hulur tangan\nWalau dunia ini penuh cabaran\n\n[Guitar Solo]\n\n[Chorus]\nMain dengan lowkey, buat hidup terasa\nHarga naik, story mahal punya\nBicara macam maya, suka duka\nTapi cuma idea buat dunia\n\n[Outro - Guitar Solo]",
+    "bpm": 100,
+    "duration": 218,
+    "keyscale": "B♭ minor",
+    "language": "ms",
+    "timesignature": "4"
 }

examples/text2music/example_35.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Rap, J-Pop, Anime, kawaii pop, EDM, Aggressive, Intense, Crisp Snare, Super Fast, Clear",
-    "lyrics": "[Intro]\nNya.\n\n[Verse]\n我 在 五 点 二 十 早 起，十 三 点 十 四 弹 会儿 琴\n习 惯 了 坐 班，习惯了 隔夜 的  剩 饭，\n习 惯 了 没有 你\n\n[Verse]\n怕 你 想 不 开，拦 在 你 的 面 前\n那 时 候 摔 得 差 点 住 院\n东 京 的 春 天 莺 莺 燕 燕\n我 说 想 不 想 来 跟 我 玩 音乐\n\n[Verse]\n带 着 我 的 朋 友 守 在 你 的 门 口\n弹 着 我 的 钢 琴 当 伴 奏\n等 你 放 学 后，陪 你   K   T   V\n端 着 我 的 红 茶 跟 你 碰 杯\n\n[Pre-Chorus]\n忽然间现实淹没了远方\n万家灯火，盖住月光\n奔走，忍受，变成了人偶\n别再对我伸出你的 双 手，会 受 伤\n\n[Chorus]\n明明都向前走，方向却渐渐不同\n时间让你我越走越近，却越来越陌生\n春 天 在 滂 沱 的 大 雨 里 飘 落\n得 了 心 太 高 脸 太 薄 的病\n\n[Bridge]\n我越难过，春日影越顶\n眼泪晃得我看不清\n埋葬了懦弱还有矫情\n却还是会在半夜摸眼睛\n\n青春期大部分时间在工 作\n用微笑换来余额几个零\n戴上了面具也明白了生活\n拼的是数字和脸更是命\n\n[Verse]\n我在五点二十早起，十三点十四弹会琴\n早上要做饭，回家时满地的瓶罐\n\n师 徒 二 人 站 在 我 的 面 前\n台 词 很 熟 练，照 着 就 念\n\n背 后 的 小 睦 扭 扭 捏 捏\n我 说 我 还 有 点 事 要 不 改 天 见\n\n然 后 你 的 双手 握 住 我 的 袖 口\n开 始 哭 着 求 我 不 要 走\n\n[Verse]\n我在下班后，忙活柴米油\n你和你的姐妹住着高楼\n\n苦 来 兮 苦，早 就 没 了\n现 实 扬 鞭，赶 着 我 向 前\n没有时间跟你分辨什么对与错\n\n[Bridge]\n没有什么对错，没有罪过\n谁不曾天真，是我太早看破\n生活一片狼藉，却又不想放弃\n一 边 聚 光 灯 下 绽 放，一 边 坠 落\n故作坚强，筑起心的墙\n越是委屈的伤口，越要藏\nLet it all out， it’s all right\n\n[Outro]\n俺 是 东 京 嘞，东 京 打 工 妹\n\n从虎之门带你转到浅草\n再从新宿转到竹桥\n\n俺 是 东 京 嘞，东 京 打 工 妹\n\n带 你 转 羽田 成田 蒲田 神田\n做 你 嘞 小 甜 甜！\n\n俺 是 东 京 嘞，东 京 打 工 妹\n带 你 转 赤 坂，带 你 转 霞 关\n恁 咋 不 早 说，今 天 不 管 饭\n"
+    "caption": "An energetic, bilingual pop-rap track built on a foundation of punchy trap drums and a deep sub-bass. The song's signature element is a highly catchy, synthesized woodwind melody with a distinct Middle Eastern or Balkan flavor, which serves as the main hook. A male vocalist delivers confident, rhythmic rap verses, seamlessly switching between English and French. The production is clean and modern, with hype-man ad-libs and vocal shouts punctuating the arrangement, creating a vibrant, club-ready atmosphere perfect for a night out.",
+    "lyrics": "[Intro]\n[Synth melody intro]\nYo, c'est Big Boss sur le beat\nMaxi sans détour\nOMG, hey baby, viens goûter au détour\nOn lève les bras, tourne le ciel\nLaisse l'énergie nous appeler\nLaisse-moi chauffer ton corps\nDécollage, gars!\n\n[Instrumental Drop]\n\n[Verse 1]\nSway, DJ, pousse ton petto\nLes basses cognent, c'est trop chaud\nLa vibe est dans ma tête, pas de règles ni de stop\nJe suis déjà hors de contrôle, ready to drop\nShort lu, ta playlist, elle déchire le trône\nJ'essaie de conduire mais la passion m'emprisonne\nLa vibe est divine, dansons sans limite\nGarde la vitesse, ce soir c'est la frénésie\n\n[Chorus]\nOn est les boss, on chauffe sans hésitation\nMaxi, Silan, Moustique dans la mission\nJusque la nuit, fais-moi vibrer\nTant que la musique monte, je ne peux plus freiner\n\n[Instrumental Drop]\n\n[Verse 2]\nOn est les boss, on chauffe sans hésitation\nMaxi, Silan, Moustique dans la mission\nJusque la nuit, fais-moi vibrer\nTant que la musique monte, je ne peux plus freiner\nDes ladies all in, te fait sourire en public\nJe suis prêt pour l'action, pas pour te faire diviser\nLa peau douce, ton corps contre le mien\nOn se donne à fond, plus rien ne nous retient\nJe t'invite à la danse, je suis ton vestin\nQuand tu danses, c'est le moment, laisse-toi porter par le refrain\nOn est là pour le show, pas besoin de mots\nJe t'emmène au-delà de ce tempo\nDanse, balance, oublie tout, on brille ensemble\nLa nuit est jeune, l'énergie nous rassemble\nAlors viens, collé-serré, laisse-toi porter\nBig Boss, on est là pour tout casser\n\n[Instrumental Drop]\n\n[Outro]\n[Synth melody fades out]",
+    "bpm": 100,
+    "duration": 135,
+    "keyscale": "G minor",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_36.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Rap, adult, male, spoken word, rapping, clear, warm, articulate, Lo-Fi Hip Hop, 100-120 BPM, Keyboard Chords, Male Rap, Lazy Rhythm, Melancholy, Rap",
-    "lyrics": "[Intro]\n夜色 很 淡 像 褪色 的 照片  \n但 记忆 却 像 刀锋 一样 锐利  \n\n[Verse 1]\n你 说过 的 甜言蜜语 现在 听来 像 最 恶毒 的 咒骂  \n你 刺进 我 心里 的 刀 现在 还 在 滴血 未 干 哪  \n慵懒 的 旋律 像 我 的 脚步 拖着 沉重 的 躯壳  \n脑海 里 循环 播放 那 画面 快 把 我 逼疯 了  \n键盘 和弦 低沉 又 忧伤 弹奏 着 我 的 绝望  \n我 曾经 的 信任 像 玻璃 一样 被 你 狠狠 地 摔 在 地上  \n不想 振作 不想 原谅 只 想 让 这 一切 都 停止  \n可 心底 有 个 声音 嘶吼 着 要 你 付出 该 有 的 代价  \n\n[Chorus]\n背叛 像 毒药 渗透 我 的 血液  \n复仇 的 火焰 在 我 眼中 燃起  \n哪怕 遍体鳞伤 哪怕 万劫不复  \n我 也 要 亲手 撕碎 你 的 幸福  \n这 是 我 的 哀歌 也 是 我 的 战书  \n键盘 的 音符 每 一下 都 带着 恨意 和 痛苦  \n\n[Verse 2]\n曾经 的 兄弟 现在 面目全非 像 个 陌生人  \n你 的 自私 像 癌细胞 一点点 吞噬 我 的 纯真  \n我 学着 你 的 样子 把 心 锁 起来 不再 轻易 相信  \n让 懒散 的 节奏 包裹 我 给 自己 一点 喘息  \n键盘 的 音色 变得 更加 阴冷 像 秋天 的 雨滴  \n冲刷 掉 所有 温情 只 剩下 彻骨 的 寒意  \n我 不会 大喊大叫 只是 默默 地 计划  \n每 一步 都 走向 让 你 后悔 的 那 一 刹那  \n\n[Chorus]\n背叛 像 毒药 渗透 我 的 血液  \n复仇 的 火焰 在 我 眼中 燃起  \n哪怕 遍体鳞伤 哪怕 万劫不复  \n我 也 要 亲手 撕碎 你 的 幸福  \n这 是 我 的 哀歌 也 是 我 的 战书  \n键盘 的 音符 每 一下 都 带着 恨意 和 痛苦  \n\n[Bridge]\n也许 复仇 不能 带来 平静  \n也许 只 会 让 我 更 堕落  \n但 如果 不 这样 做  \n我 连 活下去 的 勇气 都 没有  \n\n[Outro]\n复仇 复仇 复仇  \n直到 最后 一刻  \n懒散 地 复仇 着  "
+    "caption": "An intense, high-energy industrial metal track driven by a powerful, pounding drum machine and chugging, heavily distorted rhythm guitars. The song opens with a tense, cinematic synth arpeggio before erupting into a full-throttle verse. A commanding female lead vocal soars over the dense instrumentation, delivering a dramatic and powerful performance. The arrangement features a dynamic shift into a spoken-word bridge with a menacing, processed male vocal, adding to the dark, narrative atmosphere. The track culminates in a massive, anthemic chorus with layered vocals and a final, screamed vocal flourish before fading out with atmospheric synths and a lingering guitar motif.",
+    "lyrics": "[Intro - Synth Arpeggio & Driving Beat]\n\n[Verse 1]\nSous les cieux de la nuit, elle se lève\nLes mains caressant l'épaule en alerte\nSes doigts effleurent les vêtements qui comptent\nSon rire est une promesse qu'il comprendra\n\n[Pre-Chorus]\nElle est parée de rouge et de sang\nForte, sans force, infinie, forte\nSes épaules sont chaotiques, liées par la réalité\nL'écho de son corps gronde, la pression qui monte\n\n[Chorus]\nLa sorcière au caoutchouc se couvre des habits vêtements\nDevaient emporter les fils de son corps\nElle brandit sa lame tendue et ses yeux ne brillent plus\nQuand elle embrasse les ténèbres de son ombre\n\n[Instrumental Break]\n\n[Verse 2]\nLes chasseurs se battent dans l'ombre, mais elle ne chancelle pas\nSa sagesse est une bataille et la menace est son instinct\nElle évoque les légendes des plus terres\nElle aime la nuit, la transpirance infinie\nL'attirance, la puissance qui les hante\n\n[Instrumental Break]\n\n[Bridge - Male Spoken Word, Filtered]\nLes indices se cachent, les espoirs s'effacent\nElle est l'empereur de son château sacré\nSon père est une victime\nElle règne à sa loi\nDe la sorcière au caoutchouc, les cris de son esprit\n\n[Chorus]\nLa sorcière au caoutchouc, la danse du monde se nourrit\nDe ses coups de poing plus fin, de son rire plus fatal\nDans ce chaos de la nuit, elle est le chant de la mort\nLa sorcière au caoutchouc, la mort, la mort, la mort\n\n[Instrumental Outro with Vocal Ad-libs]\n(Ahhh-ah-ah-ahhh)\n(Ahhh-ah-ah-ahhh)\n\n[Final Verse - Layered Vocals]\nElle est parée de rouge et de sang\nForte, sans force, infinie, forte\nElle est parée de rouge et de sang\nForte, sans force, infinie, forte\nElle est parée de rouge et de sang\nForte, sans force, infinie, forte\nElle est parée de rouge et de sang\nForte, sans force, infinie, forte\nForte!\n\n[Outro - Male Spoken Word, Filtered]\nLa sorcière au caoutchouc, la mort, la mort, la mort\n[Song ends abruptly]",
+    "bpm": 100,
+    "duration": 252,
+    "keyscale": "E minor",
+    "language": "fr",
+    "timesignature": "4"
 }

examples/text2music/example_37.json

@@ -1,5 +1,10 @@
 {
     "think": true,
-    "caption": "Orchestra, Symphony, Sonata, Opera, Concerto, Rap, Beat, DJ, MC, StreetCulture",
-    "lyrics": "[verse1]\n羊皮卷轴 墨香飘 莫扎特 熬 安魂曲 通宵  \n和弦齿轮 咔哒转 比 瑞士 手表 更 精密 律动  \n八轨磁带 玩叠叠乐 披头士 炸 录音棚 天花板  \nAI 卷起 新风暴 像 灭霸 打响指 般 简单  \n\n[chorus]\n琴弦 到 代码 进化论 狂飙（skr）  \n象牙塔 被 鼠标 点爆 像 泡泡（boom）  \n灵感 加 算法 等于 王炸 大招  \n人类 心跳 才是 终极 混音 调料  \n\n[verse2]\n春之祭 召唤 百人 乐团 才够 燥  \n合成器 极客 玩电焊 焊出 赛博 神庙  \nDAW 解放 双手 钢琴卷帘 变 乐高  \n音色库 开挂 像 吃 金币 的 马里奥  \n\nAI 拆解 爵士乐 黑话 像 庖丁 解牛  \nCityPop 复古 滤镜 直接 参数 调油  \n神经网络 偷师 贝多芬 半夜 翻墙头  \n音乐 基因库 被 改写成 超频 万花筒  \n\n[chorus]  \n琴弦 到 代码 进化论 狂飙（skr）  \n象牙塔 被 鼠标 点爆 像 泡泡（boom）  \n灵感 加 算法 等于 王炸 大招  \n人类 心跳 才是 终极 混音 调料  \n\n[verse3]  \n电子琴 被 吐槽 塑料 味 超标  \n卧室 制作人 用 鼠标 单挑 整个 乐团 编制  \nAI 伴奏 刚上线 就被 键盘侠 集火  \n却 忘了 电吉他 曾被 说 是 魔鬼 的 副歌  \n\n现在 我 指尖 蹦迪 在 数据 炼丹炉  \n提示词 召唤 莫扎特 跨次元 碰杯 珍珠奶茶  \n当 比特 海洋 淹没 所有 物理 琴柱  \n最后 的 音轨 永远 连着 心脏 的 跳针  \n\n[bridge]  \n鹅毛笔 蘸着 银河 当 墨汁（绝了）  \n音浪 在 元宇宙 开 分店（疯了）  \n技术 迷雾 散成 像素 烟花  \n而 我们 始终 带着 老派 的 心跳 混搭  \n\n[chorus]  \n琴弦 到 代码 进化论 狂飙（skr）  \n象牙塔 被 鼠标 点爆 像 泡泡（boom）  \n灵感 加 算法 等于 王炸 大招  \n人类 心跳 才是 终极 混音 调料  \n\n[outro]  \n从 蒸汽 到 硅基 浪潮 我 冲浪（yo）  \n用 脑洞 接住 每个 技术 暴击（叮）  \n当 所有 设备 没电 的 凌晨 三点钟  \n最 原始 的 旋律 在 胸腔 敲击 成 龙卷风  "
+    "caption": "A classic blues-rock arrangement with a soulful, nocturnal groove. The track is built on a tight rhythm section of a round-toned bass guitar and a crisp acoustic drum kit. A clean, slightly overdriven electric guitar provides bluesy licks and chordal stabs throughout the verses. The lead male vocal is smooth and soulful with a touch of rasp, soaring into layered harmonies during the expansive chorus. The song features a melodic and expressive guitar solo with classic bends and vibrato, followed by a breakdown and a final, stripped-down vocal section that fades into a concluding guitar flourish.",
+    "lyrics": "[Intro - Guitar Riff]\n\n[Verse 1]\nCity lights, they dance for miles\nNeon dreams, they hum in style\nLate night tales in smoky bars\nWishing on these falling stars\n\n[Verse 2]\nFootsteps echo on the street\nMelodies in perfect beat\nLonely hearts with hidden scars\nStill we chase these falling stars\n\n[Chorus]\nOh, let's take this ride tonight\nThrough the shadows in the night\nFeel the rhythm of guitars\nLost in wishes, flying far\n\n[Verse 3]\nIn a world of glass and steel\nBroken dreams feel so unreal\nYet we sing beneath the stars\nWishing on these falling stars\n\n[Guitar Solo]\n\n[Bridge]\nBaselines thumping through the dark\nPianos playing in the park\nEvery note a shot of heart\nEvery whisper, falling stars\n\n[Chorus]\nOh, let's take this ride tonight\nThrough the shadows in the night\nFeel the rhythm of guitars\nLost in wishes, flying far\n\n[Outro]\nIn a world of glass and steel\nBroken dreams feel so unreal\nYet we sing beneath the stars\nWishing on these falling stars\n\n[Extended Guitar Solo and Outro]\n[Song fades out]",
+    "bpm": 100,
+    "duration": 235,
+    "keyscale": "G minor",
+    "language": "en",
+    "timesignature": "4"
 }

examples/text2music/example_38.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "A driving post-punk arrangement kicks off with layered electric guitars—one clean and arpeggiated, the other providing distorted chordal texture—over a solid bassline and powerful live drums. The male lead vocal is delivered with an angsty, strained quality that builds into an anthemic, shouted chorus where his voice cracks with emotion. Following a melodic yet noisy guitar solo filled with feedback and expressive bends, the track breaks down to its core rhythmic elements before fading out on lingering guitar noise.",
+    "lyrics": "[Intro - Arpeggiated Electric Guitar]\n\n[Verse 1]\nUnder neon lights, they flicker and fade\nLost in the hum of this restless parade\nA city's heartbeat pulses in my veins\nSearching for answers in the acid rain\nShadows stretch long, they're clawing my way\nThe night is alive but it's mine to obey\n\n[Chorus]\nOh, shadows of neon\nWhere do you run?\nBleeding out the night just to make us one\nOh, shadows of neon\nTell me the truth\nAre you the ghost or the one I chase with you?\n\n[Guitar Solo]\n\n[Verse 2]\nThe walls are whispering secrets they can't keep\nEchoes of footsteps buried six feet deep\nEvery corner's a story you'll never forget\nI'm tangled in chaos, dressed up as regret\n\n[Chorus]\nOh, shadows of neon\nWhere do you run?\nBleeding out the night just to make us one\nOh, shadows of neon\nTell me the truth\nAre you the ghost or the one I chase with you?\n\n[Outro - Extended Guitar Solo]\n[Music fades out]",
+    "bpm": 100,
+    "duration": 178,
+    "keyscale": "A minor",
+    "language": "en",
+    "timesignature": "4"
+}

examples/text2music/example_39.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An explosive fusion of chiptune and punk rock, this track erupts with a high-energy, 8-bit synth melody playing over a driving, distorted power-chord guitar riff and a punchy, tight drum beat. The lead male vocal is delivered with a clear, energetic, and slightly nasal punk-rock inflection. The song maintains a relentless pace, structured around catchy, anthemic choruses and verses that playfully list various types of coffee with satiric words. The chiptune synth hook serves as a recurring motif, reinforcing the track's video game-inspired sonic identity before ending abruptly after a final guitar flourish.",
+    "lyrics": "[Intro - Chiptune Synth Melody]\n\n[Verse 1]\nVos p'tits mots le pes, c'est pour me réconforter\nJe rentre chez moi avec rien à boire\nOn fait caca du pain chaud sur le lait qui grêle\nOn fume du shit sur le clocher des pommes\nOn oublie de nettoyer le four\n\n[Pre-Chorus]\nVos p'tits mots le pes, c'est le vin qui me délecte\nL'assiette\n'est plus un médicament\nOn danse au musée avec des pieds en terrasse\nUn Ricard qui tourne à plein poumons\nOn boit un verre pour oublier\n\n[Chorus]\nJ'aimerais mieux, j'ai pas la cocaïne\nQu'on ferme les yeux pour être tous ensemble\nLes wokos sont tristes, la bière est toxique\nLes poches sont vides mais vous y pensez\nJ'aimerais mieux, j'ai pas la cocaïne\nQu'on ferme les yeux pour être tous ensemble\nLes wokos sont tristes, la bière est toxique\nLes poches sont vides mais vous y pensez\n\n[Instrumental Break - Chiptune Synth Melody]\n\n[Verse 2]\nVos p'tits mots le pes, c'est pour me rassasier\nOn sort en berne avec des céréales\nOn boit pour oublier le cocaïne qu'on boit\nOn oublie de saluer le pot de bon matin\nLa bise aux potes, le rhum, le rhum\n\n[Pre-Chorus]\nVos p'tits mots le pes, c'est pour m'amender le cocaïne\nLa table en grand avec du café\nOn rit aux larmes quand on a rien à perdre\nOn s'amuse dans le coin de la cuisine\nÀ faire de la merde avec des souris\n\n[Chorus]\nJ'aimerais mieux, j'ai pas la cocaïne\nQu'on ferme les yeux pour être tous ensemble\nLes wokos sont tristes, la bière est toxique\nLes poches sont vides mais vous y pensez\nJ'aimerais mieux, j'ai pas la cocaïne\nQu'on ferme les yeux pour être tous ensemble\nLes wokos sont tristes, la bière est toxique\nLes poches sont vides mais vous y pensez\n\n[Outro - Chiptune Synth Melody]\n[abrupt silence]",
+    "bpm": 100,
+    "duration": 190,
+    "keyscale": "E major",
+    "language": "fr",
+    "timesignature": "2"
+}

examples/text2music/example_40.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "A driving indie rock track with dream pop undertones, built on a steady drum machine beat and an active bassline that anchors the song's momentum. Clean electric guitars weave intricate arpeggios throughout the verses; one plays rhythmic chords while another adds atmospheric lead lines soaked in delay during instrumental breaks. A clear female vocal delivers a melancholic melody in Polish, soaring into layered harmonies for the expansive chorus sections. The arrangement features dynamic shifts between sparser verses and fuller choruses, culminating in an extended outro where melodic guitar leads soar over wordless 'ah' vocals before fading out to a final sustained chord.",
+    "lyrics": "[Intro - Instrumental]\n\n[Verse 1]\nChciałbym zatrzymać czas\nGdy zatapiam się w tobie\nByleby nie palić się\nA byłem w niebie\n\n[Chorus]\nBo twoje oczy\nA ja tworzę świat\nPrzez ciebie\nNiebo niebo\nI światło jest we mnie\nBo jesteś tu\n\n[Guitar Solo]\n\n[Verse 2]\nKiedy widzę z daleka twoją twarz\nNic już nie muszę cieszyć się\nRozczaruję twoją ciszę\nI pragnę tej chwili\n\n[Chorus]\nBo twoje oczy\nA ja tworzę świat\nPrzez ciebie\nNiebo niebo\nI światło jest we mnie\nBo jesteś tu\n\n[Instrumental Bridge with Guitar Solo]\n\n[Outro]\n[Scat singing]\nAhhh...\nAhhh...\nAhhh...\nAhhh...\n[Song fades out]",
+    "bpm": 100,
+    "duration": 194,
+    "keyscale": "D major",
+    "language": "pl",
+    "timesignature": "4"
+}

examples/text2music/example_41.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An energetic and quirky funk-rock track driven by a tight, groovy bassline and a punchy drum machine beat. The song opens with a catchy, scatted vocal hook that sets a playful tone. The lead male vocal is delivered in a rapid-fire, almost rap-like cadence during the verses, shifting to a more melodic and anthemic style for the chorus. The arrangement is punctuated by bright, synthesized brass stabs and a lively saxophone and trumpet-led instrumental break. The track breaks down into a more contemplative section with clean electric guitar before rebuilding to a high-energy finale that fades out with the initial scat motif.",
+    "lyrics": "[Intro]\n[Scatting vocal sample]\nLig na...\nLig na...\nLig na...\nLig na...\n\n[Verse 1]\nCzasem chciałbym woda na prysznic\nNa lajczak zawsze zemykać\nMyśli, że mam złe wieści\nŻe to za mała kawiatura\nA może niedziela, mam chwilę później\nWtedy to poszło nie tak\nMój ziomal umyślny stek w tym cyfrowym mieście\nZa oknem jeszcze\n\n[Pre-Chorus]\nDzień jak co dzień w kaploferze\nPod goliszkiem jak na różowo\nJak już obchodzi, co jutro się dzieje\nI co nam znowu przyszło\n\n[Chorus]\nŻycie to nie te dni, ani te dni\nCzyli zatem trzy, do tego trochę chwili\nLecz jestem tym zimnym i zimnym\nZimnym i zimnym\nBo w zimie jest coś, co ma zapalić\nCztery, podpisane pod jajami\nZimy, to chyba nie obchodzi nas\nI tu każda parada jest zimna jak twoja krew\n\n[Instrumental Break with Synth Brass]\n\n[Verse 2]\nLecimy dalej\nNa celownik wziął\nA więc rano\nPod szkołą pod osłoną nocy\nPłacąc za stare nieco\nWszystko się kończy, a my możemy się przejąć\nZegarek kolejny leci\nMiasto żyje naprawdę\nTo jeszcze nie koniec\nTo jeszcze nie koniec\nNie\n\n[Bridge]\nA co to za pech, że wiecznie upiętne dni mijają\nI mimo, że już za późno\nDziś, jutro, niezniszczalny\nJuż nie zniknął\n\n[Instrumental Break with Synth Brass]\n\n[Outro]\n[Scatting vocal sample]\n[Synth brass melody]\n[Final chord]",
+    "bpm": 100,
+    "duration": 208,
+    "keyscale": "D minor",
+    "language": "pl",
+    "timesignature": "4"
+}

examples/text2music/example_42.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An explosive big band jazz arrangement kicks off with a powerful brass fanfare and a driving swing rhythm section featuring a walking upright bass and crisp ride cymbals. A commanding female vocalist enters with a theatrical, powerful delivery, her voice soaring over the dynamic horn stabs and saxophone lines. The track includes an energetic, improvisational saxophone solo over the tight rhythm section. The arrangement is dynamic, shifting between full-band shouts and more subdued passages, culminating in a dramatic final vocal statement and a classic big band flourish to finish.",
+    "lyrics": "[Intro - Full Band Fanfare]\n\n[Verse 1]\nKani\nKuni no kodoku ni omoi dashi\nKuni no yasuraki mo yoyotsu mo nashi\nTsuranuku utagoe yo\nIshisakiyo moeyo\nKanawanu negai o\n\n[Pre-Chorus]\nKujike no sakeba\nKono yo no owari\nKodoku no naka de\nKasumi ni ikiru\n\n[Chorus]\nTsuranuku utagoe yo\nKanawanu negai o\nKujike no sakeba\nKono yo no owari\nKodoku no naka de\nKasumi ni ikiru\n\n[Instrumental Break - Saxophone Solo]\n\n[Bridge]\nKujike no sakeba\nKono yo no owari\nKodoku no naka de\nKasumi ni ikiru\n\n[Chorus]\nTsuranuku utagoe yo\nKanawanu negai o\nKujike no sakeba\nKono yo no owari\nKodoku no naka de\nKasumi ni ikiru\n\n[Outro - Full Band Crescendo and Final Chord]",
+    "bpm": 100,
+    "duration": 175,
+    "keyscale": "D minor",
+    "language": "ja",
+    "timesignature": "2"
+}

examples/text2music/example_43.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An energetic and playful children's pop-rock song with a driving, upbeat tempo. The track is built on a foundation of a punchy drum machine beat and a clean electric bassline. Distorted electric guitars play catchy power-chord riffs, giving the chorus a rock edge. A clear, enthusiastic male lead vocal delivers the lyrics with a sing-along quality, supported by layered backing vocals in the chorus. The arrangement includes an instrumental break featuring a synth lead melody and a brief, stripped-down bridge before launching back into the high-energy chorus and a final guitar-led outro.",
+    "lyrics": "[Intro]\nHé, csávó!\nMicsoda csávó, a főnök királyfi, figyelj jól!\n[Verse 1]\nPörög a főnix, pörög a táj\nMindenki nevet, ez nem vitás\nCsini, Miki, vidám sztorikat\nA kocsma így rogy, a tűz benned ragad\n[Chorus]\nHé, hé, kicsi Miki, gyere vissza hajnalig\nTámadjon hát, és szórja a vizet\nHé, hé, Miki Miki, hajnalig jön, de\nNem áll meg senki, jöhet a tánc, ne késlekedj el!\n[Verse 2]\nA kocsma illatát mesteri a nevetés\nMiki meg a csini is énekel, ez az egész éves élmény\nPoci, csiki, pörög a rúzs\nMindenki vele, ezt sose fogja\n[Chorus]\nHé, hé, kicsi Miki, gyere vissza hajnalig\nTámadjon hát, és szórja a vizet\nHé, hé, Miki Miki, hajnalig jön, de\nNem áll meg senki, jöhet a tánc, ne késlekedj el!\n[Bridge]\nHé, hé, há, te Miki Miki, mesteri\nEzt mi sose hagynád\nÓ, micsoda, micsoda, Miki Miki, gyere vissza, gyere vissza!\n[Instrumental Break - Synth Solo]\n[Verse 3]\nA kocsma illatát mesteri a nevetés\nMiki meg a csini is énekel, ez az egész éves élmény\nPoci, csiki, pörög a rúzs\nMindenki vele, ezt sose fogja\n[Chorus]\nHé, hé, kicsi Miki, gyere vissza hajnalig\nTámadjon hát, és szórja a vizet\nHé, hé, Miki Miki, hajnalig jön, de\nNem áll meg senki, jöhet a tánc, ne késlekedj el!\n[Outro]\nHé, hih, teli a kocsma\nMiki Miki, Miki Miki, az öröm\nÓ, micsoda, micsoda, Miki Miki\nGyere vissza, gyere vissza!\n[Instrumental Outro with Synth Solo]",
+    "bpm": 100,
+    "duration": 181,
+    "keyscale": "C major",
+    "language": "hu",
+    "timesignature": "4"
+}

examples/text2music/example_44.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "A smooth, contemporary R&B track built on a foundation of clean, melodic electric guitar and a steady, unobtrusive drum machine beat. The song opens with a catchy, wordless vocal hook that sets a romantic, nighttime city mood. A clear male tenor vocal, delivered in a mix of Mandarin and English, glides over the arrangement, supported by lush, layered harmonies and echoed ad-libs that add depth and space to the mix. The production is polished and modern, with a focus on creating an immersive, heartfelt atmosphere through tasteful reverb and a well-balanced low end. The track concludes with a brief, melodic guitar solo that fades out, reinforcing its soulful character.",
+    "lyrics": "[Intro]\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n\n[Verse  1]\n[zh] ni2 hong2 deng1 xia4 wo3 men5 man4 bu4 zai4 ye4 wan3\n[en] (ooh-ooh-ooh-ooh)\n\n[zh] ni3 de5 yan3 shen2 xiang4 xing1 chen2 zai4 wo3 xin1 shang4 shan3\n[en] (ooh-ooh-ooh-ooh)\n\n[zh] ye4 feng1 chui1 dong4 zhe5 jie1 tou2 de5 lang4 man4\n[en] (ooh-ooh-ooh-ooh)\n\n[zh] mei3 ge4 jiao3 luo4 dou1 you3 wei4 zhi1 de5 da2 an4\n[en] (ooh-ooh-ooh-ooh)\n\n[en] [chorus]\n\n[zh] ai4 zai4 yong1 ji3 de5 di4 fang1 jing4 jing4 su4 shuo1\n[zh] ni3 de5 wen1 rou2 rang4 wo3 mei2 ban4 fa3 duo3\n[zh] shi4 jie4 tai4 da4 que4 yin1 ni3 yi3 jing1 wo3\n[zh] wu2 lun4 duo1 yuan3 ye3 bu2 hui4 mi2 shi1 zi4 wo3\n\n\n[Post-Chorus]\n[zh] ni3 shi4 wo3 de5 deng1 ta3   zhi3 yin3 wo3 qian2 xing2\n[zh] sheng1 ming4 de5 lv3 tu2 yin1 ni3 bian4 de2 an1 ding4\n[zh] tian1 bian1 de5 yun2 duo3 jiu4 zai4 tou1 tou1 kan4 ni3\n[zh] zhe4 yi1 miao3 zhong1 zai4 wo3 xin1 li3 hua4 xuan2 lv4\n[en] (ooh-ooh-ooh-ooh)\n\n[en] [verse two]\n\n[zh] ni3 de5 shou3 zhi3 hua2 guo4 wo3 xin1 tiao4 de5 mei3\n[zh] mei3 ge4 zi4 jie2 xiang4 shi4 liu2 xing1 hua2 guo4 de5 hui1\n[zh] gu3 dian3 chen2 zhong4 dan4 shi4 yin1 ni3 zhong4 yue4\n[zh] xin1 tiao4 li3 shi4 ni3 ai4 de5 jie2 zou4 mei2 yan2\n[en] (ooh-ooh-ooh-ooh)\n\n[zh] ni3 de5 shou3 zhi3 hua2 guo4 wo3 xin1 tiao4 de5 mei3\n[zh] mei3 ge4 zi4 jie2 xiang4 shi4 liu2 xing1 hua2 guo4 de5 hui1\n[zh] gu3 dian3 chen2 zhong4 dan4 shi4 yin1 ni3 zhong4 yue4\n[zh] xin1 tiao4 li3 shi4 ni3 ai4 de5 jie2 zou4 mei2 yan2\n[en] (ooh-ooh-ooh-ooh)\n\n[en] [bridge]\n\n[zh] ai4 zai4 yong1 ji3 de5 di4 fang1 jing4 jing4 su4 shuo1\n[zh] ni3 de5 wen1 rou2 rang4 wo3 mei2 ban4 fa3 duo3\n[zh] shi4 jie4 tai4 da4 que4 yin1 ni3 yi3 jing1 wo3\n[zh] wu2 lun4 duo1 yuan3 ye3 bu2 hui4 mi2 shi1 zi4 wo3\n[zh] yin1 fu2 zai4 wo3 men5 zhi1 jian1 qing1 qing1 xuan2 zhuan3\n[zh] ling2 hun2 de5 xuan2 lv4 tiao4 dong4 liu2 xia4 shi1 pian1\n[zh] zhe4 yi1 shou3 ge1 de5 ge1 ci2 ni3 lai2 ting1 jian4\n[zh] ai4 zai4 yong1 ji3 zhong1 rong2 hua4   zai4 shun4 jian1\n[en] (ooh-ooh-ooh-ooh)\n\n[en] [chorus]\n\n[zh] ai4 zai4 yong1 ji3 de5 di4 fang1 jing4 jing4 su4 shuo1\n[zh] ni3 de5 wen1 rou2 rang4 wo3 mei2 ban4 fa3 duo3\n[zh] shi4 jie4 tai4 da4 que4 yin1 ni3 yi3 jing1 wo3\n[zh] wu2 lun4 duo1 yuan3 ye3 bu2 hui4 mi2 shi1 zi4 wo3\n\n\n[Outro]\n[Instrumental with vocal ad-libs]\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n[Music fades out]",
+    "bpm": 100,
+    "duration": 202,
+    "keyscale": "C# major",
+    "language": "zh",
+    "timesignature": "4"
+}

examples/text2music/example_45.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An explosive burst of high-energy J-rock and chiptune, the track kicks off with a blistering, harmonized guitar and synth arpeggio intro reminiscent of a video game boss battle. A powerful, clear female vocal drives the verses over a tight, funky slap bassline and a driving drum beat. The arrangement builds into an anthemic, soaring chorus with layered vocals and powerful guitar chords. The track is punctuated by two virtuosic, shred-style guitar solos filled with rapid-fire licks and technical flourishes. The song concludes with a final, triumphant chorus and a flurry of guitar pyrotechnics before an abrupt end.",
+    "lyrics": "[Intro - Guitar & Synth Lead]\n\n[Verse  1]\n[zh] ye4 feng1 fu2 guo4 wo3 de5 lian3\n[zh] ni2 hong2 xia4 de5 meng4 jing4 wu2 bian1\n[zh] cheng2 shi4 jie2 zou4 xiang4 xin1 tiao4 zai4 bian4\n[zh] mei3 ge4 jiao3 bu4 dai4 zhe5 xin1 de5 xuan2 nian4\n[zh] ni2 hong2 jian4 yue4 que4 kan4 bu4 qing1 lian3\n[zh] cang2 zhe5 huan1 xiao4 he2 wu2 sheng1 de5 zhai4\n[zh] wo3 gen1 sui2 zhe5 xin1 zhong1 de5 bo1 lan2\n[zh] yue4 guo4 fan2 hua2 zhui1 sui2 zhe5 guang1 huan2\n\n\n[Chorus]\n[zh] tiao4 jin4 ye4 se4 de5 yin2 he2 xuan2 lv4\n[zh] fan1 yue4 gu1 dan1 yu3 wei4 zhi1 de5 mi2\n[zh] xin1 zai4 fei1 xiang2   zhui1 zhu2 de5 ji4 yi4\n[zh] rang4 meng4 ru2 xing1 chen2 zai4 ye4 kong1 yan2 xu4\n\n\n[Guitar & Synth Solo]\n\n[Verse  2]\n[zh] mei3 shan4 chuang1 wai4 dou1 cang2 zhe5 hua4 yu3\n[zh] guo4 qu4 wei4 lai2 dou1 qiao3 ran2 kao4 jin4\n[zh] jiao3 bu4 wu2 sheng1 dan4 ji4 yi4 tou4 ming2\n[zh] zai4 chen2 mo4 zhong1   zhao3 dao4 xin1 de5 su1 xing3\n[zh] shui2 shuo1 meng4 jing4 zong3 shi4 bu4 ming2 bai2\n[zh] zhui1 zhu2 xing1 chen2   yue4 guo4 zhe4 shan1 mai4\n[zh] yi1 pian4 guang1 ying3 jiang1 ling2 hun2 jie3 kai1\n[zh] ye4 wan3 dai4 zhe5 yi1 chang2 wu2 jin4 de5 ai4\n\n\n[Chorus]\n[zh] tiao4 jin4 ye4 se4 de5 yin2 he2 xuan2 lv4\n[zh] fan1 yue4 gu1 dan1 yu3 wei4 zhi1 de5 mi2\n[zh] xin1 zai4 fei1 xiang2   zhui1 zhu2 de5 ji4 yi4\n[zh] rang4 meng4 ru2 xing1 chen2 zai4 ye4 kong1 yan2 xu4\n\n\n[Guitar & Synth Solo]\n\n[Bridge]\n[zh] shui2 shuo1 meng4 jing4 zong3 shi4 bu4 ming2 bai2\n[zh] zhui1 zhu2 xing1 chen2   yue4 guo4 zhe4 shan1 mai4\n[zh] yi1 pian4 guang1 ying3 jiang1 ling2 hun2 jie3 kai1\n[zh] ye4 wan3 dai4 zhe5 yi1 chang2 wu2 jin4 de5 ai4\n\n\n[Chorus]\n[zh] tiao4 jin4 ye4 se4 de5 yin2 he2 xuan2 lv4\n[zh] fan1 yue4 gu1 dan1 yu3 wei4 zhi1 de5 mi2\n[zh] xin1 zai4 fei1 xiang2   zhui1 zhu2 de5 ji4 yi4\n[zh] rang4 meng4 ru2 xing1 chen2 zai4 ye4 kong1 yan2 xu4\n\n\n[Outro - Guitar & Synth Solo]\n[Song ends abruptly]",
+    "bpm": 40,
+    "duration": 224,
+    "keyscale": "E minor",
+    "language": "zh",
+    "timesignature": "2"
+}

examples/text2music/example_46.json

@@ -0,0 +1,10 @@
+{
+    "think": true,
+    "caption": "An energetic indie rock track driven by a clean, catchy electric guitar riff and a punchy, straightforward drum and bass groove. The male lead vocal is clear and powerful, delivering an anthemic melody that soars during the high-energy chorus. The arrangement follows a classic verse-chorus structure, punctuated by a brief, melodic guitar break and a dynamic bridge that builds back into a final, powerful chorus before concluding with the initial guitar riff.",
+    "lyrics": "[Intro - Guitar Riff]\n\n[Verse 1]\nWoke up in the electric city, bright light\nNeon signs and shadows we ignite\nStepping to the beat, feel the heart race\nLife's a wild ride in this endless place\n\n[Pre-Chorus]\nSynths and bass, they carry us above\nFunk rhythms mixed with electric love\nDrums pounding loud in the crowded street\nGuitar strings strum to our moving feet\n\n[Chorus]\nTurn it up, feel the electric heartbeat\nEvery pulse bringing us to our feet\nDance through the night, we're invincible\nElectric heartbeat, unstoppable\n\n[Instrumental Break - Guitar Riff]\n\n[Verse 2]\nCity lights flicker in a rhythmic trance\nUnder moon's glow, we take a chance\nVoices harmonize in a primal scream\nEvery note a spark, fueling our dream\n\n[Pre-Chorus]\nFeel the rush, let it take control\nElectric fire burning in our soul\nNo turning back, we're lost in the sound\nIn this moment, forever unbound\n\n[Chorus]\nTurn it up, feel the electric heartbeat\nEvery pulse bringing us to our feet\nDance through the night, we're invincible\nElectric heartbeat, unstoppable\n\n[Outro - Guitar Riff and Drums]\n[abrupt silence]",
+    "bpm": 100,
+    "duration": 153,
+    "keyscale": "A major",
+    "language": "en",
+    "timesignature": "4"
+}