chatterbox_wrapper.py

"""
Chatterbox Turbo TTS — ONNX Inference Wrapper
═══════════════════════════════════════════════
Orchestrates the 4-component ONNX pipeline:
  embed_tokens → speech_encoder → language_model → conditional_decoder

Optimised for lowest-latency CPU inference on 2 vCPU:
  • Sequential execution, thread count = physical cores, no spinning
  • Token list pre-allocation (avoids O(n²) np.concatenate in loop)
  • In-memory voice caching (no disk writes for uploads)
  • Robust audio loading: WAV, MP3, MPEG, M4A, OGG, FLAC, WebM
  • Sentence-level streaming for real-time playback
"""

# ── Suppress harmless transformers warnings BEFORE import ─────────
import os
import warnings

os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
warnings.filterwarnings("ignore", message=".*model of type.*chatterbox.*")

import hashlib
import io
import logging
import subprocess
import tempfile
import time
from collections import OrderedDict
from dataclasses import dataclass
from pathlib import Path
from typing import Callable, Generator, Optional

import librosa
import numpy as np
import onnxruntime as ort
import soundfile as soundfile_lib
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer

from config import Config
import text_processor

logger = logging.getLogger(__name__)

# ── Supported audio MIME types for voice upload ───────────────────
_SUPPORTED_AUDIO_EXTENSIONS = {
    ".wav", ".mp3", ".mpeg", ".mpga", ".m4a", ".mp4",
    ".ogg", ".oga", ".opus", ".flac", ".webm", ".aac", ".wma",
}


def _slugify(text: str) -> str:
    buf = []
    prev_underscore = False
    for ch in text.strip().lower():
        if ch.isalnum():
            buf.append(ch)
            prev_underscore = False
        else:
            if not prev_underscore:
                buf.append("_")
                prev_underscore = True
    slug = "".join(buf).strip("_")
    return slug or "voice"


# ═══════════════════════════════════════════════════════════════════
# Data Structures
# ═══════════════════════════════════════════════════════════════════

@dataclass
class VoiceProfile:
    """Cached speaker embedding extracted from reference audio."""
    cond_emb: np.ndarray
    prompt_token: np.ndarray
    speaker_embeddings: np.ndarray
    speaker_features: np.ndarray
    audio_hash: str = ""


class GenerationCancelled(Exception):
    """Raised when inference is cancelled by the client."""
    pass


# ═══════════════════════════════════════════════════════════════════
# LRU Voice Cache
# ═══════════════════════════════════════════════════════════════════

class _VoiceCache:
    """LRU cache for VoiceProfile objects with TTL-based expiration.

    Entries auto-expire after `ttl_seconds` (default: 1 hour).
    Re-uploading the same voice file within the TTL window returns
    the cached profile instantly — no re-encoding needed.
    """

    def __init__(self, maxsize: int, ttl_seconds: int = 3600):
        self._cache: OrderedDict[str, tuple[VoiceProfile, float]] = OrderedDict()
        self._maxsize = maxsize
        self._ttl = ttl_seconds

    def _evict_expired(self):
        """Remove all entries older than TTL."""
        now = time.time()
        expired = [k for k, (_, ts) in self._cache.items() if now - ts > self._ttl]
        for k in expired:
            del self._cache[k]
            logger.debug(f"Voice cache expired: {k[:8]}…")

    def get(self, key: str) -> Optional[VoiceProfile]:
        self._evict_expired()
        if key in self._cache:
            profile, ts = self._cache[key]
            remaining = self._ttl - (time.time() - ts)
            self._cache.move_to_end(key)
            logger.info(f"Voice cache HIT: {key[:8]}… (expires in {remaining:.0f}s)")
            return profile
        return None

    def put(self, key: str, profile: VoiceProfile):
        self._evict_expired()
        if key in self._cache:
            self._cache.move_to_end(key)
        else:
            if len(self._cache) >= self._maxsize:
                evicted_key, _ = self._cache.popitem(last=False)
                logger.debug(f"Voice cache evicted (LRU): {evicted_key[:8]}…")
        self._cache[key] = (profile, time.time())
        logger.info(f"Voice cache STORED: {key[:8]}… (TTL: {self._ttl}s, size: {len(self._cache)})")

    @property
    def size(self) -> int:
        return len(self._cache)


# ═══════════════════════════════════════════════════════════════════
# Audio Loading (robust multi-format support)
# ═══════════════════════════════════════════════════════════════════

def _load_audio_bytes(audio_bytes: bytes, sr: int = 24000) -> np.ndarray:
    """Load audio from raw bytes, supporting WAV/MP3/MPEG/M4A/OGG/FLAC/WebM.

    Strategy: try soundfile (fast, native) → librosa (ffmpeg backend) → ffmpeg CLI.
    """
    buf = io.BytesIO(audio_bytes)

    # 1) Try soundfile (handles WAV, FLAC, OGG natively — fastest)
    try:
        audio, file_sr = soundfile_lib.read(buf)
        if audio.ndim > 1:
            audio = audio.mean(axis=1)  # stereo → mono
        if file_sr != sr:
            audio = librosa.resample(audio.astype(np.float32), orig_sr=file_sr, target_sr=sr)
        return audio.astype(np.float32)
    except Exception:
        buf.seek(0)

    # 2) Try librosa (handles MP3 via audioread + ffmpeg backend)
    try:
        audio, _ = librosa.load(buf, sr=sr, mono=True)
        return audio.astype(np.float32)
    except Exception:
        buf.seek(0)

    # 3) Fallback: use ffmpeg CLI to convert to WAV in memory
    try:
        proc = subprocess.run(
            ["ffmpeg", "-i", "pipe:0", "-f", "wav", "-ac", "1", "-ar", str(sr), "pipe:1"],
            input=audio_bytes, capture_output=True, timeout=30,
        )
        if proc.returncode == 0 and len(proc.stdout) > 44:
            wav_buf = io.BytesIO(proc.stdout)
            audio, _ = soundfile_lib.read(wav_buf)
            return audio.astype(np.float32)
    except Exception:
        pass

    raise ValueError(
        "Could not decode audio file. Supported formats: "
        "WAV, MP3, MPEG, M4A, OGG, FLAC, WebM, AAC. "
        "Please upload a valid audio file."
    )


# ═══════════════════════════════════════════════════════════════════
# Main Wrapper
# ═══════════════════════════════════════════════════════════════════

class ChatterboxWrapper:

    def __init__(self, download_only: bool = False):
        self.cfg = Config
        os.makedirs(self.cfg.MODELS_DIR, exist_ok=True)

        logger.info(f"Downloading ONNX models (dtype={self.cfg.MODEL_DTYPE}) …")
        self._model_paths = self._download_models()

        if download_only:
            return

        logger.info(
            f"Creating ONNX Runtime sessions "
            f"(intra_op_threads={self.cfg.CPU_THREADS}, workers={self.cfg.MAX_WORKERS}) …"
        )
        opts = self._make_session_options()
        providers = ["CPUExecutionProvider"]

        self.embed_session   = ort.InferenceSession(self._model_paths["embed_tokens"],       sess_options=opts, providers=providers)
        self.encoder_session = ort.InferenceSession(self._model_paths["speech_encoder"],      sess_options=opts, providers=providers)
        self.lm_session      = ort.InferenceSession(self._model_paths["language_model"],      sess_options=opts, providers=providers)
        self.decoder_session = ort.InferenceSession(self._model_paths["conditional_decoder"], sess_options=opts, providers=providers)

        logger.info("Loading tokenizer …")
        self.tokenizer = AutoTokenizer.from_pretrained(self.cfg.MODEL_ID)

        self._voice_cache = _VoiceCache(
            maxsize=self.cfg.VOICE_CACHE_SIZE,
            ttl_seconds=self.cfg.VOICE_CACHE_TTL_SEC,
        )

        self._builtin_voice_profiles: dict[str, VoiceProfile] = {}
        self._builtin_voice_bytes: dict[str, bytes] = {}
        self._builtin_voice_by_hash: dict[str, VoiceProfile] = {}
        self._voice_alias_to_id: dict[str, str] = {}
        self._builtin_voice_catalog: list[dict] = []
        self._default_voice_id: str = "default"

        logger.info("Loading built-in voices (HF default + local samples) …")
        self.default_voice = self._load_builtin_voices()

        logger.info("✅ ChatterboxWrapper ready")

    # ─── Model download ──────────────────────────────────────────

    def _download_models(self) -> dict:
        """Download all 4 ONNX components + weight files from HuggingFace."""
        components = ("conditional_decoder", "speech_encoder", "embed_tokens", "language_model")
        paths = {}
        for name in components:
            paths[name] = self._download_component(name, self.cfg.MODEL_DTYPE)
        return paths

    def _download_component(self, name: str, dtype: str) -> str:
        if dtype == "fp32":
            filename = f"{name}.onnx"
        elif dtype == "q8":
            filename = f"{name}_quantized.onnx"
        else:
            filename = f"{name}_{dtype}.onnx"

        graph = hf_hub_download(
            self.cfg.MODEL_ID, subfolder="onnx", filename=filename,
            cache_dir=self.cfg.MODELS_DIR,
        )
        # Download companion weight file
        try:
            hf_hub_download(
                self.cfg.MODEL_ID, subfolder="onnx", filename=f"{filename}_data",
                cache_dir=self.cfg.MODELS_DIR,
            )
        except Exception:
            pass  # Some quantized variants embed weights in-graph
        return graph

    # ─── Session configuration (optimised for 2 vCPU) ─────────────

    def _make_session_options(self) -> ort.SessionOptions:
        opts = ort.SessionOptions()
        # Sequential execution: no parallel graph scheduling overhead
        opts.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
        # Match physical cores exactly (2 for HF Space free tier)
        opts.intra_op_num_threads = self.cfg.CPU_THREADS
        opts.inter_op_num_threads = 1
        # Full graph optimisations (constant folding, fusion, etc.)
        opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
        # Disable thread spinning — wastes CPU cycles on busy-wait
        opts.add_session_config_entry("session.intra_op.allow_spinning", "0")
        opts.add_session_config_entry("session.inter_op.allow_spinning", "0")
        # Enable memory optimisations
        opts.enable_cpu_mem_arena = True
        opts.enable_mem_pattern = True
        opts.enable_mem_reuse = True
        return opts

    # ─── Built-in voices (HF default + local samples) ────────────

    def _download_hf_default_voice_bytes(self) -> bytes:
        path = hf_hub_download(
            self.cfg.DEFAULT_VOICE_REPO,
            filename=self.cfg.DEFAULT_VOICE_FILE,
            cache_dir=self.cfg.MODELS_DIR,
        )
        return Path(path).read_bytes()

    def _list_local_voice_paths(self) -> list[Path]:
        wrapper_dir = Path(__file__).resolve().parent

        # Support both module-level and repo-root deployment layouts.
        candidates = []
        for d in (wrapper_dir, Path.cwd().resolve(), wrapper_dir.parent):
            try:
                resolved = d.resolve()
            except Exception:
                continue
            if resolved.is_dir() and resolved not in candidates:
                candidates.append(resolved)

        voices: list[Path] = []
        seen_real_paths: set[str] = set()
        for root in candidates:
            try:
                entries = sorted(root.iterdir(), key=lambda x: x.name.lower())
            except Exception:
                continue

            for p in entries:
                if not p.is_file():
                    continue
                if p.suffix.lower() not in _SUPPORTED_AUDIO_EXTENSIONS:
                    continue
                real_path = str(p.resolve())
                if real_path in seen_real_paths:
                    continue
                seen_real_paths.add(real_path)
                voices.append(p)

        logger.info(
            "Local voice scan complete: %s files across %s",
            len(voices),
            [str(x) for x in candidates],
        )
        return voices

    def _make_unique_voice_id(self, preferred: str) -> str:
        base = _slugify(preferred)
        candidate = base
        idx = 2
        while candidate in self._builtin_voice_profiles:
            candidate = f"{base}_{idx}"
            idx += 1
        return candidate

    def _register_builtin_voice(
        self,
        *,
        preferred_id: str,
        display_name: str,
        source: str,
        source_ref: str,
        audio_bytes: bytes,
        is_default: bool = False,
    ) -> str:
        if not audio_bytes:
            raise ValueError("Voice file is empty")

        voice_id = self._make_unique_voice_id(preferred_id)
        audio_hash = hashlib.md5(audio_bytes).hexdigest()

        profile = self._voice_cache.get(audio_hash)
        if profile is None:
            audio = _load_audio_bytes(audio_bytes, sr=self.cfg.SAMPLE_RATE)
            profile = self._encode_audio_array(audio, audio_hash=audio_hash)
            self._voice_cache.put(audio_hash, profile)
        else:
            # Keep hash attached to cached profile for metadata/voice-key usage.
            profile.audio_hash = audio_hash

        self._builtin_voice_profiles[voice_id] = profile
        self._builtin_voice_bytes[voice_id] = audio_bytes
        if audio_hash:
            self._builtin_voice_by_hash[audio_hash] = profile

        aliases: list[str] = []
        for alias in (voice_id, _slugify(Path(display_name).stem)):
            if alias not in self._voice_alias_to_id:
                self._voice_alias_to_id[alias] = voice_id
                aliases.append(alias)

        if is_default:
            self._default_voice_id = voice_id
            self._voice_alias_to_id["default"] = voice_id
            if "default" not in aliases:
                aliases.append("default")

        self._builtin_voice_catalog.append(
            {
                "id": voice_id,
                "display_name": display_name,
                "source": source,
                "source_ref": source_ref,
                "aliases": aliases,
                "voice_key": audio_hash,
            }
        )
        return voice_id

    def _load_builtin_voices(self) -> VoiceProfile:
        # 1) HF default voice (kept as true default fallback)
        hf_bytes = self._download_hf_default_voice_bytes()
        self._register_builtin_voice(
            preferred_id="default_hf_voice",
            display_name=self.cfg.DEFAULT_VOICE_FILE,
            source="huggingface",
            source_ref=f"{self.cfg.DEFAULT_VOICE_REPO}:{self.cfg.DEFAULT_VOICE_FILE}",
            audio_bytes=hf_bytes,
            is_default=True,
        )

        # 2) Local voice samples placed next to app files
        for path in self._list_local_voice_paths():
            # Avoid duplicate entry if someone also copied default_voice.wav locally.
            if path.name == self.cfg.DEFAULT_VOICE_FILE:
                continue
            try:
                self._register_builtin_voice(
                    preferred_id=path.stem,
                    display_name=path.name,
                    source="local",
                    source_ref=str(path.name),
                    audio_bytes=path.read_bytes(),
                    is_default=False,
                )
            except Exception as e:
                logger.warning(f"Skipping local voice {path.name}: {e}")

        default_profile = self._builtin_voice_profiles.get(self._default_voice_id)
        if default_profile is None:
            raise RuntimeError("Default built-in voice could not be initialized")

        logger.info(
            f"Built-in voices loaded: {len(self._builtin_voice_catalog)} "
            f"(default={self._default_voice_id})"
        )
        return default_profile

    def list_builtin_voices(self) -> list[dict]:
        """Return metadata for startup-preloaded voices."""
        return [dict(v) for v in self._builtin_voice_catalog]

    @property
    def default_voice_name(self) -> str:
        return self._default_voice_id

    def resolve_voice_id(self, voice_name: Optional[str]) -> str:
        if voice_name is None:
            return self._default_voice_id
        key = _slugify(str(voice_name))
        if not key:
            return self._default_voice_id
        voice_id = self._voice_alias_to_id.get(key)
        if voice_id is None:
            available = ", ".join(sorted(self._voice_alias_to_id.keys()))
            raise ValueError(f"Unknown voice '{voice_name}'. Available: {available}")
        return voice_id

    def get_builtin_voice(self, voice_name: Optional[str]) -> VoiceProfile:
        voice_id = self.resolve_voice_id(voice_name)
        profile = self._builtin_voice_profiles[voice_id]
        if profile.audio_hash:
            self._voice_cache.put(profile.audio_hash, profile)
        return profile

    def get_builtin_voice_bytes(self, voice_name: Optional[str]) -> Optional[bytes]:
        voice_id = self.resolve_voice_id(voice_name)
        return self._builtin_voice_bytes.get(voice_id)

    def get_builtin_voice_by_hash(self, audio_hash: str) -> Optional[VoiceProfile]:
        return self._builtin_voice_by_hash.get((audio_hash or "").strip())

    # ─── Voice encoding ──────────────────────────────────────────

    def encode_voice_from_bytes(self, audio_bytes: bytes) -> VoiceProfile:
        """Encode reference audio from raw bytes (in-memory, no disk write).

        Accepts: WAV, MP3, MPEG, M4A, OGG, FLAC, WebM, AAC, WMA, Opus.
        """
        audio_hash = hashlib.md5(audio_bytes).hexdigest()
        cached = self._voice_cache.get(audio_hash)
        if cached is not None:
            logger.info(f"Voice cache hit: {audio_hash[:8]}…")
            return cached

        # Robust multi-format audio loading
        audio = _load_audio_bytes(audio_bytes, sr=self.cfg.SAMPLE_RATE)

        # Validate duration
        duration = len(audio) / self.cfg.SAMPLE_RATE
        if duration < self.cfg.MIN_REF_DURATION_SEC:
            raise ValueError(
                f"Reference audio too short ({duration:.1f}s). "
                f"Minimum: {self.cfg.MIN_REF_DURATION_SEC}s"
            )
        if duration > self.cfg.MAX_REF_DURATION_SEC:
            audio = audio[: int(self.cfg.MAX_REF_DURATION_SEC * self.cfg.SAMPLE_RATE)]

        profile = self._encode_audio_array(audio, audio_hash=audio_hash)
        self._voice_cache.put(audio_hash, profile)
        return profile

    def _encode_audio_array(self, audio: np.ndarray, audio_hash: str = "") -> VoiceProfile:
        """Run speech_encoder on a float32 mono audio array."""
        audio_input = audio[np.newaxis, :].astype(np.float32)
        cond_emb, prompt_token, speaker_emb, speaker_feat = self.encoder_session.run(
            None, {"audio_values": audio_input}
        )
        return VoiceProfile(
            cond_emb=cond_emb,
            prompt_token=prompt_token,
            speaker_embeddings=speaker_emb,
            speaker_features=speaker_feat,
            audio_hash=audio_hash,
        )

    # ─── Full generation (non-streaming) ──────────────────────────

    def generate_speech(
        self,
        text: str,
        voice: Optional[VoiceProfile] = None,
        max_new_tokens: Optional[int] = None,
        repetition_penalty: Optional[float] = None,
    ) -> np.ndarray:
        """Generate complete audio for the given text."""
        voice = voice or self.default_voice
        text = text_processor.sanitize(text.strip()[: self.cfg.MAX_TEXT_LENGTH])
        if not text:
            raise ValueError("Text is empty after sanitization")

        tokens = self._generate_tokens(
            text, voice,
            max_new_tokens or self.cfg.MAX_NEW_TOKENS,
            repetition_penalty or self.cfg.REPETITION_PENALTY,
        )
        return self._decode_tokens(tokens, voice)

    # ─── Streaming generation ─────────────────────────────────────

    def stream_speech(
        self,
        text: str,
        voice: Optional[VoiceProfile] = None,
        max_new_tokens: Optional[int] = None,
        repetition_penalty: Optional[float] = None,
        is_cancelled: Optional[Callable[[], bool]] = None,
    ) -> Generator[np.ndarray, None, None]:
        """Yield audio chunks sentence-by-sentence for real-time streaming.

        Each sentence is independently processed through the full pipeline
        so the first chunk arrives as fast as possible (low TTFB).

        Args:
            is_cancelled: Optional callable that returns True to abort generation.
                          Checked between chunks and every 25 autoregressive steps.
        """
        voice = voice or self.default_voice
        text = text_processor.sanitize(text.strip()[: self.cfg.MAX_TEXT_LENGTH])
        if not text:
            return

        sentences = text_processor.split_for_streaming(text)
        _max = max_new_tokens or self.cfg.MAX_NEW_TOKENS
        _rep = repetition_penalty or self.cfg.REPETITION_PENALTY
        _check = is_cancelled or (lambda: False)

        for i, sentence in enumerate(sentences):
            # Check cancellation between chunks
            if _check():
                logger.info("Generation cancelled by client (between chunks)")
                return
            if not sentence.strip():
                continue
            t0 = time.perf_counter()
            try:
                tokens = self._generate_tokens(sentence, voice, _max, _rep, _check)
                if _check():
                    return
                audio = self._decode_tokens(tokens, voice)
                elapsed = time.perf_counter() - t0
                audio_duration = len(audio) / self.cfg.SAMPLE_RATE
                rtf = elapsed / audio_duration if audio_duration > 0 else 0
                logger.info(
                    f"Chunk {i + 1}/{len(sentences)}: "
                    f"{len(sentence)} chars → {audio_duration:.1f}s audio "
                    f"in {elapsed:.2f}s (RTF: {rtf:.2f}x)"
                )
                yield audio
            except GenerationCancelled:
                logger.info(f"Generation cancelled mid-token at chunk {i + 1}")
                return
            except Exception as e:
                logger.error(f"Error on chunk {i + 1}: {e}")
                raise

    # ─── Autoregressive token generation (OPTIMISED) ──────────────

    def _generate_tokens(
        self,
        text: str,
        voice: VoiceProfile,
        max_new_tokens: int,
        repetition_penalty: float,
        is_cancelled: Callable[[], bool] = lambda: False,
    ) -> np.ndarray:
        """Run embed → LM autoregressive loop. Returns raw token array.

        Optimisations:
          • Token list instead of repeated np.concatenate (O(n) → O(1) append)
          • Unique tokens set for inline repetition penalty (avoids exponential penalty bug)
          • Pre-allocated attention mask for zero-copy slicing
          • Correct dimensional routing for step 0 prompt processing
        """
        input_ids = self.tokenizer(text, return_tensors="np")["input_ids"].astype(np.int64)

        # Pre-allocate collections
        token_list: list[int] = [self.cfg.START_SPEECH_TOKEN]
        unique_tokens: set[int] = {self.cfg.START_SPEECH_TOKEN}
        penalty = repetition_penalty

        past_key_values = None
        attention_mask_full = None
        seq_len = 0
        
        for step in range(max_new_tokens):
            if step > 0 and step % 25 == 0 and is_cancelled():
                raise GenerationCancelled()
            
            embeds = self.embed_session.run(None, {"input_ids": input_ids})[0]

            if step == 0:
                # Prepend speaker conditioning
                embeds = np.concatenate((voice.cond_emb, embeds), axis=1)
                batch, seq_len, _ = embeds.shape
                
                past_key_values = {
                    inp.name: np.zeros(
                        [batch, self.cfg.NUM_KV_HEADS, 0, self.cfg.HEAD_DIM],
                        dtype=np.float16 if inp.type == "tensor(float16)" else np.float32,
                    )
                    for inp in self.lm_session.get_inputs()
                    if "past_key_values" in inp.name
                }
                
                # Pre-allocate full attention mask
                attention_mask_full = np.ones((batch, seq_len + max_new_tokens), dtype=np.int64)
                attention_mask = attention_mask_full[:, :seq_len]
                
                # Step 0 requires position_ids matching prompt sequence length
                position_ids = np.arange(seq_len, dtype=np.int64).reshape(batch, -1)
            else:
                # O(1) zero-copy slice for subsequent steps
                attention_mask = attention_mask_full[:, : seq_len + step]
                # Single position ID for the single new token
                position_ids = np.array([[seq_len + step - 1]], dtype=np.int64)

            # Language model forward pass
            logits, *present_kv = self.lm_session.run(
                None,
                dict(
                    inputs_embeds=embeds,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    **past_key_values,
                ),
            )

            # ── Inline repetition penalty + token selection ───────
            last_logits = logits[0, -1, :].copy()  # shape: (vocab_size,)

            # Apply repetition penalty strictly to unique tokens to prevent over-penalization
            for tok_id in unique_tokens:
                if last_logits[tok_id] < 0:
                    last_logits[tok_id] *= penalty
                else:
                    last_logits[tok_id] /= penalty

            next_token = int(np.argmax(last_logits))
            token_list.append(next_token)
            unique_tokens.add(next_token)

            if next_token == self.cfg.STOP_SPEECH_TOKEN:
                break

            # Update state for next step
            input_ids = np.array([[next_token]], dtype=np.int64)
            for j, key in enumerate(past_key_values):
                past_key_values[key] = present_kv[j]

        return np.array([token_list], dtype=np.int64)

    # ─── Token → audio decoding ───────────────────────────────────

    def _decode_tokens(self, generated: np.ndarray, voice: VoiceProfile) -> np.ndarray:
        """Decode speech tokens to a float32 waveform at 24 kHz."""
        # Strip START token; strip STOP token if present
        tokens = generated[:, 1:]
        if tokens.shape[1] > 0 and tokens[0, -1] == self.cfg.STOP_SPEECH_TOKEN:
            tokens = tokens[:, :-1]

        if tokens.shape[1] == 0:
            return np.zeros(0, dtype=np.float32)

        # Prepend prompt token + append silence
        silence = np.full(
            (tokens.shape[0], 3), self.cfg.SILENCE_TOKEN, dtype=np.int64
        )
        full_tokens = np.concatenate(
            [voice.prompt_token, tokens, silence], axis=1
        )

        wav = self.decoder_session.run(
            None,
            {
                "speech_tokens": full_tokens,
                "speaker_embeddings": voice.speaker_embeddings,
                "speaker_features": voice.speaker_features,
            },
        )[0].squeeze(axis=0)

        return wav

    # ─── Warmup ───────────────────────────────────────────────────

    def warmup(self):
        """Run a short inference to warm up ONNX sessions and JIT paths."""
        try:
            t0 = time.perf_counter()
            _ = self.generate_speech("Hello.", self.default_voice, max_new_tokens=32)
            logger.info(f"Warmup done in {time.perf_counter() - t0:.2f}s")
        except Exception as e:
            logger.warning(f"Warmup failed (non-critical): {e}")