T

app.py

@@ -1,8 +1,4 @@
 # coding=utf-8
-# Qwen3-TTS Gradio Demo for HuggingFace Spaces with Zero GPU
-# Supports: Voice Design, Voice Clone (Base), TTS (CustomVoice)
-#import subprocess
-#subprocess.run('pip install flash-attn==2.7.4.post1', shell=True)
 import os
 import sys
 import logging
@@ -18,33 +14,26 @@ import uuid
 import random
 import whisper
 import librosa
+from opencc import OpenCC
+
 # 配置日志
 logging.basicConfig(
     level=logging.INFO,
     format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
-    handlers=[
-        logging.StreamHandler(sys.stdout)
-    ]
+    handlers=[logging.StreamHandler(sys.stdout)]
 )
-# 屏蔽第三方库的冗余日志
-logging.getLogger("httpx").setLevel(logging.WARNING)
-logging.getLogger("urllib3").setLevel(logging.WARNING)
-logging.getLogger("httpcore").setLevel(logging.WARNING)
-logging.getLogger("gradio").setLevel(logging.WARNING)
-
 logger = logging.getLogger("Qwen3-TTS-Demo")
 
+# 初始化简繁转换器
+cc = OpenCC('t2s')
+
 HF_TOKEN = os.environ.get('HF_TOKEN')
-login(token=HF_TOKEN)
+if HF_TOKEN:
+    login(token=HF_TOKEN)
 
-# Model size options
 MODEL_SIZES = ["0.6B", "1.7B"]
-
-# Speaker and language choices for CustomVoice model
-SPEAKERS = [
-    "Aiden", "Dylan", "Eric", "Ono_anna", "Ryan", "Serena", "Sohee", "Uncle_fu", "Vivian"
-]
 LANGUAGES = ["Auto", "Chinese", "English", "Japanese", "Korean", "French", "German", "Spanish", "Portuguese", "Russian"]
+
 def seed_everything(seed=42):
     random.seed(seed)
     np.random.seed(seed)
@@ -55,42 +44,14 @@ def seed_everything(seed=42):
     torch.backends.cudnn.benchmark = False
 
 def get_model_path(model_type: str, model_size: str) -> str:
-    """Get model path based on type and size."""
     return snapshot_download(f"Qwen/Qwen3-TTS-12Hz-{model_size}-{model_type}")
 
-
-# ============================================================================
-# GLOBAL MODEL LOADING - Load all models at startup
-# ============================================================================
-logger.info("正在加载所有模型到 CUDA...")
-
-# # Voice Design model (1.7B only)
-# logger.info("正在加载 VoiceDesign 1.7B 模型...")
-# voice_design_model = Qwen3TTSModel.from_pretrained(
-#     get_model_path("VoiceDesign", "1.7B"),
-#     device_map="cuda",
-#     dtype=torch.bfloat16,
-#     token=HF_TOKEN,
-#     attn_implementation="kernels-community/flash-attn3",
-# )
-
-# # Base (Voice Clone) models - both sizes
-# logger.info("正在加载 Base 0.6B 模型...")
-# base_model_0_6b = Qwen3TTSModel.from_pretrained(
-#     get_model_path("Base", "0.6B"),
-#     device_map="cuda",
-#     dtype=torch.bfloat16,
-#     token=HF_TOKEN,
-#     attn_implementation="kernels-community/flash-attn3",
-# )
-
-@functools.lru_cache(maxsize=1) # 只缓存当前正在使用的模型，节省显存
+@functools.lru_cache(maxsize=1)
 def load_model(model_type, model_size):
-    logger.info(f"正在按需加载 {model_type} {model_size} 模型...")
     path = get_model_path(model_type, model_size)
     return Qwen3TTSModel.from_pretrained(
         path,
-        device_map="cuda", # 注意：在 ZeroGPU 环境下，这行只有在被装饰的函数内执行才有效
+        device_map="cuda",
         dtype=torch.bfloat16,
         token=HF_TOKEN,
         attn_implementation="kernels-community/flash-attn3"
@@ -98,67 +59,14 @@ def load_model(model_type, model_size):
 
 @functools.lru_cache(maxsize=1)
 def load_whisper_model(model_name="large-v3"):
-    logger.info(f"正在加载 Whisper 模型: {model_name}...")
-    # whisper.load_model 会自动处理下载和缓存
     model = whisper.load_model(model_name, device="cuda" if torch.cuda.is_available() else "cpu")
-    logger.info("Whisper 模型加载成功！")
     return model
 
-# logger.info("正在加载 Base 1.7B 模型...")
-# base_model_1_7b = Qwen3TTSModel.from_pretrained(
-#     get_model_path("Base", "1.7B"),
-#     device_map="cuda",
-#     dtype=torch.bfloat16,
-#     token=HF_TOKEN,
-#     attn_implementation="kernels-community/flash-attn3",
-# )
-
-# CustomVoice models - both sizes
-# logger.info("正在加载 CustomVoice 0.6B 模型...")
-# custom_voice_model_0_6b = Qwen3TTSModel.from_pretrained(
-#     get_model_path("CustomVoice", "0.6B"),
-#     device_map="cuda",
-#     dtype=torch.bfloat16,
-#     token=HF_TOKEN,
-#     attn_implementation="kernels-community/flash-attn3",
-# )
-
-# logger.info("正在加载 CustomVoice 1.7B 模型...")
-# custom_voice_model_1_7b = Qwen3TTSModel.from_pretrained(
-#     get_model_path("CustomVoice", "1.7B"),
-#     device_map="cuda",
-#     dtype=torch.bfloat16,
-#     token=HF_TOKEN,
-#     attn_implementation="kernels-community/flash-attn3",
-# )
-
-logger.info("所有模型加载成功！")
-
-# Model lookup dictionaries for easy access
-# BASE_MODELS = {
-#     "0.6B": base_model_0_6b,
-#     # "1.7B": base_model_1_7b,
-# }
-
-# CUSTOM_VOICE_MODELS = {
-#     "0.6B": custom_voice_model_0_6b,
-#     "1.7B": custom_voice_model_1_7b,
-# }
-
-# ============================================================================
-
-
 def _normalize_audio(wav, eps=1e-12, clip=True):
-    """Normalize audio to float32 in [-1, 1] range."""
     x = np.asarray(wav)
-
     if np.issubdtype(x.dtype, np.integer):
         info = np.iinfo(x.dtype)
-        if info.min < 0:
-            y = x.astype(np.float32) / max(abs(info.min), info.max)
-        else:
-            mid = (info.max + 1) / 2.0
-            y = (x.astype(np.float32) - mid) / mid
+        y = x.astype(np.float32) / max(abs(info.min), info.max)
     elif np.issubdtype(x.dtype, np.floating):
         y = x.astype(np.float32)
         m = np.max(np.abs(y)) if y.size else 0.0
@@ -166,39 +74,27 @@ def _normalize_audio(wav, eps=1e-12, clip=True):
             y = y / (m + eps)
     else:
         raise TypeError(f"Unsupported dtype: {x.dtype}")
-
     if clip:
         y = np.clip(y, -1.0, 1.0)
-
     if y.ndim > 1:
         y = np.mean(y, axis=-1).astype(np.float32)
-
     return y
 
-
 def _audio_to_tuple(audio):
-    """Convert Gradio audio input to (wav, sr) tuple."""
     if audio is None:
         return None
-
     if isinstance(audio, tuple) and len(audio) == 2 and isinstance(audio[0], int):
         sr, wav = audio
         wav = _normalize_audio(wav)
         return wav, int(sr)
-
     if isinstance(audio, dict) and "sampling_rate" in audio and "data" in audio:
         sr = int(audio["sampling_rate"])
         wav = _normalize_audio(audio["data"])
         return wav, sr
-
     return None
 
-
-
-
 @spaces.GPU
 def infer_voice_design(part, language, voice_description):
-    """Single segment inference for Voice Design."""
     voice_design_model = load_model("VoiceDesign","1.7B")
     seed_everything(42)
     wavs, sr = voice_design_model.generate_voice_design(
@@ -210,13 +106,8 @@ def infer_voice_design(part, language, voice_description):
     )
     return wavs[0], sr
 
-
-
 @spaces.GPU
-def infer_voice_clone( part, language,audio_tuple,ref_text,use_xvector_only):
-    """Single segment inference for Voice Clone using reference audio."""
-    # tts = BASE_MODELS[model_size]
-    # seed_everything(42)
+def infer_voice_clone(part, language, audio_tuple, ref_text, use_xvector_only):
     tts = load_model("Base", "0.6B")
     voice_clone_prompt = tts.create_voice_clone_prompt(
         ref_audio=audio_tuple,
@@ -228,97 +119,81 @@ def infer_voice_clone( part, language,audio_tuple,ref_text,use_xvector_only):
         language=language,
         voice_clone_prompt=voice_clone_prompt,
         max_new_tokens=2048,
-        # 核心参数：固定 seed
         seed=42, 
-        temperature=0.3, # 配合低温度，音色会更稳
+        temperature=0.3,
         top_p=0.85
     )
     return wavs[0], sr
 
 @spaces.GPU
 def infer_voice_clone_from_prompt(part, language, prompt_file_path):
-    """Single segment inference for Voice Clone using pre-extracted prompt."""
-    logger.info("正在加载音频特征文件...")
     loaded_data = torch.load(prompt_file_path, map_location='cuda', weights_only=False)
-    
-    # 兼容旧版本直接保存对象的情况
     if isinstance(loaded_data, list) and len(loaded_data) > 0 and isinstance(loaded_data[0], VoiceClonePromptItem):
         voice_clone_prompt = loaded_data
     elif isinstance(loaded_data, list) and len(loaded_data) > 0 and isinstance(loaded_data[0], dict):
-        # 从字典列表重建对象
         voice_clone_prompt = [VoiceClonePromptItem(**item) for item in loaded_data]
     else:
-        # 尝试作为单个对象处理
          voice_clone_prompt = loaded_data
-
-    # 维度校正：确保 ref_code 是 2D 的 (Time, Q)
     if isinstance(voice_clone_prompt, list):
         for item in voice_clone_prompt:
             if item.ref_code is not None and item.ref_code.ndim == 3:
-                # [1, T, Q] -> [T, Q]
                 item.ref_code = item.ref_code.squeeze(0)
-
-    logger.info("音频特征文件加载成功。")
-
     tts = load_model("Base", "0.6B")
-    logger.info(f"克隆音频，目标文本：{part}")
     wavs, sr = tts.generate_voice_clone(
         text=part,
         language=language,
         voice_clone_prompt=voice_clone_prompt,
         max_new_tokens=2048,
-        # 核心参数：固定 seed
         seed=42, 
-        temperature=0.3, # 配合低温度，音色会更稳
+        temperature=0.3,
         top_p=0.85
     )
     return wavs[0], sr
 
 @spaces.GPU
-def extract_voice_clone_prompt(ref_audio,ref_text,use_xvector_only):
-    logger.info("正在提取参考音频特征（仅执行一次）...")
+def extract_voice_clone_prompt(ref_audio, ref_text, use_xvector_only):
     tts = load_model("Base", "0.6B")
     seed_everything(42)
     audio_tuple = _audio_to_tuple(ref_audio)
     if audio_tuple is None:
         return None, "错误：需要参考音频。"
-
-    # if not use_xvector_only and (not ref_text or not ref_text.strip()):
-    #     return None, "错误：未启用 '仅使用 x-vector' 时需要参考文本。"
     r_text = ref_text
     uxo = use_xvector_only
-    
-    # 如果没有提供参考文本且未开启仅 x-vector 模式，尝试使用 Whisper 自动识别
     if not r_text or (isinstance(r_text, str) and not r_text.strip()):
         whisper_size = "base"
-        logger.info(f"未提供参考文本，开始使用 Whisper 自动识别。模型: {whisper_size}")
         try:
             whisper_model = load_whisper_model(whisper_size)
             audio_data, sr = audio_tuple
-            
-            # 仅为 Whisper 识别进行重采样，不影响原始 audio_tuple
             if sr != 16000:
-                logger.info(f"Whisper 识别：临时重采样音频 {sr}Hz -> 16000Hz")
                 whisper_audio = librosa.resample(audio_data, orig_sr=sr, target_sr=16000)
             else:
                 whisper_audio = audio_data
-                
             result = whisper_model.transcribe(whisper_audio)
-            r_text = result["text"].strip()
-            logger.info(f"Whisper 识别成功：{r_text}")
+            
+            res_val = result.get("text", "")
+            if isinstance(res_val, list) and len(res_val) > 0:
+                res_val = res_val[0]
+            if not isinstance(res_val, str):
+                res_val = str(res_val)
+            r_text = cc.convert(res_val.strip())
             uxo = False
         except Exception as e:
             logger.error(f"Whisper 识别失败: {str(e)}", exc_info=True)
-            return None, f"错误：语音识别失败且未提供参考文本。{str(e)}"
-
+            uxo = True
+            # return None, f"错误：语音识别失败且未提供参考文本。{str(e)}"
+    
+    r_text_str = ""
+    if isinstance(r_text, str):
+        r_text_str = r_text.strip()
+    elif isinstance(r_text, list) and len(r_text) > 0 and isinstance(r_text[0], str):
+        r_text_str = r_text[0].strip()
+        
+    logger.info(f"语音识别成功 ：{r_text_str}")
     voice_clone_prompt_items = tts.create_voice_clone_prompt(
         ref_audio=audio_tuple,
-        ref_text=r_text.strip() if (isinstance(r_text, str) and r_text) else None,
+        ref_text=r_text_str if r_text_str else None,
         x_vector_only_mode=uxo
     )
-    logger.info("参考音频特征提取完成。")
-    
-    # 转换为字典列表保存，避免对象序列化问题
     prompt_data = []
     for item in voice_clone_prompt_items:
         prompt_data.append({
@@ -328,413 +203,120 @@ def extract_voice_clone_prompt(ref_audio,ref_text,use_xvector_only):
             "icl_mode": item.icl_mode,
             "ref_text": item.ref_text
         })
-
-    # 生成唯一的文件名
     file_id = str(uuid.uuid4())[:8]
     file_path = f"voice_clone_prompt_{file_id}.pt"
-    
-    # 保存到文件
     torch.save(prompt_data, file_path)
-    logger.info(f"voice_clone_prompt 已保存到: {file_path}")
-    
     return file_path
-# @spaces.GPU(duration=60)
-# def infer_custom_voice(model_size, part, language, speaker, instruct):
-#     """Single segment inference for Custom Voice."""
-#     tts = CUSTOM_VOICE_MODELS[model_size]
-#     wavs, sr = tts.generate_custom_voice(
-#         text=part,
-#         language=language,
-#         speaker=speaker.lower().replace(" ", "_"),
-#         instruct=instruct.strip() if instruct else None,
-#         non_streaming_mode=True,
-#         max_new_tokens=2048,
-#     )
-#     return wavs[0], sr
 
-
-def generate_voice_design(text, language, voice_description, progress=gr.Progress(track_tqdm=True)):
-    """Generate speech using Voice Design model (1.7B only)."""
+def generate_voice_design(text, language, voice_description):
     if not text or not text.strip():
         return None, "错误：文本不能为空。"
     if not voice_description or not voice_description.strip():
         return None, "错误：语音描述不能为空。"
-
-    logger.info(f"开始 Voice Design 生成任务。语言: {language}, 文本长度: {len(text)}, 描述: {voice_description}")
     try:
         wav, sr = infer_voice_design(text.strip(), language, voice_description)
-        logger.info("Voice Design 生成任务完成...")
         return (sr, wav), "语音设计生成成功！"
     except Exception as e:
         logger.error(f"Voice Design 生成失败: {str(e)}", exc_info=True)
-        return None, f"错误: {type(e).__name__}: {e}"
-
+        return None, f"错误: {e}"
 
-def generate_voice_clone(ref_audio, ref_text, target_text, language, use_xvector_only, model_size, progress=gr.Progress(track_tqdm=True)):
-    """Generate speech using Base (Voice Clone) model."""
-    if not target_text or not target_text.strip():
+def generate_voice_clone(ref_audio, ref_text, target_text, language, use_xvector_only):
+    t_text = target_text.strip() if isinstance(target_text, str) else ""
+    if not t_text:
         return None, "错误：目标文本不能为空。"
-
     audio_tuple = _audio_to_tuple(ref_audio)
     if audio_tuple is None:
         return None, "错误：需要参考音频。"
-
-    if not use_xvector_only and (not ref_text or not ref_text.strip()):
+    r_text = ref_text.strip() if isinstance(ref_text, str) else ""
+    if not use_xvector_only and not r_text:
         return None, "错误：未启用 '仅使用 x-vector' 时需要参考文本。"
-
-    logger.info(f"开始 Voice Clone 生成任务。模型大小: {model_size}, 语言: {language}, 目标文本长度: {len(target_text)}, 仅使用 x-vector: {use_xvector_only}")
     try:
-        wav, sr = infer_voice_clone(target_text.strip(), language, audio_tuple, ref_text, use_xvector_only)
-        logger.info("Voice Clone 生成任务完成...")
+        wav, sr = infer_voice_clone(t_text, language, audio_tuple, r_text, use_xvector_only)
         return (sr, wav), "语音克隆生成成功！"
     except Exception as e:
         logger.error(f"Voice Clone 生成失败: {str(e)}", exc_info=True)
-        return None, f"错误: {type(e).__name__}: {e}"
+        return None, f"错误: {e}"
 
-def generate_voice_clone_from_prompt_file(prompt_file_path, target_text, language, progress=gr.Progress(track_tqdm=True)):
-    """Generate speech using Base (Voice Clone) model with pre-extracted prompt file."""
-    if not target_text or not target_text.strip():
+def generate_voice_clone_from_prompt_file(prompt_file_path, target_text, language):
+    t_text = target_text.strip() if isinstance(target_text, str) else ""
+    if not t_text:
         return None, "错误：目标文本不能为空。"
-    
     if not prompt_file_path:
         return None, "错误：需要提供音频特征文件。"
-
-    logger.info(f"开始 Voice Clone 生成任务（使用特征文件）。语言: {language}, 目标文本长度: {len(target_text)}, 特征文件: {prompt_file_path}")
     try:
-        wav, sr = infer_voice_clone_from_prompt(target_text.strip(), language, prompt_file_path)
-        logger.info("Voice Clone 生成任务完成...")
+        wav, sr = infer_voice_clone_from_prompt(t_text, language, prompt_file_path)
         return (sr, wav), "语音克隆生成成功（使用特征文件）！"
     except Exception as e:
         logger.error(f"Voice Clone 生成失败: {str(e)}", exc_info=True)
-        return None, f"错误: {type(e).__name__}: {e}"
-
+        return None, f"错误: {e}"
 
 @spaces.GPU
-def infer_whisper_audio(audio_path, model_size="large-v3"):
-    """Transcribe audio using Whisper model."""
+def infer_whisper_audio(audio_path, model_size="base"):
     if not audio_path:
         return "错误：请上传音频文件或进行录音。"
-    
-    logger.info(f"开始 Whisper 语音识别任务。模型: {model_size}, 音频路径: {audio_path}")
     try:
         model = load_whisper_model(model_size)
-        
-        # 使用 transcribe 方法进行转录
-        # whisper 会自动处理音频加载和重采样
         result = model.transcribe(audio_path)
         
-        text = result["text"]
-        logger.info(f"Whisper 识别完成。文本长度: {len(text)}")
-        return text.strip()
+        res_val = result.get("text", "")
+        if isinstance(res_val, list) and len(res_val) > 0:
+            res_val = res_val[0]
+        if not isinstance(res_val, str):
+            res_val = str(res_val)
+        
+        return cc.convert(res_val.strip())
     except Exception as e:
         logger.error(f"Whisper 识别失败: {str(e)}", exc_info=True)
-        return f"识别出错: {type(e).__name__}: {e}"
+        return f"识别出错: {e}"
 
-
-# def generate_custom_voice(text, language, speaker, instruct, model_size, progress=gr.Progress(track_tqdm=True)):
-#     """Generate speech using CustomVoice model with segment-based GPU allocation."""
-#     if not text or not text.strip():
-#         return None, "错误：文本不能为空。"
-#     if not speaker:
-#         return None, "错误：说话人不能为空。"
-
-#     logger.info(f"开始 Custom Voice 生成任务。模型大小: {model_size}, 语言: {language}, 说话人: {speaker}, 指令: {instruct}, 文本长度: {len(text)}")
-#     try:
-#         text_parts = split_text(text.strip())
-#         logger.info(f"文本已切分为 {len(text_parts)} 段。")
-#         all_wavs = []
-#         sr = 24000
-
-#         for i, part in enumerate(progress.tqdm(text_parts, desc="正在生成分段")):
-#             logger.info(f"正在处理第 {i+1}/{len(text_parts)} 段文本...")
-#             wav, current_sr = infer_custom_voice(model_size, part, language, speaker, instruct)
-#             all_wavs.append(wav)
-#             sr = current_sr
-
-#         combined_wav = np.concatenate(all_wavs)
-#         logger.info("Custom Voice 生成任务完成，正在合并音频...")
-#         return (sr, combined_wav), "语音生成成功！"
-#     except Exception as e:
-#         logger.error(f"Custom Voice 生成失败: {str(e)}", exc_info=True)
-#         return None, f"错误: {type(e).__name__}: {e}"
-
-
-# Build Gradio UI
 def build_ui():
-    theme = gr.themes.Soft(
-        font=[gr.themes.GoogleFont("Source Sans Pro"), "Arial", "sans-serif"],
-    )
-
-    css = """
-    .gradio-container {max-width: none !important;}
-    .tab-content {padding: 20px;}
-    """
-
-    with gr.Blocks(theme=theme, css=css, title="Qwen3-TTS Demo") as demo:
-        gr.Markdown(
-            """
-# Qwen3-TTS Demo
-A unified Text-to-Speech demo featuring three powerful modes:
-- **Voice Design**: Create custom voices using natural language descriptions
-- **Voice Clone (Base)**: Clone any voice from a reference audio
-- **ASR (Whisper)**: Accurate speech-to-text using OpenAI's Whisper model
-- **TTS (CustomVoice)**: Generate speech with predefined speakers and optional style instructions
-Built with [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) by Alibaba Qwen Team.
-"""
-        )
-
+    theme = gr.themes.Soft(font=[gr.themes.GoogleFont("Source Sans Pro"), "Arial", "sans-serif"])
+    with gr.Blocks(theme=theme, title="Qwen3-TTS Demo") as demo:
+        gr.Markdown("# Qwen3-TTS Demo")
         with gr.Tabs():
-            # Tab 3: ASR (Whisper)
             with gr.Tab("ASR (Whisper)"):
-                gr.Markdown("### 语音识别 (Speech Recognition)")
-                gr.Markdown("使用 OpenAI Whisper 模型将语音转换为文本。")
-                
                 with gr.Row():
-                    with gr.Column(scale=1):
-                        asr_audio_input = gr.Audio(
-                            label="输入音频 (录音或上传)",
-                            type="filepath", # Whisper 需要文件路径
-                            sources=["microphone", "upload"]
-                        )
-                        asr_model_size = gr.Dropdown(
-                            label="Whisper 模型大小",
-                            choices=["base", "small", "medium", "large-v3"],
-                            value="large-v3",
-                            interactive=True,
-                            info="越大越准，但速度越慢"
-                        )
-                        asr_btn = gr.Button("开始识别 (Transcribe)", variant="primary")
-                    
-                    with gr.Column(scale=1):
-                        asr_text_output = gr.Textbox(
-                            label="识别结果",
-                            lines=10,
-                            show_copy_button=True
-                        )
-                
-                asr_btn.click(
-                    infer_whisper_audio,
-                    inputs=[asr_audio_input, asr_model_size],
-                    outputs=[asr_text_output],
-                    api_name="infer_whisper"
-                )
-
-            # Tab 1: Voice Design (Default, 1.7B only)
+                    with gr.Column():
+                        asr_audio_input = gr.Audio(label="输入音频", type="filepath", sources=["microphone", "upload"])
+                        asr_model_size = gr.Dropdown(label="Whisper 模型大小", choices=["base", "small", "medium", "large-v3"], value="large-v3")
+                        asr_btn = gr.Button("开始识别", variant="primary")
+                    with gr.Column():
+                        asr_text_output = gr.Textbox(label="识别结果", lines=10, show_copy_button=True)
+                asr_btn.click(infer_whisper_audio, inputs=[asr_audio_input, asr_model_size], outputs=[asr_text_output])
             with gr.Tab("Voice Design"):
-                gr.Markdown("### Create Custom Voice with Natural Language")
                 with gr.Row():
-                    with gr.Column(scale=2):
-                        design_text = gr.Textbox(
-                            label="Text to Synthesize",
-                            lines=4,
-                            placeholder="Enter the text you want to convert to speech...",
-                            value="It's in the top drawer... wait, it's empty? No way, that's impossible! I'm sure I put it there!"
-                        )
-                        design_language = gr.Dropdown(
-                            label="Language",
-                            choices=LANGUAGES,
-                            value="Auto",
-                            interactive=True,
-                        )
-                        design_instruct = gr.Textbox(
-                            label="Voice Description",
-                            lines=3,
-                            placeholder="Describe the voice characteristics you want...",
-                            value="Speak in an incredulous tone, but with a hint of panic beginning to creep into your voice."
-                        )
-                        design_btn = gr.Button("Generate with Custom Voice", variant="primary")
-
-                    with gr.Column(scale=2):
-                        design_audio_out = gr.Audio(label="Generated Audio", type="numpy")
-                        design_status = gr.Textbox(label="Status", lines=2, interactive=False)
-
-                design_btn.click(
-                    generate_voice_design,
-                    inputs=[design_text, design_language, design_instruct],
-                    outputs=[design_audio_out, design_status],
-                    api_name="generate_voice_design"
-                )
-
-            # Tab 2: Voice Clone (Base)
+                    with gr.Column():
+                        design_text = gr.Textbox(label="目标文本", lines=4, value="It's in the top drawer... wait, it's empty?")
+                        design_language = gr.Dropdown(label="语言", choices=LANGUAGES, value="Auto")
+                        design_instruct = gr.Textbox(label="语音描述", lines=3, value="Speak in an incredulous tone.")
+                        design_btn = gr.Button("开始生成", variant="primary")
+                    with gr.Column():
+                        design_audio_out = gr.Audio(label="生成音频", type="numpy")
+                        design_status = gr.Textbox(label="状态", interactive=False)
+                design_btn.click(generate_voice_design, inputs=[design_text, design_language, design_instruct], outputs=[design_audio_out, design_status])
             with gr.Tab("Voice Clone (Base)"):
-                # Section 1: Extract Voice Features
                 gr.Markdown("### 1. 提取音频特征")
-                gr.Markdown("上传参考音频并提取特征，保存为文件供后续使用。")
                 with gr.Row():
-                    with gr.Column(scale=2):
-                        extract_ref_audio = gr.Audio(
-                            label="参考音频",
-                            type="numpy",
-                        )
-                        extract_ref_text = gr.Textbox(
-                            label="参考文本（参考音频的文字内容）",
-                            lines=2,
-                            placeholder="输入参考音频中的确切文字...",
-                        )
-                        extract_xvector = gr.Checkbox(
-                            label="仅使用 x-vector（无需参考文本，但质量较低）",
-                            value=False,
-                        )
+                    with gr.Column():
+                        extract_ref_audio = gr.Audio(label="参考音频", type="numpy")
+                        extract_ref_text = gr.Textbox(label="参考文本", lines=2)
+                        extract_xvector = gr.Checkbox(label="仅使用 x-vector", value=False)
                         extract_btn = gr.Button("提取音频特征", variant="primary")
-
-                    with gr.Column(scale=2):
-                        extract_file_out = gr.File(label="下载特征文件 (.pt)")
-                        extract_status = gr.Textbox(label="状态", lines=2, interactive=False)
-
-                extract_btn.click(
-                    extract_voice_clone_prompt,
-                    inputs=[extract_ref_audio, extract_ref_text, extract_xvector],
-                    outputs=[extract_file_out],
-                    api_name="extract_voice_clone_prompt"
-                )
-
-                gr.Markdown("---")
-
-                # Section 2: Generate Voice from Features
-                gr.Markdown("### 2. 使用特征文件生成语音")
-                gr.Markdown("上传之前提取的特征文件，快速生成语音（无需重复提取特征）。")
+                    with gr.Column():
+                        extract_file_out = gr.File(label="特征文件 (.pt)")
+                extract_btn.click(extract_voice_clone_prompt, inputs=[extract_ref_audio, extract_ref_text, extract_xvector], outputs=[extract_file_out])
+                gr.Markdown("### 2. 使用特征文件生成")
                 with gr.Row():
-                    with gr.Column(scale=2):
-                        prompt_file = gr.File(
-                            label="音频特征文件 (.pt)",
-                        )
-                        prompt_target_text = gr.Textbox(
-                            label="目标文本（要用克隆音色合成的文字）",
-                            lines=4,
-                            placeholder="输入要让克隆音色说话的文字...",
-                        )
-                        prompt_language = gr.Dropdown(
-                            label="语言",
-                            choices=LANGUAGES,
-                            value="Auto",
-                            interactive=True,
-                        )
+                    with gr.Column():
+                        prompt_file = gr.File(label="特征文件 (.pt)")
+                        prompt_target_text = gr.Textbox(label="目标文本", lines=4)
+                        prompt_language = gr.Dropdown(label="语言", choices=LANGUAGES, value="Auto")
                         prompt_btn = gr.Button("使用特征文件生成", variant="primary")
-
-                    with gr.Column(scale=2):
-                        prompt_audio_out = gr.Audio(label="生成的音频", type="numpy")
-                        prompt_status = gr.Textbox(label="状态", lines=2, interactive=False)
-
-                prompt_btn.click(
-                    generate_voice_clone_from_prompt_file,
-                    inputs=[prompt_file, prompt_target_text, prompt_language],
-                    outputs=[prompt_audio_out, prompt_status],
-                    api_name="generate_voice_clone_from_prompt"
-                )
-
-                gr.Markdown("---")
-
-                # Section 3: Traditional Voice Clone (Original)
-                gr.Markdown("### 3. 传统音色克隆（直接使用参考音频）")
-                gr.Markdown("直接上传参考音频生成语音（每次都需要提取特征）。")
-                with gr.Row():
-                    with gr.Column(scale=2):
-                        clone_ref_audio = gr.Audio(
-                            label="参考音频",
-                            type="numpy",
-                        )
-                        clone_ref_text = gr.Textbox(
-                            label="参考文本",
-                            lines=2,
-                            placeholder="输入参考音频中的确切文字...",
-                        )
-                        clone_xvector = gr.Checkbox(
-                            label="仅使用 x-vector",
-                            value=False,
-                        )
-
-                    with gr.Column(scale=2):
-                        clone_target_text = gr.Textbox(
-                            label="目标文本",
-                            lines=4,
-                            placeholder="输入要让克隆音色说话的文字...",
-                        )
-                        with gr.Row():
-                            clone_language = gr.Dropdown(
-                                label="语言",
-                                choices=LANGUAGES,
-                                value="Auto",
-                                interactive=True,
-                            )
-                            clone_model_size = gr.Dropdown(
-                                label="模型大小",
-                                choices=MODEL_SIZES,
-                                value="1.7B",
-                                interactive=True,
-                            )
-                        clone_btn = gr.Button("克隆并生成", variant="primary")
-
-                with gr.Row():
-                    clone_audio_out = gr.Audio(label="生成的音频", type="numpy")
-                    clone_status = gr.Textbox(label="状态", lines=2, interactive=False)
-
-                clone_btn.click(
-                    generate_voice_clone,
-                    inputs=[clone_ref_audio, clone_ref_text, clone_target_text, clone_language, clone_xvector, clone_model_size],
-                    outputs=[clone_audio_out, clone_status],
-                    api_name="generate_voice_clone"
-                )
-
-            # # Tab 3: TTS (CustomVoice)
-            # with gr.Tab("TTS (CustomVoice)"):
-            #     gr.Markdown("### Text-to-Speech with Predefined Speakers")
-            #     with gr.Row():
-            #         with gr.Column(scale=2):
-            #             tts_text = gr.Textbox(
-            #                 label="Text to Synthesize",
-            #                 lines=4,
-            #                 placeholder="Enter the text you want to convert to speech...",
-            #                 value="Hello! Welcome to Text-to-Speech system. This is a demo of our TTS capabilities."
-            #             )
-            #             with gr.Row():
-            #                 tts_language = gr.Dropdown(
-            #                     label="Language",
-            #                     choices=LANGUAGES,
-            #                     value="English",
-            #                     interactive=True,
-            #                 )
-            #                 tts_speaker = gr.Dropdown(
-            #                     label="Speaker",
-            #                     choices=SPEAKERS,
-            #                     value="Ryan",
-            #                     interactive=True,
-            #                 )
-            #             with gr.Row():
-            #                 tts_instruct = gr.Textbox(
-            #                     label="Style Instruction (Optional)",
-            #                     lines=2,
-            #                     placeholder="e.g., Speak in a cheerful and energetic tone",
-            #                 )
-            #                 tts_model_size = gr.Dropdown(
-            #                     label="Model Size",
-            #                     choices=MODEL_SIZES,
-            #                     value="1.7B",
-            #                     interactive=True,
-            #                 )
-            #             tts_btn = gr.Button("Generate Speech", variant="primary")
-
-            #         with gr.Column(scale=2):
-            #             tts_audio_out = gr.Audio(label="Generated Audio", type="numpy")
-            #             tts_status = gr.Textbox(label="Status", lines=2, interactive=False)
-
-            #     tts_btn.click(
-            #         generate_custom_voice,
-            #         inputs=[tts_text, tts_language, tts_speaker, tts_instruct, tts_model_size],
-            #         outputs=[tts_audio_out, tts_status],
-            #         api_name="generate_custom_voice"
-            #     )
-
-        gr.Markdown(
-            """
----
-**Note**: This demo uses HuggingFace Spaces Zero GPU. Each generation has a time limit.
-For longer texts, please split them into smaller segments.
-"""
-        )
-
+                    with gr.Column():
+                        prompt_audio_out = gr.Audio(label="生成音频", type="numpy")
+                        prompt_status = gr.Textbox(label="状态", interactive=False)
+                prompt_btn.click(generate_voice_clone_from_prompt_file, inputs=[prompt_file, prompt_target_text, prompt_language], outputs=[prompt_audio_out, prompt_status])
     return demo
 
-
 if __name__ == "__main__":
-    demo = build_ui()
-    demo.launch()
+    build_ui().launch()

requirements.txt

@@ -13,3 +13,4 @@ spaces
 numpy
 kernels
 openai-whisper
+opencc-python-reimplemented