加whisper

app.py

@@ -15,6 +15,7 @@ from qwen_tts import Qwen3TTSModel
 import functools
 import uuid
 import random
+import whisper
 # 配置日志
 logging.basicConfig(
     level=logging.INFO,
@@ -93,6 +94,14 @@ def load_model(model_type, model_size):
         attn_implementation="kernels-community/flash-attn3"
     )
 
+@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"),
@@ -408,6 +417,28 @@ def generate_voice_clone_from_prompt_file(prompt_file_path, target_text, languag
         return None, f"错误: {type(e).__name__}: {e}"
 
 
+@spaces.GPU
+def infer_whisper_audio(audio_path, model_size="large-v3"):
+    """Transcribe audio using Whisper model."""
+    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()
+    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}"
+
+
 # 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():
@@ -454,12 +485,48 @@ def build_ui():
 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.
 """
         )
 
         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.Tab("Voice Design"):
                 gr.Markdown("### Create Custom Voice with Natural Language")

requirements.txt

@@ -10,6 +10,6 @@ soundfile
 sox
 onnxruntime
 spaces
-torch
 numpy
-kernels
+kernels
+openai-whisper