ffmpeg pre-convert inputs to wav (torchaudio 2.0 cannot decode m4a); show_error

app.py

@@ -4,7 +4,10 @@
 上传带噪音频（口播、讲座录音等），输出降噪后的 wav。
 模型在首次请求时加载并缓存，之后常驻内存。
 """
+import shutil
+import subprocess
 import threading
+import uuid
 
 import gradio as gr
 
@@ -24,13 +27,32 @@ def get_model():
     return _model, _df_state
 
 
+def to_wav(audio_path):
+    """把任意格式（m4a/mp3/flac 等）经 ffmpeg 转成 48k 单声道 wav。
+
+    df 自带的 torchaudio 2.0 sox 后端解不了 m4a/AAC，这是报错的根源。
+    若容器里没有 ffmpeg，则原样返回，让 load_audio 自己试。
+    """
+    if shutil.which("ffmpeg") is None:
+        return audio_path
+    out = f"/tmp/conv_{uuid.uuid4().hex}.wav"
+    r = subprocess.run(
+        ["ffmpeg", "-y", "-v", "error", "-i", audio_path, "-ar", "48000", "-ac", "1", out],
+        capture_output=True,
+        text=True,
+    )
+    if r.returncode != 0:
+        raise gr.Error(f"音频转码失败：{r.stderr.strip()[-300:]}")
+    return out
+
+
 def denoise(audio_path):
     if not audio_path:
         return None
     from df.enhance import enhance, load_audio, save_audio
 
     model, df_state = get_model()
-    audio, _ = load_audio(audio_path, sr=df_state.sr())
+    audio, _ = load_audio(to_wav(audio_path), sr=df_state.sr())
     enhanced = enhance(model, df_state, audio)
     out = "/tmp/enhanced.wav"
     save_audio(out, enhanced, df_state.sr())
@@ -50,4 +72,4 @@ with gr.Blocks(title="音频降噪增强") as demo:
 
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(show_error=True)