dsh: convert any audio/video (m4a/mp3/mp4/aac) to WAV via ffmpeg before whisper; verified m4a + mp3 transcribe

dsh/app.py

@@ -411,6 +411,32 @@ async def _whisper(data: bytes, fname: str, mime: str) -> str:
         return (r.json() or {}).get("transcript", "") or ""
 
 
+def _to_wav(data: bytes, fname: str) -> tuple[bytes, str]:
+    """Convert arbitrary audio/video (mp3, m4a, aac, mp4, ogg…) to 16kHz mono WAV
+    via the container's ffmpeg, so whisper always receives a plain WAV. Returns
+    (original data, original name) untouched if already WAV/FLAC or on failure."""
+    ext = "." + fname.rsplit(".", 1)[-1].lower() if "." in fname else ""
+    if ext in {".wav", ".flac"}:
+        return data, fname
+    tmp = f"/tmp/conv_{uuid4().hex}"
+    src = tmp + (ext or ".bin")
+    Path(src).write_bytes(data)
+    out = tmp + ".wav"
+    try:
+        r = subprocess.run(
+            ["ffmpeg", "-v", "error", "-y", "-i", src,
+             "-ar", "16000", "-ac", "1", "-c:a", "pcm_s16le", out],
+            capture_output=True, timeout=300)
+        if r.returncode == 0 and Path(out).is_file() and Path(out).stat().st_size > 100:
+            return Path(out).read_bytes(), "converted.wav"
+        return data, fname  # fall back to raw (whisper's own ffmpeg may still decode)
+    except Exception:  # noqa: BLE001
+        return data, fname
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+
+
+
 @app.post("/api/tool/transcribe")
 async def transcribe_audio(file: Annotated[UploadFile, File()]) -> dict:
     """Speech -> text. Save the audio into the workspace, then whisper it."""
@@ -423,7 +449,8 @@ async def transcribe_audio(file: Annotated[UploadFile, File()]) -> dict:
     audio_path = up / fname
     audio_path.write_bytes(data)
     try:
-        transcript = await _whisper(data, file.filename or "audio.wav", file.content_type or "audio/wav")
+        wav, wav_name = _to_wav(data, file.filename or "audio")
+        transcript = await _whisper(wav, wav_name, "audio/wav")
     except Exception as exc:  # noqa: BLE001
         return {"ok": False, "error": f"Whisper failed: {exc}", "path": f"uploads/{fname}"}
     if not transcript.strip():
@@ -464,7 +491,8 @@ async def ask_file(path: Annotated[str, Form()], question: Annotated[str, Form()
 
     if ext in AUDIO_EXT and ext not in VIDEO_EXT:
         try:
-            transcript = await _whisper(fp.read_bytes(), fp.name, "audio/wav")
+            wav, wav_name = _to_wav(fp.read_bytes(), fp.name)
+            transcript = await _whisper(wav, wav_name, "audio/wav")
         except Exception as exc:  # noqa: BLE001
             return {"ok": False, "error": f"Whisper failed: {exc}"}
         if not transcript.strip():