FIX: pre-encode mp3→wav + preloaded audio dict for pyannote (FFmpeg 8 / torchcodec 0.7 broken)

ROOT CAUSE
- whisperx venv has torchcodec 0.7.0 + PyTorch 2.8.0 → torchcodec dylibs target
  FFmpeg 4-7 (libavutil.56-.59) but Mac brew now ships FFmpeg 8 (libavutil.60).
- All audio loading via torchcodec silently fails (warning but hangs/returns 0).

FIX
- /transcribe endpoint: pre-encode uploaded mp3/webm → 16kHz mono WAV via
  subprocess ffmpeg (whisperx.load_audio then reads clean PCM file).
- _do_transcribe diarize branch: pass {'waveform': tensor, 'sample_rate': int}
  dict to pyannote Pipeline instead of file path → skips torchcodec entirely.

Tested
- chunk_0.mp3 (4MB, 5min): diarize=0 → 200 OK + hallucination filter triggered
- Confirmed working in live HTTP POST at 15:40 HKT

transcribe_server.py

@@ -62,12 +62,31 @@ async def transcribe(file: UploadFile = File(...), language: str = "zh", diarize
     try:
         shutil.copyfileobj(file.file, tmp)
         tmp.close()
+        # BUGFIX (2026-08-11): torchcodec 0.7.0 in venv can't load FFmpeg 8 dylibs
+        # (libavutil.59 missing — brew only has .60). Re-mux to 16kHz mono WAV
+        # via subprocess ffmpeg so whisperx.load_audio gets a clean PCM file.
+        wav_tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        wav_tmp.close()
+        try:
+            r = subprocess.run(
+                ["ffmpeg", "-y", "-i", tmp.name, "-ac", "1", "-ar", "16000",
+                 "-acodec", "pcm_s16le", wav_tmp.name],
+                capture_output=True, timeout=120
+            )
+            if r.returncode != 0 or not Path(wav_tmp.name).exists() or Path(wav_tmp.name).stat().st_size < 1000:
+                raise RuntimeError(f"ffmpeg re-encode failed: {r.stderr.decode()[:200]}")
+            audio_path = wav_tmp.name
+        except Exception as e:
+            log.warning(f"WAV re-encode failed ({e}), trying raw mp3")
+            audio_path = tmp.name
         # Run CPU-heavy whisperx in thread to not block health endpoint
         loop = asyncio.get_running_loop()
-        result = await loop.run_in_executor(None, _do_transcribe, tmp.name, language, do_diarize)
+        result = await loop.run_in_executor(None, _do_transcribe, audio_path, language, do_diarize)
         return result
     finally:
         Path(tmp.name).unlink(missing_ok=True)
+        if 'wav_tmp' in dir() and Path(wav_tmp.name).exists():
+            Path(wav_tmp.name).unlink(missing_ok=True)
 
 
 def _do_transcribe(audio_path: str, language: str, do_diarize: bool):
@@ -83,8 +102,16 @@ def _do_transcribe(audio_path: str, language: str, do_diarize: bool):
     segments_with_speakers = []
     if do_diarize:
         try:
+            import torch
             diarize_model = get_diarize()
-            diar_segments = diarize_model(audio)
+            # BUGFIX (2026-08-11): pyannote's torchcodec can't load (FFmpeg 8 only has
+            # libavutil.60, venv needs .59). Pre-load audio as tensor dict to bypass.
+            if isinstance(audio, np.ndarray):
+                audio_input = {"waveform": torch.from_numpy(audio).unsqueeze(0).float(),
+                               "sample_rate": 16000}
+                diar_segments = diarize_model(audio_input)
+            else:
+                diar_segments = diarize_model(audio)
             result = whisperx.assign_word_speakers(diar_segments, result)
             for seg in result["segments"]:
                 segments_with_speakers.append({