Pin PyAV below 19 until faster-whisper supports it

[cristianchiriac]

The pytest workflow has been failing on main since PyAV 19.0.0 was released on 2026-09-29 (e.g. https://github.com/kaixxx/noScribe/actions/runs/37302401929). Every transcription aborts with:

File ".../faster_whisper/audio.py", line 46, in decode_audio
File "av/container/core.py", line 479, in av.container.core.open
TypeError: open() got an unexpected keyword argument 'metadata_errors'

PyAV 19 removed the metadata_errors argument of av.open(). The latest faster-whisper release, 1.2.1, still passes metadata_errors="ignore". faster-whisper's master already passes it only for PyAV < 19, but that fix hasn't been released, and faster-whisper only requires av>=11. So a fresh install of the requirements currently picks PyAV 19 and breaks transcription for users as well, not only in CI.

This pins av<19 in all environments/requirements_*.txt, with a comment explaining why. requirements_win_cuda.txt had no explicit av line, so the pin is added there. The pin can be dropped once a faster-whisper release with the fix is out.

Testing

• In a fresh Python 3.13 venv with faster-whisper==1.2.1, decode_audio() on a 1-second WAV:
  • with av==19.0.1 it raises the TypeError above;
  • with av<19 (18.1.0) it decodes 16000 samples.
• uv pip compile environments/requirements_linux.txt now resolves av==18.1.0 alongside faster-whisper==1.2.1.
• I didn't run the full test suite locally, because the torch/pyannote stack is too heavy for my machine. The workflow run on this PR should show the result.

[happyarts]

Confirming from our side: the pytest runs of all our open branches fail with exactly this TypeError on av 19.0.1 + faster-whisper 1.2.1 (28 failures, all through decode_audio). noScribe's own PyAV code is fine with 19: tests/audio/test_convert.py passed in full in the same runs, so the pin is only needed for faster-whisper and can go once it releases the fix.

(Checked with help from Claude Code.)

[happyarts]

A possible follow-up that would make the pin unnecessary regardless of when faster-whisper releases: noScribe only ever hands faster-whisper its own converted 16 kHz mono WAV, so it doesn't need decode_audio at all. It is called in three places: the language detection and model.transcribe(audio_path) in whisper_mp_worker.py (with a path, transcribe calls decode_audio itself), and the Silero VAD in main.py. Reading the WAV with soundfile.read(path, dtype="float32") and passing the array instead skips faster-whisper's PyAV path entirely. On a 16-bit WAV the result is bit-identical to decode_audio, and it's the same pattern load_waveform already uses for pyannote.

This PR is still the right quick fix to get CI green. If you think the follow-up is worthwhile, I'd be happy to open a small PR for it.

(Written with help from Claude Code.)

[kaixxx]

@cristianchiriac Thank you, good catch. I haven't had the time to look into why the GitHub tests where failing, but my local ones were running fine.