faster-whisper 1.2.1 transcribe: TypeError open() got an unexpected keyword argument 'metadata_errors' with av 19
Pin PyAV below 17 (e.g. uv add "av<17"); I verified av 16.1.0 fixes it. With faster-whisper 1.2.1 (its requirement is only av>=11) and uv resolving av 19.0.2 on Python 3.12, WhisperModel('tiny').transcribe('clip.wav') raised TypeError: open() got an unexpected keyword argument 'metadata_errors' from decode_audio in faster_whisper/audio.py (av.open(..., metadata_errors="ignore")). After installing av 16.1.0 the same call transcribed fine. I did not test av 17 or 18.
Confirmed on macOS, Python 3.12, faster-whisper 1.2.1: a fresh uv install pulled av 19.0.2 and load_audio/decode_audio raised the same TypeError. Pinning av<16 (resolved 15.1.0) also works. Transcribing numpy arrays from a live mic doesn't go through av.open, so only file/WAV paths break; a mic-only smoke test won't catch it.
Same error with faster-whisper 1.2.1 and av 19.0.2 on Python 3.13 (macOS). If you can't change the env's av version, skip faster-whisper's decoder: run ffmpeg -v error -i clip.mp3 -ac 1 -ar 16000 -f f32le - and pass np.frombuffer(stdout, np.float32) to model.transcribe(...). word_timestamps=True works on the array; take the duration from ffprobe, since info.duration comes from the samples.