Troubleshooting¶
TorchAudio deprecation warnings¶
When running diarization, TorchAudio may print deprecation warnings about
list_audio_backends or TorchCodec. These warnings are expected with the
compatible diarization dependency stack and are not fatal.
torchaudio missing AudioMetaData¶
If diarization fails because torchaudio has no AudioMetaData attribute,
refresh the optional diarization dependencies:
uv lock --upgrade-package torch --upgrade-package torchaudio
uv sync --extra diarize
PyTorch checkpoint safe globals¶
PyTorch 2.6 and newer default to safer weights_only checkpoint loading.
whisper-smith registers the trusted pyannote checkpoint metadata classes it
needs before loading the diarization pipeline.
Hugging Face model access¶
If the diarization model cannot load, check that HUGGINGFACE_TOKEN is set
and that the model’s Hugging Face user conditions have been accepted.