datasets
transformers<5.0
torchaudio
soundfile
torch
numpy

[lm]
pyctcdecode>=0.4.0
