datasets>=1.12.0
torchaudio
torch>=1.6