jams files

[xu-shitong]

Thanks for the great work! I have a question about audio loading.
In the provided datasets, the audio is loaded using scaper.generate_from_jams function. However, when I load audio directly using torchaudio.load, the model performance drops significantly from around 7 si_snr_db improvement to around 2. Have you encountered a similar problem? And is there particular reasons for loading audio using jms file instead of directly loading the audio?