sampling rate used in training

[vra]

Hi @Hertin @BradyFU, I hope this message finds you well.

I noticed that snac decode script was uploaded into this repo, in which a sampling rate of 24000 is used to dump audio waves. However, the sampling rate in training scripts is set to 16000, which is different from previous setting. I guess the training script need to be updated to match the real sampling rate of snac audios? If so, is there any other parameter to update to reproduce the LUCY pretrained model? I really appreciate your help.

[Hertin]

I downsampled all the audios to 16khz before loading them using the dataloader. You can add a resampling step using torchaudio/librosa etc in the dataloader yourself.