Does this model support training with a 48k sampling rate?

[WangQin-wq]

I converted the dataset to 48k and modified the code, and it can train normally, but the results on the test set are incorrect

[DM-Rookie]

It should be supported in theory. Are the loss and SDR on the validation set looking good for 48kHz compared to 44.1kHz? If so, take a look at the inference part. The current code downsamples everything to 44100.

       # Convert to model's expected format
       mix = convert_audio(mix, sample_rate, 44100, self.separator.audio_channels)