The output wav file of the speech enhancement task is distorted.

Thank you for all your help. I tried the vctk_noisy voice enhancement task in egs2. However, the output wav file (e.g. espnet/egs2/vctk_noisy/enh1/exp/enh_train_raw/enhanced_tt_2spk/logdir/output.1/wavs/1/p232_001.wav) is very distorted.

The loss is decreasing, so it seems to be learning.

Is there something wrong with my settings? I would like to know why the audio is distorted and what to do about it.

[Emrys365]

@Emrys365 Thank you very much for your reply.

I have figured out why the output audio is distorted. If I change inference_args="--normalize_output_wav true" to inference_args="--normalize_output_wav false" in enh.sh, the output audio is distorted.

What does normalization mean?

I see. You could check out espnet2/bin/enh_inference.py to understand how it works. Basically, it rescales the estimated audio to be in range [-1.0, 1.0]. If your model is trained with some scale-invariant loss, it is likely the model output will have a large magnitude, and normalization is needed in this case.

[Emrys365]

what is the amount of training data (say ____ hours)?

The amount of training data is about 8.5 hours.

I think it is too small. If possible, please increase the amount to at least tens of hours.