How to make training samples to let the model have the ability to dereveberate and get a stable model

[gugogi]

HI Xiaobin, I try to use your code to train denoise and dereveration model. But I find my model's output is not stable（the voice is not continuous） when SNR is low .
I try to make training (samples = clean * rir + noise) and (label = clean * rir[:100]) . Is this operation right? I want to know how you make these kind of samples ， thank you ! Another question: would you mind sharing your loss when training the model? Thanks a lot!

[Xiaobin-Rong]

Hi, your operation looks correct. It might be normal for the voice to be over-suppressed when the SNR is low, as this is a common issue with DNN-based SE. By the way, how low is your SNR?

[AlanCc1994]

Hi, your operation looks correct. It might be normal for the voice to be over-suppressed when the SNR is low, as this is a common issue with DNN-based SE. By the way, how low is your SNR?

对于低SNR及小信号的情况下，语音容易被过度抑制，无论是降噪任务还是回声消除任务，如何改善这种情况？请赐教哈哈