search

seorim0 / DCCRN-with-various-loss-functions

DCCRN with various loss functions
MIT License
91 stars 22 forks source link

Hi,excuse me,I want to konw why the SI-SNR loss doesn't decrease when I used reverberation dataset? #8

Closed yyd19948 closed 1 year ago

seorim0 commented 2 years ago

Hi. Could you please explain a little more? I don't know the correlation between using reverberation dataset and not reducing the SI-SNR loss.

yyd19948 commented 2 years ago

When the training data has no reverberation, the SI-SNR can be reduced to -17, but after adding the reverberation to the training data, it will be reduced to 6 and will not move.

When the training data has no reverberation, the SI-SNR can be reduced to -17, but after adding the reverberation to the training data, it will be reduced to 6 and will not move.

seorim0 commented 2 years ago

The loss range may differ because the training dataset has become more difficult for DNN to learn as reverberation is added to the training dataset. What was the result of the training? Was the model not able to reduce noise or reverberation?

lixinlun commented 2 years ago

it is not easy to answer his question so far, but i do remember, that the train-dataset in the original paper DCCRN(by tan) has reverberation. "In detail, at each training epoch, we rst convolve speech and noise with a room impulse response (RIR) randomly-selected from a simulated 3000-RIR set by the image method [32], and then the speech-noise mixtures are generated dynamically by mixing reverb speech and noise at random SNR between -5 and 20 dB."

yyd1994 may change the train-dataset very slowly and careful and observe the difference, which might be a big project...........