details about the noisy_train.mat and target_train.mat

[zyy-fc]

Thank you for your share! However, I want to know the details about how to constructi the "noisy_train" and "target_train". What is the meaning of "NUM_TOKENS" ? What about the NUM_CHANNELS in "noisy_train" and "target_train" ??

[zyy-fc]

I am really impressed by your work and got a few questions in terms of how you process the training data.

Actually, I can not understanding the meaning of "Dimension: [24570, NUM_TOKENS]" in readme.md

I guess the 24570 represents the length of noisy signal. But, how you concatenate the multi-channel signals?

[auspicious3000]

NUM_TOKENS is batch_size. The enhancement model is single channel, so you can use any speech enhancement method as a drop in replacement. 24570 is the length of noisy signal based on the model structure. You can refer to the following paper for more details. 《Speech enhancement using Bayesian Wavenet》 The input data of the beamforming algorithm is multichannel. That's where NUM_CHANNELS comes into play.

[zyy-fc]

Thank you for your reply.

1. To train the enhancement model, the input should be single-channel. However, is the input to enhancement model the one channel signal of multichannel signal into beamforming ?
2. Is the multichannel signal not used in training stage ? According to the section "Enhancement Network Configurations" in your paper, the simulated data are multichannel signals.
3. Could you tell me the value of your batch_size? I have read the code of 《Speech enhancement using Bayesian Wavenet》. Thank you for sharing ! I guess the batch_size is 312, so you just used 312 utterances as training set ?
4. Could you share your "noisy_train.mat" and "target_train.mat"? Thus, I can understand the complete process better. I am new to deep learning. So I am honored to be able to communicate with you. Thank you again!

[auspicious3000]

1. The enhancement model is pretrained independently using single channel signal simulated under different SNR and room configurations. It can be any general enhancement model.
2. See above.
3. The batch size is 1 due to memory limit.
4. I am unable to share the training files because they are too large. I am sure you can create the training files by simply following the paper and readme. The deep learning part of this work is very straight forward. The key insight is the iterative process between two classes of methods.

Hope this helps.

[zyy-fc]

Could you tell me the version of tensorflow that you used?

[auspicious3000]

I was using tensorflow 1.3, but I think it will probably work with 1.4 or higher. Since the beamforming part is not implemented in Tensorflow, you could virtually use any framework.