naplab/danet

Deep Attractor Network (DANet) for single-channel speech separation

danet's Issues

I only have timit dataset,how to implement this code?

hello,your code is so brief...but I only have timit dataset ,so how to implement this code？thank you for your prompt reply！！！

Data generation code

hi,
I wanted to try out the code, how can I generate the data, is the code for data generation also uploaded to GitHub

read the h5

I meet the problem that the dataset cannot be read...
Is it the dataset should be converted to the HDf5 so that the dataset can be recognised？but I don't how to convert the dataset... can you help me?

The trained model is not working well.

Thanks very much for your code! That really help my a lot. I generated the data set according to your paper requirements. I applied the data to the training model code you provided and found that it did not achieve the expected results. The Mixed audio file cannot separate even a little. The separated audio file has only the difference in sound size from the original file.

Deciding threshold re keep top 90% T-F bins

I am bit confused in how to compute threshold to keep 90% of T-F bins.
I have tried
threshold = min(a) + 0.1 (max(a) - min(a)), where a is matrix of
but it does not work well if max value to too large. Could you please give some detail on this?