Dabi Ahn
Dabi Ahn
@sayelaoda My environment was linux and mac.
@vrosato The loss graph I've got was a lot like that of yours. When you set the smooth rate to 0.99, you probably see a decreasing loss graph. Actually I...
@vrosato I used only ikala (was better than mixing two) and I remember that over 20k steps training was enough to get a generalized model. I hope you get a...
@balenko1992 yes i did, but the amount of each is not the same like that in the paper.
@balenko1992 hi. I don't load all the dataset to gpu at a time. Instead, I sample a proper number of wavfiles for each step. one epoch is done by repeating...
@balenko1992 because I consider the training as step-based, epoch is not much meaningful. but if the number of data is 100 and NUM_WAVFILE=10, one epoch would be 100/10 = 10...
@dankorg I don't get it. what does 16bit or 32bit mean?
@dankorg oh, I understood. I updated the code and then it will write wav file with 16bit/16kHz ;) Thank you for the issue.
@sigpro The same architecture can be used for removing noise from speech or extracting single speech from dialog as well ;) Take a look at this reference paper: https://posenhuang.github.io/papers/Joint_Optimization_of_Masks_and_Deep%20Recurrent_Neural_Networks_for_Monaural_Source_Separation_TASLP2015.pdf
@sigpro yes i'll it if have some dataset ;) Do you know any dataset that has multi-channel ?