xuchenglin28 / speaker_extraction_spex Goto Github PK

Hi author! I want to be a little sure!
When we test a brand new sample, we don't have the reference speech. Are we going to use the audio mix as the speaker encoder input?

Hi Author,

Do you have any plan or schedule to share your implementation of SpEx and SpEx+?

Hi! Thanks for this great wok :D ! I have a question: When the number of speakers in classification task is too much (about more than 2000 speakers), will it affect the performance of the entire model !

I have also trained the model following the step you mentioned in README, and use the default setting except the parameter of batchsize (10 for 2 GPU). The SI-SDR was about 12dB , witch is different with the result in the paper.

Hi, thanks for the great work. I have some question about the results. I have trained the model following the step you mentioned in README, and use the default setting except the parameter of batchsize (24 for 2 GPU). The SI-SDR was about 12dB , witch is different with the result in the paper.