speech mixture

[DShuaiG]

Dear friend.
I have one question. When I use this code to deal with Singing and voice mixture, it works well. But when I use it to process speech mixture, it seems to have no effect. Can you guide why？

Regards.

[leimao]

Hi @DShuaiG,
It has been a long time since I did this project. If I recall, I used the spectrogram as the input to the neural network for separation. It is likely that the frequencies of background music and human (singing) voice are quite different so that it is easy to separate with a simple RNN without too many human orchestrations. However, the frequencies of two human voices could be quite similar and therefore it will be a much harder task for the neural network to do. Assuming the data preprocessing and postprocessing you have done is correct, this neural network architecture and training configurations might not be quite suitable for speech mixture separation, although I have never tested it personally.
Best,
Lei