search

JusperLee / Dual-Path-RNN-Pytorch

Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation implemented by Pytorch
Apache License 2.0
417 stars 65 forks source link

About computing loss -SI-SNR #32

Closed Reid-Cui7 closed 2 years ago

Reid-Cui7 commented 3 years ago

Thank you for your great work in Speech Separation, I've been benifited a lot from your awesome-speech-separation repository. I am a little confused when computing loss: In model->loss.py->function sisnr(): you return 20log10(...）at last, but when computing Si-SNR, we should use 10log10(...), should we ? I can achieve about 19 in Si-SNR according to your code, but if I change to 10, the result falls

JusperLee commented 2 years ago

if you look the sisnr equation, you will find this equation: 10log()^2=20log()

Reid-Cui7 commented 2 years ago

OH, man! How silly I am, I just focus on the factor, did not get into it. Thanks for your patience