fakufaku
commented
2 years ago This metric is actually also used to train systems working in the STFT domain. The signals are first brought back to time-domain by doing iSTFT. This is not really an issue. Now, my experience (which is similar to what is found in the literature), is that the time-domain metric gives better performance than those applied in the STFT domain directly.
Lots of BSS work make efforts to find coefficients in Short-Time Fourier Transform (STFT) domain, and multiply them with signal in the same domain. I think that this metric is evaluated in time domain, and wondering how can use it directly in STFT domain (maybe need to find out the frequency mapping of distortion filter?). If it worked, this metric can be directly extend to these neural network in STFT domain? Thank for your brilliant creative work!