what is the purpose of "first" in _separate_frame function about denoiser HOT 4 CLOSED

One frame is 36ms of audio, but the stride between frames is 16ms, so that there is 20ms of overlap. A naive implementation would recompute the convolutions for this 20ms in common. Instead here, we only compute the convolutions for the new part of the audio and merge the two (cache from the end of the previous frame + newly computed).
first means is it the first frame of audio we are processing, in which case there is no cache from the previous frame and we need to compute the convolution over the entire input.

from denoiser.

thanks for your quick response!
why does it need overlap? is it for the resampling or reducing latency?
thanks!

from denoiser.

Overlaps improve the quality of the signal, I can't give a specific reason but it is often the case with convolutions. If you take spectrogram models, the STFT overlaps by 50 to 75% usually, so that the prediction for each time step is going to be averaged from the prediction of multiple frames.

from denoiser.

i got it, thanks for your response! i will close this issue.

from denoiser.