yl4579
commented
2 years ago I think this is the common boundary effect in real-time audio processing caused by incorrect phase alignment. The easiest way to deal with it is a weighted average of the part that it misaligns, maybe just a few hundreds of samples. It can be realized with the following code assuming you use a window of 300 samples:
buffer_weight = np.linspace(1, 0, 300)
new_wave[0:300] = buffer_weight * prev_wave[-300:-1] + (1 - buffer_weight) * new_wave[0:300]
Pathos0925
dragen1860
Would you have any optimization tips for using this with live audio? I am processing 200 ms chunks at a time. The f0 model was performing poorly with just that so I now append it to the last 2~ seconds first then chop the end after converting which gives better results, but there are still some artifacts. How can I get the end of one converted frame to match up closer to the start of the next? Right now it sounds like this https://drive.google.com/file/d/1TTvrysYW4khnQ7eJVDpJrS0OCobV4O0o/view?usp=sharing The spectrogram: https://i.imgur.com/LC1YbLI.png
Thank you!