search

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.4k stars 1.07k forks source link

Not separating background vocals #91

Open lucellent opened 4 years ago

lucellent commented 4 years ago

This may have been previously reported, but for some reason Demucs combines "Other" with background vocals from the acapella track. I don't know if this is intentional, but I think most will agree that it would be better if it doesn't do it.

Background vocals should be part of "Vocals", not "Other" (the synth track).

Do you think this can be fixed?

adefossez commented 4 years ago

Hey @lucellent , sadly I don't think this can be fixed, as there is either not enough tracks with background vocals or they were wrongly assigned to the other source in the dataset. Can you try again adding -n light_extra to the command line, just before the filename? This uses extra training data that I have curated myself, so I know the background vocals were assigned to the vocals source.

lucellent commented 4 years ago

Thank you for the suggestion. I just tried light extra but except that it's slightly worse in quality, the background vocals are still separated and merged with the synths.

iamnoob1 commented 4 years ago

Try -n light_extra --shifts=100

adefossez commented 4 years ago

The --shifts option is unlikely to change much the output. While it can remove small artifacts, in that case it seems there is a clear mistake by the model. @lucellent , do you have an example of song for which there is this problem?

lucellent commented 4 years ago

The --shifts option is unlikely to change much the output. While it can remove small artifacts, in that case it seems there is a clear mistake by the model. @lucellent , do you have an example of song for which there is this problem?

All of my songs, actually. Do I send an example song without processing or one that I've already demucs-ed?

0xBEEEF commented 4 years ago

Seems to affect not only background voices but also saxophones and other brass instruments in general. These are mostly interpreted as voices, although they are not voices at all. This is really interesting, because you slowly notice how the model differs from other instruments.