Agree it can be useful in some cases. But it's a little uncertain when it comes to background vocals - you can hear them, but they're quiet. Not anything like fully preserving them, as there are thousands of ways to add, mix and master background vocals, it's almost impossible to know them all. That's probably why anjok also isn't fully confident in releasing a model trained on tv tracks.
multiman said something about using an "extras model" in his use of Demucs. He may be doing Demucs with command-line. This is not to say it's for backing vocals but would be wondering if that is already in the colab version or if not could you get with multiman to see if it's possible to add it? If only for other improvements? Post referenced is here. http://literecords.com/showthread.ph...upd-2020-06-09)
Yes, that's why there is "demucs_extra" in the code :)
If you want, you can try the regular model, just remove "_extra" from the name, but I highly doubt it will be better. The extra model was trained with more songs.
Also, there is another model that sometimes provides better drums and bass (mostly on songs with a lot of acoustic drums and guitars), it's called "tasnet_extra". Just replace that in the code where you put your song name, though it takes more time to finish than demucs_extra.