Quote Originally Posted by Anjok View Post
That's a good question and it's something I would have to experiment with. I think one way I can potentially make this happen is to train the AI on a dataset consisting of only full mixes paired with their official TV track counterparts. To have a model as effective as the ones I've shared, it would have to consist of at least 200 pairs as well. The GUI that's being developed will have a drop-down of models to choose from, so you'll be able to toggle between a karaoke model and a full vocal removal model.
That would be absolutely fantastic. Some songs leave backing vocals currently already but again, it depends on the mix. As for the earlier mention songs someone mentioned with talking parts, I have one such old country classic I'd like to get devocaled. I may have to send it to you since I bought a retail CD in order for you to have the full frequency of the track. But it's called "Phantom 309." Really would mean a lot to get it devocaled properly.