3 ms·
Open-source AI tool quickly isolates the vocals in any song
- DyslexicAtheist 7y agooriginal deezer PR post (didn't get much attention here): https://news.ycombinator.com/item?id=21456673 https://news.ycombinator.com/item?id=21456673 spleeter project: https://github.com/deezer/spleeter https://github.com/deezer/spleeter
- detaro 7y agoDiscussion of the project, 175 comments: https://news.ycombinator.com/item?id=21431071 https://news.ycombinator.com/item?id=21431071
- ksaj 7y agoI'd thought about this kind of process: Take a stereo image, and subtract left from right to make a mask. It'll essentially be the sound from anything that wasn't centered (usually main vocals and a few other things that are centered). Then use that resulting file as a mask for both the right and left again, which will isolate the middle, leaving just the vocals and some artifacts from other centered audio. One of these (the left or the right channel) is likely to be cleaner than the other, depending on how/where the stereo stuff was mixed. Listen or analyze them to determine if one is better than the other, and then continue working on that one. With the better of the two resulting "center" tracks, use this AI tool to clean up any remaining artifacts that don't belong there. In a way, it's a hybrid analog/digital method of vocal extraction, but it would probably produce a lot cleaner of a vocal track. And of course if you wanted a good stereo backing track to sing Karaoke over, you can use that clean vocal extraction as a mask for both the original right and left channels, which in theory would produce a better karaoke file than the original centered-audio elimination. I think this would in essence be a modified/improved version of Dolby 3-Stereo.