Radiobuzz wrote:Deezer's new tool, Spleeter, can separate several audio tracks from mp3. Is anyone here willing to try it out? It can separate up to 5 'stems', I started trying it on Tomorrow's Harvest on my office but it makes my crappy PC crash. It would be interesting to hear what can come out of it, maybe we can find something interesting.
https://github.com/deezer/spleeter
What this does differently is use a machine learning algorithm trained on a huge dataset of music to be able to know the difference between sounds in a way that goes beyond the usual things like stereo placement and frequency. A piano and a human voice overlap all over the place but this splits them apart pretty cleanly. Now, that model is more useful for, you know, your standard pop or rock track. But you can train the model yourself if you have enough source material to put through it. (Which I don't). So theoretically, this could be even better if you wanted to really invest the time training it.
I've run a few tunes through this already, and yeah it is interesting, though I've not come up with anything new from BoC's stuff that we haven't already heard. That's not because it's not any good, like I say - I've been making mashups for years and this thing is INSANELY good at splitting things that were pretty much "forget it" till now - it's just that most stuff has been discovered already. It works better on stuff ripped straight off the CDs, MP3 encoding leaves a bunch of psycho-acoustic compression glitter behind.
What it did do was confirm a bunch of vocal samples that had already been found on BoCpages.