Anyone know of a tool (or tools) that can take an existing song and output the stems? For example, separate voice, guitar, keyboard, drums, etc into separate files
2 Likes
MPC has it now, Serato has it, there plugins for your DAW, Ableton currently has it in beta and should be released soon.
2 Likes
I use spectralayers 12 pro myself.
My favorite of what’s out there but remember it’s also your audio source. Some songs can track the band with two mics into a recorder and it will be harder to separate those sounds.
2 Likes
I’m sure there’s another thread that’s been going for ages with stem discussions in it, stem roller / traktor / dj pro / logic / etc etc
2 Likes
Koala Sampler on iPad can do this.
4 Likes
Stem roller, but I’m not positive that it’s available anymore. I read that it was no longer supported but that was here on elektronauts and it was a while back, so I don’t know if the information is credible.
The MPC stems output seems kinda weak compared to stem roller. What I’ve heard from serato sounded ok.
You can do it for free online but the output is not always the best and sometimes you have to add a bunch of silence to the end or sometimes the beginning because they want you to pay for full versions so they only output a percentage of it, but since it doesn’t truncate the file at the silence you can cheat the system by making your file overly long so the percentage which is output contains your entire audio stream.
If I can think of some others, I’ll update this but that’s what I can think of at the moment. I’ve never tried some of the others mentioned so can’t comment.
1 Like
Here you go:
1 Like
If you’re handy with Python at all (or really just know how to handle installing and resolving dependencies), you can use something like audio-separator to DIY. It’s not going to produce perfect results for every stem (depends heavily on the source track), but can certainly produce usable results, especially if you make multiple passes using different models.
An example from tonight: I used the melband_roformer_instvox_duality_v2 model to split the original track (an early 80s Grace Jones track FWIW) into vocal and instrumental tracks, then used htdemucs_6s to break the instrumental into multiple parts, and then used MDX23C-DrumSep-aufr33-jarredou to break the drum track from the prior stage into multiple parts. The vocal separation was very good…grabbing the slight echo and the reverb on the vocal as well, leaving a pretty clean instrumental. The other stages had various artifacts that could be mostly sorted out with manual edits…ex: it has trouble determining the difference between a tom and a snare or kick, similar story for the hats and the snare, and some of the low end from the kick ends up on the bass track in a few places.
Caveat: It helps if you’ve got a decent GPU to leverage. Still doable with CPU only, but you’ll be waiting much longer for the processing to finish at higher quality settings.
1 Like