Jump to content

Avatar

Facebook Research "Demucs" to separate mp3 into stems


  • Please log in to reply
6 replies to this topic

#1 mscbuck

mscbuck
  • 3 posts

Posted June 25, 2020 - 06:37 PM

Hey all,

 

I thought I'd post about a recent tool that was actually created by Facebook that has the ability to accurately separate an mp3 into its stems. I've actually been using it a ton for just practicing drums in general, but obviously would work great in the context of what we're all about here.

 

The code from their paper entitled: "Music Source Separation in the Waveform Domain" is here: https://github.com/f...research/demucs

 

Included is the Python environment needed to take a MP3 and split it into stems for drums, bass, vocals, and "other" (often the guitar). It can do this with pretty remarkable accuracy in certain cases. I've posted an example from a Metallica song. You'll see that sometimes you get some bleeding in the guitar/vocals, etc., but using this in combination with potentially the full song as the "backing track", you could probably get something really high quality. I've tended to have better luck with bass + drums when it comes to the quality of the stems that are generated. Depending on the track all of the stems won't have high quality, but figured I'd at least demonstrate an example here of what you can do with just a pretty simple install!

 

Here's a link to the stems that were generated from Spit Out the Bone (Sorry, fixed the link here)

https://mega.nz/folder/fkpi2KqS#QxcyQ9F9mixkjcq1xGSzhQ



#2 YoshiIsDaBomb

YoshiIsDaBomb
  • 81 posts

Posted June 27, 2020 - 04:04 AM

The stems download requires a decryption key.

#3 mscbuck

mscbuck
  • 3 posts

Posted June 27, 2020 - 02:32 PM

The stems download requires a decryption key.

Sorry about that, updated the link above and here it is with the key on the end:

 

https://mega.nz/folder/fkpi2KqS#QxcyQ9F9mixkjcq1xGSzhQ



#4 Shroud

Shroud
  • 477 posts
  • Location
    EU

Posted June 27, 2020 - 05:31 PM

Sounds similar to Spleeter, which also requires python and conda, except that it doesn't separate piano. Would be interesting to compare results.

#5 doa

doa
  • 300 posts
  • Location
    Northern CA

Posted June 27, 2020 - 07:27 PM

The samples sound good, does it have a GUI?



#6 ManicPixieGirl

ManicPixieGirl
  • 172 posts

Posted June 27, 2020 - 10:01 PM

I've posted ezstems.com numerous times. Get with the times. It's the same thing, only less hassle. 



#7 mscbuck

mscbuck
  • 3 posts

Posted June 27, 2020 - 10:44 PM

I've posted ezstems.com numerous times. Get with the times. It's the same thing, only less hassle. 

 

It's actually not the same thing. It's the same concept, but both have very different architectures under them (note: I am a data scientist by trade). Each probably have their advantages, but I've found Demucs to be better than Spleeter. Can see comparison among the numerous architectures here: https://ai.honu.io/papers/demucs/). However, I think in the context of Phase Shift, I'll take the higher noted quality even if there is a little bit more bleeding. Though I'll give you the ease of use of EZStems. Just offering another option out there






0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users