Deezer's Spleeter

Deezer just released its “source separation algorithm” named Spleeter.

Its aim is to help separate tracks from music using pretrained deep-learning models (think “Artificial Intelligence”, or… just “Magic” :no_mouth:), and not by using high pass filters and so on.

So we might have new ways to remove the bass track from our favorite songs! :smiley:

It’s open source, so anyone can mess with it. Unfortunately, those “anyone” are mostly developers now since it’s not a product, it’s published “as is” and you need at least to know how to install development tools to be able to test it.

The good news is that I’m a dev and I will definitely test this in the following days. :smirk:

I’m very exited by that and I’ll post here my findings and results. :slight_smile:


That looks interesting. Let us know you findings.
Will this eventually be available to the public?


@NicolasConstant That’s pretty amazing that this technology has come this far. when you have a chance to work with it, I would love to hear what the resulting files sound like. I would also like to know how long it takes to process a single song. It says, “100 times faster than real time” (using a GPU) which sounds like being able to process a 5 minute song in 3 seconds.

(5 [minutes] x 60 [second] / 100 = 3 [seconds])

That falls in the Holly Shit category of fast. I look forward to reading whatever you learn about this.

@studio Yes, it is already available to the public but not as a finished/user friendly product. Based on what little documentation I read, it’s already functional. However, to use it now, you would need some specialized computer skills to get it working.

My guess is 1 year+ before we see a user friendly product. Since it was released with an MIT license it is open for anybody to use or even to include in retail products. I wouldn’t be surprised to see stand alone products that let you select a song, run the program, and give you a number of files back based on what you tell it to do.

For instance, tell it you want to isolate the bass, it will then give you two music files. One with everything but the bass line and one with only the bass line. That would be pretty great for learning and transcription in addition to giving you a play along track.

I also wouldn’t be surprised to see this including in lots of other products.

Okay Google, Play Bohemian Rhapsody.
Okay Eric, Playing Bohemian Rhapsody.
[music plays]
Okay Google, isolate vocals.
Okay Eric, isolating Bohemian Rhapsody vocals.
[Freddy sings a capella]

Okay, this is probably a lot more than a year away. The legalities would be a mess.


Wow cool, keep us posted! I just bought Izotops “Music Rebalance” plugin which does similar algorhythm-based source separation, curious how they compare!


Like @eric.kiser said, we’ll need to wait that someone creates an open source or commercial application that make it usable to everyone. :slight_smile:

That’s not “hard” btw (I have ideas how it could be done), it’s mostly a question of time (for the development) and money (to host it).

So far I successfully set the non-GPU version on my computer (wasn’t easy, got lots of errors and I’m not a CUDA/Python dev so I had to search a bit to fix things).
I processed “Can’t stop” and “Feel good inc” as test samples, through the 1 file => bass/drums/vocals/other files process.
You can find the result here (you’ll need 7zip to open it).

The result is more messy that I was hoping for, but I’ll have to merge those files without the bass line to see how it compete to the regular “audio range removal” technique (or commercial plugins like @JoshFossgreen’s one). :slight_smile:

If you guys want me to test some mp3 files you may own, just send them to me via Private Message (using transfernow for the file for example), I guess it may not be a good idea to share publicly the original files. :wink:


Hmm. Been kind of looking for an excuse for a mac development project. Don’t really have time at the moment though. On mobile so no github right now but what license did they use for the library?


It’s MIT licensed :slight_smile:


Oh cool, that’s good. Thanks!


This is killer.
Please oh please keep us posted. If you can get any supposed bass tracks from it, please post them so we can give a listen!
It would be cool to do it with something where the bass line is simple and easy to hear - something like My Girl - just to check it’s accuracy. If you use it on some super slammin’ fancy line, I wouldn’t know if it worked or not until I had transcribed said fancy line to compare it.


@NicolasConstant These files were pretty dirty. The Can’t Stop bass file was silent and the Feel Good Inc. bass file was pretty unusable. The other files were better but not great.

The assessment of 1+ years till a user friendly product was based on what I was reading. Based on the files you shared and actually hearing it, It sounds like they have a much longer road ahead.

Could my files have been corrupted to some degree or were you hearing the same thing?


Hey I have created a website where you can use Spleeter without the headache of setting it up. It uses CUDA GPU processing so it will process a song in 5-10 seconds.


That’s really cool, good job @ezstems!


This is really neat! I didn’t know what a stem was. It looks like the same or similar thing to a separate track for a song? Thanks!


That’s great! I knew someone would do such integration at some time. :slight_smile:


The “stem” is related to the separated track outputed by the tool:

  • Vocals (singing voice) / accompaniment separation (2 stems)
  • Vocals / drums / bass / other separation (4 stems)
  • Vocals / drums / bass / piano / other separation (5 stems)

@ezstems Nice job on the site.

1 Like

Looks really cool! Does it suitable for those uploaded music? I have synced some Spotify music to my Deezer library. Those Spotify songs are downloaded by the audfree Spotify to MP3 downloader and saved in MP3 formats.


Hi @lisabrown, welcome. This is an older thread, I am not sure if many have used it or followed it.
Maybe someone will answer you, but if nobody is using it, they may not.

Would you like to post in out “introduce yourself” thread so others can get to know you?