Thanks to AI, Deezer has made an open source (FREE) python library called "Spleeter" which can split songs into either into either

• 2 - Vocals + the rest,
• 4 - Vocals + Bass + Drums + the rest, or
• 5 - Vocals + Bass + Drums + Piano + the rest

(number) of STEM tracks. I just found it out when I was watching a video on yt. I know how shit DIY mp3 acapellas sound, so I igonored it once. Again it popped up in my recommendations. I thought lets check, and indeed it was truly amazing. I mean the hours and hours required to extract an acapella with Kn0ck0ut, etc. are gone. Now since it has been released only 2 weeks ago, most people don't yet know it. There is not any application/plugin made for it yet. Its a way long process to get it working on our PC for most of us.

That problem too is solved, go to melody.ml and your song file. It will take 5-10 mins. max depending on how long the song is, where you live, server load etc. Melody.ml has 2 and 4 stem tracks option only. Artifacts are truly less considering only MP3 can be uploaded and stems will be in MP3 ofc. No bitrate limit. I quickly dropped a few 192kbps YT rips and the stems I got were far better than any DIY acapella I have ever heard.

And the quality of this will keep on improving automatically as this is AI-powered. And you can give your email id to get notified when your tracks are ready. You can also check the website any time later to see if it is done. RIP to acapella performers.

Links
https://melody.ml/
https://deezer.io/releasing-spleeter-deezer-r-d-source-separation-engine-2b88985e797e
https://github.com/deezer/spleeter

 

Hey, bro, how you doing? Some time ago someone posted a post about this. I'm learning python and I have it installed on the PC. So I went to the page, download Anaconda (mini) and another app. The case is that in principle I throw error of lack of dll in the folders of system 32 and wow64...I could correct that error. Then I download all the complements of spleeter. And I tried to separate the tracks but in the DOS of anaconda throws me many errors . And when I go to output . There are two tracks and they sound the same. It was a headache for me and I quit.

 

Installed this library when it came out. Install with conda. Only thing I dislike is that it takes a 4-5 min for one track. If you have a good gpu, use the library with your gpu, the extraction process will be a lot faster, or use docker. And remember shit in is shit out, meaning dont throw mp3 files at it, but use youtube-dl, spotify, or what ever source file in .wav format.

 

DIY acapellas can be better because spleeter has 11k bandwidth on its default models.

 

really liking this

 

Hi, I have created a website where you can use Spleeter without having to set it up. Check out https://ezstems.com and let me know what you think!

 

I tried to use these methods and this is my review:

https://ezstems.com
When I upload song there is a message saying: SyntaxError: Unexpected end of JSON input
The screen remains stuck on “processing your song…” with wheel turning: after 10mins I closed the page: no way to have stems in here

https://melody.ml/
Site is quite fast and the algorithm certainly guesses the different sources, but the result is terrible: I tried 2 songs and the stems are really bad in quality.
After 2 conversion the site asks for money to continue uploading songs

https://github.com/deezer/spleeter
too complex for me

 

Spleeter will be intergrated in the next version of Acon Digital Acoustica, it will be easier to use:

 

Worked for me.. Not bad vocal split will try some more tracks

 

Wow, this is great news, thx !

 

You know what, sometimes this fraud phonicmind works better for me.
Hehe

 

Finally I managed to use Ezstems:
Quality is not so bad, stems have several artifacts (this is due to the algorithm) but I must say that I am amazed at what I can get out of it.
Nothing can ever replace the quality of the stems taken from the original master, but in the end this website offers an interesting service

 

iZotope RX7 Rebalance do the job too

 

Facebook is also working on an open source "demixing" tool based on deep learning, python code is available here: https://github.com/facebookresearch/demucs

 

https://moises.ai/

Guess I should add some explanation...
Account is "currently" free and it saves all the files you have separated under your account. Also, you can use links to youtube videos as your sound source.

 

Last edited: Jan 16, 2020

Thank you to those that tried https://ezstems.com and said good things.

Sorry about your first failed attempt, I had an influx of users which temporally caused problems on the server. I have now patched a few things and am running a larger server instance.

I am constantly working on https://ezstems.com (legit every night), creating new features and adding functionality from user feedback.

If anyone has problems or feedback please let me know. Also I feature an artist on the homepage every week so send me a link to your work if you would like it featured.

I soon hope to implement a CUDA processing server which will process songs 100x faster.

Thanks everyone.