Yep, that's what they use

 

izotope have done this too. from my minimal experience using it i still think its better used in combination with other techniques to get things right, but it has definitely made things a whole lot easier and saved many hours for people who have a use for it.
personally i dont have much use for tools like this though.

 

The extra khz is fake/ extrapolated since the models are still 11k. If you want REAL 22k, use https://github.com/facebookresearch/demucs , or import this into google colab. It's not better all the time, but it's an alternative.

Can't remember what I used, but both songs in my mashup https://hearthis.at/jzgmyrcb/guru-kaze20e/ were done using ML. This is a particularly hard song to separate.

 

okay... please try to understand:
It is for those people who can not find seperate recordings.

 

works like a charm, I created a Automator Workflow which makes spleeter available as service in the finder

 

Could you please share this workflow service?

 

Mostly with the right tools you can simply extract audio files from games! I do this often to get some good sounds.

 

Last edited: Mar 15, 2020

Yeah! iZotope did some AI source separation before but compare to Spleeter, the latter is on a different league.

Neat idea! I would like to hear more on how this automation thing works as I'm not macOS user. Is it like a one-click macro that executes a set of commands?

 

Agree. This thing works much better on many songs (sounds worse on others), but it isn't as popular as Splitter for... whatever reason

 

interesting, i will have to try it out. would you consider it usable on its own?

 

@tun yeah it's very easy and stable. Just use one of the many web services we talked about on this thread to get a taste. The training models they used are based on freely available multitracks on the internet. What this means is that even though these AIs were not trained with commercial songs, they still do a pretty good job at extracting stems out of a full mix.


UPDATE: I checked out Facebook Research's demucs, as mentioned by @junh1024, and the models are the same. But no one has made a web service for that one yet. And it is waveform based extraction unlike Spleeter, which is spectral based.
No good or bad. Just different solutions for solving the same problem. I always liked having choices!

 

Last edited: Mar 15, 2020

The future is bright for AI separation/isolation and not so bright for us musicians and engineers. Because in 10 or 20 years, every kid on the planet will be able to deconstruct your mixes on their iPhones down to that ugly reverse cymbal you always put before the chorus to mask your inability to write a drum fill.
But hey... get over it!

Instead of fighting back and suing them, give away the music and the stems for free and charge them 10 times more for a dongle to play it. Problem solved!

 

I guess technology is a double-edged sword

 

I

Interesting, how did you do it?

 

Hi, I know you're on a Mac, but maybe you could still help a clueless non-coder on a windows box:

Installed Anaconda, tried installing with conda install -c conda-forge spleeter-gpu, the thing coughed up "Solving environment: failed with initial frozen solve," ground some more checking installed packages, errored out. Probably a dumb move, but tried installing with pip, that seemed to work, but
TL;DR [from the console]:
"Successfully installed spleeter-gpu-1.4.9 tensorflow-gpu-1.14.0

(base) C:\Users\me>spleeter separate -i spleeter/C:\Users\me\Desktop\audio\some_song.wav -p spleeter:2stems -o output
INTEL MKL ERROR: The operating system cannot run %1. mkl_intel_thread.dll.
Intel MKL FATAL ERROR: Cannot load mkl_intel_thread.dll."

Any ideas?

P.S. Could this be it (finding wrong version)?

 

Last edited: Mar 16, 2020

thats awesome. i have not really kept up with these technologies, i didnt realise they had come so far. to me, the idea of doing this kind of thing well using non-spectral techniques blows my mind, i had no idea we were there already :D

i have been reading this thread assuming the results were pretty much the same as izotopes. even if i have no use for it i will be trying out some tests to see what the hype is about.

there is a remix i could do of a very old track for my wife that would almost definitely get me laid :D

 

its not just a reverse cymbal! i also reversed the sound, applied reverb, then reversed it again!!! who showed you my music anyway???

 

@Lock it down I'm on Windows but I have no idea what's causing that error.
If you're not comfortable with commands, maybe you should have a look at https://ezstems.com It's free mate!

The main reason for wanting to install Spleeter on your computer is to train your own models. But it's not easy as clicking a button and importing WAV files. You will need to use another program for encoding the data and then input some more commands to introduce the new models to Spleeter.

 

We assume and expect that you upload your own music only. WTF DOES THAT MEAN?!
So called "hypocrite warning"?

OF COURSE the only aim and reason behind this kind of technology consist in extracting instruments and vocal tracks from your favourite tunes. First song I uploaded was "Barrel of a gun" (Depeche Mode), just because (LOL). By the way, still "meh" results, alas. Nice try, though. It reminds me of all the "Celemony DNA" hype, some years ago. Still dreaming for now.