Spleeter. Isolating bass on audio recordings

[Woodinblack]

I only recently (which recent is yesterday) heard about spleeter, although it has been round for a couple of years. I would be very surprised if it hadn't taken off hugely if it wasn't a massive pain to install.

Spleeter is an AI bit of software (literally a tensorflow model if that means anything to you) which you put audio through and it attempts to split the sound in the component instruments. The amazing thing is that it doesn't do that bad a job of it and is thus, pretty close to magic.

Unfortunately it is one of those python 'right version of python, right version of every package manager, right version of everything else and a guess at what the errors mean' but after failing to get it to work on the PC, I got it to work on the mac tonight.

I actually did it because I wanted to isolate the bass line from 'learning to fly' by pink floyd, as the bass is a chapman stick. It does a fairly good job except it does tend to remove the higher parts.

I tried 'fool for your loving', on it as more posts have been done on this forum about it than anything else. Its a command line thing, so it works like this:

% spleeter separate -o ~/Music/audio-out -p spleeter:4stems  /Users/woody/Music/iTunes/iTunes\ Music/Whitesnake/Best\ Of\ Whitesnake/01\ Fool\ For\ Your\ Loving.m4a
INFO:spleeter:File /Users/woody/Music/audio-out/01 Fool For Your Loving/bass.wav written succesfully
INFO:spleeter:File /Users/woody/Music/audio-out/01 Fool For Your Loving/vocals.wav written succesfully
INFO:spleeter:File /Users/woody/Music/audio-out/01 Fool For Your Loving/drums.wav written succesfully
INFO:spleeter:File /Users/woody/Music/audio-out/01 Fool For Your Loving/other.wav written succesfully

And the output, well, obviously I can't put it all up here, it is copyrighted stuff after all, but here are bits of the sections in one audio file as examples..

mix.mp3

Obviously it is not the sort of quality you want to listen to isolated, but it is convincingly pulled apart and lets you hear stuff that is hard to hear otherwise. 

- edit -

Should have added the links. Heres it release page

https://deezer.io/releasing-spleeter-deezer-r-d-source-separation-engine-2b88985e797e

and here is the source page

https://github.com/deezer/spleeter

although if you are setting it up rather than editing it, python (max 3.7 as it uses tensorflow) will download the files

 

Edited March 4, 2021 by Woodinblack

[Jus Lukin]

-

Edited March 11, 2022 by Jus Lukin

[pineweasel]

I'd not heard of this. Very impressive. Looking forward to having a go later.

[MartinB]

It'd be interesting to compare the results with this one:

 

[Woodinblack]

1 hour ago, MartinB said:

It'd be interesting to compare the results with this one:

 

Will give it a try tonight. Obviously that one is easier as you only have to upload, but its also less free and flexable.

[Saul Panzer]

I've used it on and off for a year or so now, apart from issues with some older muddy mp3s it's pretty great.

Does a decent job ripping voices from background noise which is what I used it for originally, clearing up audio for transcription work, docs and lawyers love recording notes while driving the loudest diesel sounding jeeps with the windows open I guess.

Edited March 5, 2021 by Saul Panzer

[Woodinblack]

Turns out it is the same tool, the website is just a portal to it (as is the app).

Pretty powerful thing really, I was looking on line for bass tabs for Skunk Anansies Twisted, and every one of them is wrong, as is the first few bass lessons I saw on it, and it is very clear by listening to the isolated bass.

They all do the standard triplets on the vocal lines, when its just basically voice, bass and drums. What noone has right is that it isn't a triplet, and after hearing it isolated, it is quite clear it is. Its a pair of notes played after the beat, so it is like a triplet but with no note on the beat, so the last 2 of the triplet.

Going to be useful!