Creating multitracks with A.I. (Innuendo, vocals only)

A couple of days ago Deezer released their open source code Spleeter which uses deep learning algorythms and artificial intelligence to seperate Vocals from Backing Tracks of any song. It separates a maximum of 5 stems; vocals, drums, bass, and 2 others (haven't tried yet) Ofcourse, this is only in its early stage but my first result sounds promising.

Imagine being able to remove “property of Queen Productions”

That was interesting but flawed as it never removed enough of the original source.

It DID make me appreciate the soundscapes they'd created even more - simply hearing that isolated guitar trill... that was beyond interesting. Stems would be amazing.

Would you get a better result if you used the 5.1 stems from the German Greatest Flix DVD?

Great, more vocal snippets for Freddie Goes To The Shops memes.

Amazing!!!
Thanks

Nah.. a ghost hovering above your ears... nothing revealed. A bit unsettling actually...

Phenomenal sound!

Can it remove the vocals and leave the rest?

That's amazing! some voice voice Freddie had.. I thought you did a good job cheers.

We will see a CGI Freddie in the future moving/ acting to concerts that never shot on film. Scary and fascinating at the same time.

How do you do it?

You chose the best Queen song too! Lol. Well in my opinion anyway!

brians wig wrote:

Can it remove the vocals and leave the rest?

That would be a lot easier technically.

This process outputs (in my case) two files in a new folder.

1. vocals.wav - not the main vocals but all vocals in a song
2. accompaniment.wav - all the rest

If you extract the maximum of separated files you get the following

1. vocals.wav
2. bass.wav
3. drums.wav
4. piano.wav
5. accompaniment.wav (which is all that's left)

As soon as I have it up and running again (there's a codec issue as of yesterday) I will try extracting all five stems from Innuendo.