PPT mouthpieces

Software AI to Remove Instruments from Recordings

rhysonsax

Well-Known Member
Café Supporter
Messages
5,727
Location
Surrey, UK
I stumbled on this video reviewing an AI-driven software tool that seems to be able to isolate individual instruments quite well in recordings. All the examples are older small band jazz recordings, including vocal + rhythm section (1' 10"), tenor saxophone + rhythm section (5' 22") and piano trio (7' 27").

View: https://youtu.be/zzPqr1HcMow?si=12MuWbweNQjg1Gm-


I am impressed and think it would already be useful to me, both for transcribing and for making backing tracks.

And the technology is only going to get better.

Rhys
 
Last edited:
I can see the use.

And once they can isolate all the instruments they can autotune everything, and fix timing problems. Most recorded, commercial singers are already autotuned, apparently.
 
I just tried the preview demo, choosing 'piano' on a 30 second excerpt of piano, drums, bass, trumpet and tenor sax. It seems to do pretty well on isolating the piano (removing the rhythm and horns) , and not as well on removing the piano and keeping the other instruments. So it does seem like a handy tool but you'll have to put some arranging work into getting a useful backing track.

That said there are last I checked at least a couple of open source stem separation tools out there, I just don't know if this tool is that much better it's worth the $$.
 
I've been using a program called ripxdaw PRO for a longer time to separate solo parts from recordings.
It does a decent job for my purposes. I've tried lalal.ai and it doesn't seem better.
What I cannot understand are the positive comments on youtube about singers singing somebody elses songs - Frank Sinatra singing Bohemian Rhapsody for example. To my ears, it sounds horrible, AI or not.
 
Most recorded, commercial singers are already autotuned, apparently.
This is pretty much 100% and has been for decades now. The term autotune is misleading as whilst you can “select all” and double click for 100% tuning to A440, it unsurprisingly sounds rather obvious and can also corrupt certain bits of audio.

Manual tuning can still be pretty quick (depending upon how wayward your singer is) and tuning choices are made by the engineer.

Autotune can be used on live performances and has been for years.
 
I've been using a program called ripxdaw PRO for a longer time to separate solo parts from recordings.
It does a decent job for my purposes. I've tried lalal.ai and it doesn't seem better.

Could you tell us what your purposes are ?

Rhys
 
I've used lalai for a while now and find it very useful to isolate the sax. My primary purpose is to hear the phrasing for more accurate transcription.

I also use the free version of moises which is very good and you can split unto 5 stems. I occasionally use the stems without the vocals/instrument for backing tracks. The paid version can do more but I've not explored it or needed it as yet.
 
Ripx got a good score in this review, We tested 5 of the best stem separation software tools (and the best one was free)

I gave that article's highest rated tool Gaudio a go, on the same track I uploaded to lalai. The downside is it takes a while for your upload to process in their queue. The upside is...well there wasn't really an upside as far as I could tell, the results were similar to lalai. Individual separation is OK (though with that same weird tone quality at times where it seem like someone is sucking the sound through a straw), but there are a lot of artifacts when you listen to a track with instruments removed.

So it does seem good to have the stem separation in a tool like a DAW. I think if someone released a "musician's practice app" that did this (kind of like Transcribe! in ease of use -- I guess Moises is basically that?), less complex than a DAW, that would be very useful.
 
Last edited:
So I have a recital next month and realized I could try to make a backing track from the original tune with these tools, instead of using iReal or another play-along. Here's a sample of my first draft:

Sample of backing track for Song for my Father

After playing around and not being very happy with any individual result I ended up blending the results of three different tools: Gaudio, Lalalai, and demucs (which I downloaded). All of them seemed to do best isolating the drums, and worse at piano and bass.

Is this any better than iReal or a play-along (of which there are many for a popular tune like this)? Hard to say IMO. The iReal track is fine and the play-alongs are too, and the sound quality is better, but there is a feel in the original recording which I think the play-alongs I've heard struggle to duplicate (and iReal definitely doesn't).

Also it doesn't help that I'm trash at audio editing ;). I had never thought of trying to improve in that area but after this experience I'm thinking about it more!

I would be interested in hearing some samples from anyone else here who cares to give this a try, it's quite fun if you're into this sort of thing.
 
Back
Top Bottom