A.I. To Make Backing Tracks?

tbleen

With all of this new AI stuff out there, I'm wondering if anybody has come across a tool that cleanly removes the lead instruments of a track. I was learning Fanou's "How About You" and was thinking how cool it would be to use Benji's rhythm as a backing track. While I'm being slightly lazy in creating my own backing track, it just wouldn't have the same power and feel as one of the Wintersteins in the studio.

Buco

Well that'll be the day. I don't think it's here yet, at least as easily accessible consumer product, but I'm sure it's coming.

In a way that's exactly what was done to create the latest Beatles tune, Now and Then. All that was available as far as John's voice was a cassette tape with him singing and playing the piano plus with the sound of the TV in the background. AI tech was used to recognize the voice and isolate just that as a separate track in the studio.

lucky

I’ve not tried any of them but a quick google throws up a number of demixing apps including free ones and professional ones with free trials.

adrian

The baseline technology ("audio source separation") already exists, but the logistical issue is that most (all?) of today's products prioritize removing vocals.

The more sophisticated ones can separate into "vocals" + "drums" + "bass" + "everything else". So if you put a gypsy jazz recording through it, at best you'd get "bass" + "everything else" — meaning it's not helpful if you want to isolate the rhythm guitar from the lead guitar.

To my knowledge there's no product yet that gets more granular with separation. But, yeah, I'd expect we'll see this in a few years, maybe less!

Adrian

flaco

Moises.ai is the one I’ve played around with. Adrian is spot on that these AI models are trained on pop and rock music. They don’t recognize the instrumentation of gypsy jazz since they are looking for drums, electric bass, electric guitar, etc. Even just trying to isolate the bass track doesn’t really work for the acoustic bass, and it seems like it only picks up about half the notes. The concept is there and technical ability is there, they just need to spend the time training the model on gypsy mass instrumentation.

Tydides

Anything but the latest gen isn’t gonna be usable but I got pretty good results from one — so long as you use the djanog electric recordings with drums. And even then it’s bass and drums but still very cool.

Use VB-Splitz

https://audiostrip.co.uk/isolate

tbleen

Thanks everyone. With so much AI progressing so quickly its hard to keep track of all the tools. Sounds like I'm overestimating its capability for this niche use, for now. Well...at least I can remove the silly vocals to Nagasaki and Shine now?

JasonS

Moises is the best one I've used. It can definitely work for instrumental music when the frequency/attack separation between the instruments is great enough. I used a track that had drums, upright, & sax and it had no problem separating them. The difficult with gypsy jazz would mostly be with the rhythm guitar since it often overlaps with the other instruments in timbre and frequency space.

billyshakes

Well...at least I can remove the silly vocals to Nagasaki and Shine now?

I'm usually all for lyrics in the songs but, yes, the lyricist for Nagasaki seemed to run out of nonsense rhymes so just started throwing out random syllables. Then again, given that it was Harry Warren, who also wrote Jeepers Creepers and Chattanooga Choo Choo, and Cheerful Little Earful, the rhyming onomatopoeia should come as no surprise. That said, I do like Freddy Taylor's performance and voice.

I wonder if someone tried to make a dance called the "wicky wacky woo" after hearing this? Or could it be alluding to something more cringey?

wim

Think it ended up as the name for a popular cocktail (a sort of fruit punch tiki-bar drink)?

I dislike Freddy Taylor's performance on Nagasaki. The actual melody of the song is interesting, but he just raps on a single note over the most interesting part instead of actually hitting the melody notes. That cool turnaround was the one thing this tune had going for it, and lacking that, this Django version is pretty dull and forgettable

Myckalizter

@wim compared to Django we all sound pretty mortal. As for Freddy Taylor's performance it's honestly pretty good and if it was good enough for Django and Stephane I'd say it's good enough for me. I enjoy it just as much as the rest of the recording. Most importantly Freddy Taylor doesn't lose any of the swing and that's one of the reasons why he ended up in a studio to record multiple times with Django, which speaks for itself. Maybe he would have interpreted the melody differently at a lower tempo, anyway it is pretty good as-is if you ask me.