Meta Segment Anything Model Audio

Basically the same thing musicians said about the synth and music made by computers back in the day

That’s pretty much been the story since the Neolithic revolution though?

"Don't be curmudgeonly. Thoughtful criticism is fine, but please don't be rigidly or generically negative."

"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."

https://news.ycombinator.com/newsguidelines.html

very cool idea

It shall also allow to make re-recordings in higher quality of stuff that are impossible to find in good quality. Like that cover that that band played only once at that obscure concert and that was recorded on an old tape. Or many very old reggae songs: although many from Jamaica/Kingston had great recordings (there was know-how and great recording studios there) there's also a shitload of old reggae songs that are just barely listenable to because the recording is so poor (and, no, it's not an artistic choice by the artist: it's just, you know, a crappy recording).

From the papers I've read, the stem separation models all seem to train off what seems like a fairly small dataset that doesn't have great instrument representation.

I wonder if you could assemble a big corpus of individual solo instruments, then permute a cacophonous mix of them. IIRC the main training dataset is comprised of a limited number of real songs. But I think a model trained on real songs might struggle with more "out there" harmonies and mixes.

The problem of track isolation is sometimes underconstrained, and so any AI system that does this will probably invent "neat parts" for us to hear that weren't necessarily in the original recording. It feels like using super-resolution models to notice details about your great-grandma's wedding dress.

Footsteps worked pretty well when I tried that on the other hand.

Maybe I'm totally misinterpreting, but the chart I'm looking at says "Net Win Rate of SAM Audio vs. SoTA Separation (text prompted)", so perhaps a lower number means that the alternative model is better?

It’s really a shame how popular it was to mar shows with this… I saw a DVD set of a show once with a no-laugh-track version. It sucked because the actors pause for the laughs after each line. This is bad enough with the laugh track in place, but if it’s just dead air it makes every scene feel awkward.

For me it either worked great or not at all. Extracting footsteps, the air conditioner noise, voices, one particular persons voice (identified by gender), all worked great (across multiple clips for most of those).

A few prompts failed almost entirely though, "train noises", "background noise" and "clatter"... so definitely sensitive to either prompting or the kind of noise being extracted.

It realy depends on your runtime environment, but I agree it would be nice to have some references with commonly used setups.

Indeed. I've tried to run it locally this but couldn't get it running on my measly gaming-spec workstation.

It's seems you need lot's of ram and vram. Reading the issues on github[1], it does not seem many others have had success in using this effectively:

- someone with a 96 Gb VRAM RTX 6000 Pro had cuda oom issues

- someone somehow made it work on a RTX 4090 somehow, but RTF processing time was 12...

- someone with a RTX 5090 managed to use it, but with clips no longer than 20s

It seems utility of the model for hobbyist with consumer grade cards will be low.

[1]: https://github.com/facebookresearch/sam-audio/issues/24