I hope we keep making progress in isolating tracks in music. I love listening to stems of my favorite songs, I find all sorts of neat parts I missed out on. Listening to isolated harmonies is cool too.
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.
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).
As someone recording myself playing music, I've been meaning to see if any of these tools are good enough yet to not only separate vocals from another instrument (acoustic guitar for example), but do so without any loss of fidelity (or least not a perceivable one).
The reason I'm interested in this is because recording with multiple microphones (one on guitar, one on the vocal), has it's own set of problems with phase relationship and bleed between the microphones, which causes issues when mixing.
Being able to capture a singing guitarist with a single microphone placed in just the right spot, but still being able to process the tracks individually (with EQ, compression, reverb, etc), could be really helpful.
This is incredible! I wouldn't have thought it was possible to cleanly separate tracks like that. I wonder to what extent the model is filling in gaps, akin to Samsung's "ultra zoom" moon.
I use moises frequently for track separation for learning songs. It does pretty dang well. I was shocked that the score of moises is ranked way worse than just about everything else, including lalal.ai, which I also used before buying moises. Perhaps lalal.ai has gotten better since I last tried it.
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?
Now that I go back and read it again I agree with you. Presumably "win rate" means what percent of the time did the SAM model (Meta's new one) beat the other tool over some set of examples.
super amazing demo performance being able separate out music voice and background noises. do you have to explicitly specify what type of noise to separate?
Playing with the background I tried to Isolate just the espresso machine and the train sounds in one of their demos and it seemed to fail. Maybe not the desired use case, but I thought it was odd that I could break it so easily on the sample material.
Footsteps worked pretty well when I tried that on the other hand. I wonder if lot of it has to do with how well the model understands what the english description of the sound should sound like...
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.
Would be interesting to leverage the non spoken/environment noises to guide what level of detail and style of speech a chatbot replied with, for instance being more casual, gentle, with a touch more detail if in a quiet home/office environment, but more curt and concise with emphasized diction if the person is traveling, such as in a noisy train concourse.
People tend to do that subconsciously but bots ignorantly wittering on can be annoying and hard to use because they miss the cues.
100%. The music world has gone through the "but what will we do now?" at least 6-7 times. Music videos ("video killed the radio star"), sampling, the DAW (and time aligning), home studios, auto tune, plugins and amp simulators, napster/piracy, etc, etc.
This one rankles me because of a) the benefits piracy has (third world consumers can now discover you, for starters) and b) the absolute bad faith way in which the industry acts, screwing over artists, unethically going after Pirate Bay by making it into a trade war with Sweden (I think)
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.
It's not clear from the blog post, the git page, and most other places if this will run on, even in big-O:
* CPU
* 16GB GPU
* 240GB server (of the type most business can afford)
* Meta/Google/Open AI/Anthropic-style data center
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
The reason I'm interested in this is because recording with multiple microphones (one on guitar, one on the vocal), has it's own set of problems with phase relationship and bleed between the microphones, which causes issues when mixing.
Being able to capture a singing guitarist with a single microphone placed in just the right spot, but still being able to process the tracks individually (with EQ, compression, reverb, etc), could be really helpful.
- This feature is awesome for sample-based music
- Sample music is not what it was due to difficulties related to legal rights
- This model was probably created by not giving a damn about said rights
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.
"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
This one rankles me because of a) the benefits piracy has (third world consumers can now discover you, for starters) and b) the absolute bad faith way in which the industry acts, screwing over artists, unethically going after Pirate Bay by making it into a trade war with Sweden (I think)