USS-Bytedance (any; esp. SFX)
https://github.com/bytedance/uss
(COMMAND: "conda install -c intel icc_rt" SOLVES the LLVM ERROR)
You provide e.g. a sample of any instrument or SFX, and the AI separates it solo from a song or movie fragment you choose to separate.
It works in mono. You need to process right and left channel separately.
Update 29.04.25 (Python No such file or directory fix; thx epiphery)
https://colab.research.google.com/drive/1rfl0YJt7cwxdT_pQlgobJNuX3fANyYmx?usp=sharing
(old) ByteDance USS with Colab by jazzpear94
https://colab.research.google.com/drive/1lRjlsqeBhO9B3dvW4jSWanjFLd6tuEO9?usp=share_link
(old) Probably mirror (fixed March 2024):
https://colab.research.google.com/drive/1f2qUITs5RR6Fr3MKfQeYaaj9ciTz93B2
errors out with:
“sed: can't read /usr/local/lib/python3.10/dist-packages/uss/config.py: No such file or directory”) (2025)
It works (much) better than zero-shot (not only “user-friendly wise”).
Better results, and It divides them into many categories.
Great for isolating SFX', worse for vocals than current vocal models. Even providing acapella didn't give better results than current instrumental models. It just serves well for other purposes.
"Queries [so exemplary samples] for ByteDance USS taken from the DNR dataset. Just download and put these on your drive to use them in the Colab as queries [as similarly sounding sounds from your songs to separate]."
https://www.dropbox.com/sh/fel3hunq4eb83rs/AAA1WoK3d85W4S4N5HObxhQGa?dl=0
Also, grab some crowd samples from here:
Q&A by Bas Curtis and jazzpear
Q: What is the difference between running with and without the usage of reference query audio?
A: Query audio lets you input audio for it to reference and extract similar songs based upon (like zeroshot but way better) whereas without a query auto splits many stems of all kinds without needing to feed it a query.
Q: Let's say there is this annoying flute you wanna get rid off...
and keep the vocals only....
You feed a snippet of the flute as reference, so it tries to ditch it from the input?
A: Quite the reverse. It extracts the flute only which ig you could use to invert and get rid of it