search

Audio-AGI / AudioSep

Official implementation of "Separate Anything You Describe"
https://audio-agi.github.io/Separate-Anything-You-Describe/
MIT License
1.64k stars 118 forks source link

Is something broken? This is really bad. #36

Open nathanodle opened 1 year ago

nathanodle commented 1 year ago

Tried locally and didn't get much, then tried HF space and the model didn't really seperate anything. Was it overfit on the demo data? Barely any difference between input and output on a random song from my library with organ, drums, vocal.

mus-com-psy commented 10 months ago

Sorry I have to agree with @nathanodle. When I tried this on a few basic examples, it did not work very well at all. I will try it a couple more times. Please @liuxubo717 provide us suggestions if you think we are doing something wrong.

coldinjection commented 8 months ago

"separate anything" is indeed an overstatement according to my test results. The only acceptable use case I found for this model is to separate speech (and not singing or vocals) that is already distinct from noise or background music

Beliathal commented 7 months ago

Tested thus far with various audio sources, tried separating noise from speech, music from vocals, etc... Results are mediocre at best, probably could achieve similar things with UVR5.

MonolithFoundation commented 3 days ago

The result is extremly bad, please avoid use this to save time.

Reminder from an disappointed user.