This repository contains the official implementation of "Separate Anything You Describe".
We introduce AudioSep, a foundation model for open-domain sound separation with natural language queries. AudioSep demonstrates strong separation performance and impressive zero-shot generalization ability on numerous tasks such as audio event separation, musical instrument separation, and speech enhancement.
Check the demonstration examples in the Demo Page!
We will be releasing the code & pre-training models & evaluation benchmark very soon!
Please stay tuned for further updates and more details!