Project ◆ Hidden gem
AudioSep
Audio-AGI/AudioSep · Python · MIT
Text-queried universal audio separation
Unlike fixed four-stem Demucs, AudioSep separates whatever you describe in language — instruments, effects, niches.
7.3 / 100 · Score
★ 1.9k Stars
+-1 30-day growth
25 min Time to first run
Not required API key
Score
Trend 13
Activity 8
Quality 65
Community 29
Relevance 100
Setup at a glance
DeployCli
GPU8 GB
API keyNot required
UI languages
Doc languagesen
Install
git clone https://github.com/Audio-AGI/AudioSep.git && pip install -r requirements.txt Deep dive
AudioSep is "Separate Anything You Describe": instead of fixed stems (vocals/drums/bass), you describe the sound you want isolated in natural language — a dog barking, a specific instrument, a speaker — and the model separates that.
It is universal sound separation trained on large-scale audio-language data, which moves separation from a closed stem set to an open-vocabulary task.
The project publishes a demo page, a Hugging Face Space for browser trials, and checkpoints on Google Drive for local runs.
Use cases
Pull a described instrument out of a mix
Clean sound design beds
Custom stem workflows
Good for / Not ideal for
Good for
Sound designers and remixers
Not ideal for
Users who only need vocals/drums/bass/other