Skip to content

AudioSep

Separate audio sources with natural language queries

Visit official site

What audiosep is

The model is an open-domain audio source separation system that takes natural language queries as input.

What you can do with it

  • Pull a guitar sound out of a mix
  • Separate a singing voice from a track
  • Enhance speech inside a noisy recording

Key features

  • A text encoder paired with a separation model
  • Audio event and instrument separation tasks
  • Strong zero-shot generalization described

Additional details

  • Training uses a large multimodal dataset
  • Evaluation covers separation and enhancement
  • Source code and benchmarks are planned for release

Pros & cons

✓ What we like

  • Natural language queries drive the separation
  • Source code, benchmarks and pretrained models are planned for release

! What to watch out for

  • The architecture relies on two separate components
  • Queries are given as captions or text labels

FAQ

What are the two components?

A text encoder and a separation model are named in the material.

What queries are supported?

Audio captions or text labels are listed as queries.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review