Separate Anything You Describe¶
Summary¶
Text-conditional source separation using CLAP (Contrastive Language-Audio Pretraining) embeddings. A user provides a natural language query (e.g., "banjo" or "separate the violin solo") and AudioSep separates the matching source from the mixture. Uses a ResUNet decoder conditioned on CLAP text embeddings. This is the first system to enable open-vocabulary source separation without predefined stem categories.
Key Claims¶
- CLAP embeddings provide effective conditioning for open-vocabulary source separation
- A single model can separate arbitrary sources described in natural language
- Outperforms fixed-stem models on described sources, even for instrument categories not in training
- Enables separation queries that fixed-stem models cannot express (e.g., "the second violin")
Related¶
- ../entities/audiosef — the tool
- ../entities/clap — CLAP: Contrastive Language-Audio Pretraining
- ../concepts/query-based-source-separation — the broader concept
- ../entities/demucs — fixed-stem alternative