Learning Where to Split: Dynamic Band Splitting for Audio Source Separation
Dynamic Band Splitting adapts audio frequency bands, gaining 29% relative SI-SDR on DnR v2.
Dynamic Band Splitting adapts frequency-band boundaries and widths to each audio mixture while keeping a fixed band-token budget. The model scores candidate boundaries inside frequency blocks, selects a set number with hard Top-K, pools content-dependent tokens, and later expands them back to the STFT grid for mask decoding. It improves separation over static counterparts on DnR v2, MUSDB18-HQ, and EchoSet. DBS-BSRNN reports a 29.0% relative gain in average SI-SDR on DnR v2.
- DBS predicts mixture-specific frequency boundaries under a fixed token budget.
- Hard Top-K selects a prescribed number of boundaries.
- It outperforms static band splitting on three datasets.
- DBS-BSRNN gains 29.0% relative average SI-SDR on DnR v2.
- Code and audio demos are publicly available.
From the original article
1 min read at huggingface.co
Audio source separation models must preserve fine spectral structure while keeping the cost of the separator tractable. Band-split architectures reduce this cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a token, but the same boundaries are used for every mixture.
By Bingzhe Zhang, Kai Li, Xiaolin Hu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.