Learning Where to Split: Dynamic Band Splitting for Audio Source Separation
Bingzhe Zhang, Kai Li, Xiaolin Hu
Tsinghua University · Chinese Institute for Brain Research
Published October 10, 2026 · on the feed October 16, 2026 · ▲ 137 · arXiv 2610.14993
The abstract
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. Because informative spectral structure varies with source content, fixed boundaries can merge discriminative components or spend tokens where little resolution is needed. We introduce Dynamic Band Splitting (DBS), an end-to-end framework that adapts boundary locations and band widths to each mixture under a fixed band-token budget. DBS predicts content-conditioned boundary scores within frequency blocks, selects a prescribed number of boundaries using hard Top-K, and applies content-dependent pooling to obtain compact band tokens. After separation, it smooths and expands these tokens to the original STFT grid, refines them along frequency, and decodes complex masks with a frequency-shared head. DBS improved separation over the corresponding static models on DnR v2, MUSDB18-HQ, and EchoSet, with DBS-BSRNN achieving relative gain of 29.0% in average SI-SDR on DnR v2. These results and boundary analyses on DnR v2 show that content-adaptive frequency partitioning benefits multiple separator backbones under a fixed band-token budget. Code and audio demos are available at https://cslikai.cn/Dynamic-Band-Splitting
The Brief
2:37A short podcast summary of the paper. Written and voiced by AI.
Transcript
Learning Where to Split. Dynamic Band Splitting for Audio Source Separation. By Bingzhe Zhang and two others, from Tsinghua University and the Chinese Institute for Brain Research. Published on the tenth of October. Its code repository is worked out: about forty source files, with a readme and a licence.
An audio source separation model has two pressures on it. It must preserve fine spectral structure, and it must keep the cost of the separator tractable. Band-split architectures handle the cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a single token. The catch is that the same boundaries are used for every mixture.
Yet the informative spectral structure changes with what the sources contain. With fixed boundaries, one of two things can go wrong. Discriminative components that should stay apart get merged into one band, or tokens are spent on parts of the spectrum where little resolution is needed. A boundary that suits one mixture can be the wrong boundary for the next.
Dynamic Band Splitting, or DBS, lets the model learn where to split. It is an end-to-end framework that adapts both the boundary locations and the band widths to each mixture. What stays fixed is the band-token budget. The number of tokens the separator has to process does not grow; only where those tokens are placed on the frequency axis changes from one mixture to the next.
On the way in, DBS predicts content-conditioned boundary scores within frequency blocks. It selects a prescribed number of boundaries using hard Top-K, and it applies content-dependent pooling to turn each resulting band into a compact band token. The separator then works on those tokens.
On the way out, the tokens have to return to the spectrum's original resolution. DBS smooths and expands them back to the original STFT grid, refines them along frequency, and decodes complex masks with a head that is shared across frequency.
DBS improved separation over the corresponding static models on three datasets: DnR v2, MUSDB18-HQ and EchoSet. The largest figure reported is for DBS-BSRNN, with a relative gain of twenty-nine percent in average SI-SDR on DnR v2. Boundary analyses on DnR v2 accompany the scores, and together the authors read them as showing that content-adaptive frequency partitioning benefits multiple separator backbones under a fixed band-token budget.
Band splitting already made a trade: coarser tokens in exchange for a cheaper separator. This work keeps that trade but stops fixing its terms in advance. The budget is the same and the separators are the same; what moves is the line between one band and the next, drawn afresh for every mixture according to what is in it.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a fixed frequency map laid over every sound.
You are pulling speech, music and effects apart from one recorded mixture. Working on the full spectrogram keeps fine detail but makes the separator expensive.
The road ahead
- 1 · The problem ~6 min Grouping frequencies into bands makes a separator cheap, but one fixed grouping for every recording can blur what matters and waste effort elsewhere.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Keep the number of band tokens fixed, but let the model move the band boundaries and widths for each mixture, then rebuild full-resolution masks.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Score gaps, keep the top ones, pool bands into tokens, separate, expand back to every bin, and decode complex masks.
- 6 · The experiment ~6 min You’ll guess: On DnR v2, how large was the relative gain in average SI-SDR for DBS-BSRNN over the published BSRNN result?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.