AI Trend Research
EN

← Key papers

Paper review · October 17, 2026

Learning Where to Split: Dynamic Band Splitting for Audio Source Separation

Speech & audio Advanced Upvotes 131 cs.SD Tsinghua Machine Learning Group

The paper proposes DBS, which predicts band boundaries for each input instead of using fixed frequency bands while keeping the number of tokens the same, and improves the performance of several band-split separation models on three datasets.

Problem

Band-split source separation models such as BSRNN divide the spectrum into frequency bands predefined by humans and compress each band into a single token. But which frequency regions carry the most information differs depending on the audio content. Fixed boundaries can mix components that need to be told apart into one band, or waste tokens on regions with little information.

Method

The frequency axis is divided into 4 fixed blocks, and the number of bands (token budget) for each block is set. Within each block, temporal convolution, magnitude statistics and a BiLSTM predict boundary scores between adjacent frequency bins. Hard Top-K selects a fixed number of boundaries to form bands. Each band becomes one token through content-dependent weighted pooling and is fed into an existing separation backbone (BSRNN, TIGER, BS-RoFormer, MelRoFormer). After separation, the tokens are smoothly joined using boundary probabilities and then expanded back to the original STFT resolution. The result is refined along the frequency direction, and a head shared across all frequencies predicts a complex mask. The discrete boundary selection is trained with a straight-through estimator, and boundary density is controlled with a ratio regularization loss.

Results

Limitations

Although the number of tokens stays the same, the boundary prediction and reconstruction modules increase computation. On DnR v2, MACs rose from 76.6→124.0 G/s for BSRNN and 20.4→31.6 G/s for TIGER. The MUSDB18-HQ experiments separate 4 stems with one model and no external data, so they are hard to compare directly with the original papers' numbers, which used a dedicated model for each source. The EchoSet comparison relies on numbers reported in other papers. Some design choices, such as the number of blocks, are set by hand.

Why it matters

It is a general-purpose module that can be added to band-split models used in applications such as music remixing, separating dialogue and sound effects in film, and hearing aids. It shows that fixed frequency splitting can be replaced by learning. The added computation, however, needs to be weighed as well.

Key terms

Band-split architecture
A design that splits the frequency axis of a spectrogram into several bands and compresses each band into one fixed-size token, reducing the sequence length the separation model must process.
STFT (short-time Fourier transform)
A transform that cuts audio into short segments and computes the frequency components of each segment, turning it into a time-frequency spectrogram.
Hard Top-K + Straight-Through Estimator
Discretely selecting the K highest-scoring items blocks gradient flow. This technique therefore uses the discrete selection in the forward pass and lets gradients flow through a smooth approximation in the backward pass so the model can be trained.
SI-SDR
A metric (in dB) that measures how close a separated source is to the reference source, regardless of loudness. Higher is better.

Original

Authors: Bingzhe Zhang, Kai Li, Xiaolin Hu · Posted 2026-10-10 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.