DBS: Neural Network Decides Where to Cut the Spectrum

Researchers at Tsinghua University introduced Dynamic Band Splitting (DBS) — an end-to-end method for audio source separation where frequency band boundaries are predicted based on the content of each mix (hard Top-K) with a fixed budget of 128 band tokens, rather than being predefined.

Compiled by Sergey KostenchukPublished 2026-10-18Updated 2026-10-19

2026-10-18 Research HuggingFace

🤖 DBS: Neural Network Decides Where to Cut the Spectrum

A team from Tsinghua University described Dynamic Band Splitting (DBS): frequency band boundaries are predicted for each mix (hard Top-K, budget of 128 band tokens) rather than being predefined. On DnR v2, DBS-BSRNN achieves 9.07 dB SI-SDR (+29% compared to static BSRNN), while DBS-TIGER reaches 10.13 dB.

🌍 Static spectrum partitioning is a systemic bottleneck for band-split separators: adaptive boundaries with the same token budget provide +1.4–1.9 dB SI-SDR/SDR on BSRNN, TIGER, BS-RoFormer, and MelRoFormer backbones in film, music, and speech. The module integrates into separators without increasing computational cost.

👤 The demo is now open: Hugging Face Space JusperLee/DBS-Demo accepts audio or video and displays speech, music, and SFX separately along with the predicted band boundaries. DBS-TIGER weights and code (Apache 2.0) are open source.

Source 1: https://arxiv.org/abs/2610.14993 Source 2: https://github.com/JusperLee/Dynamic-Band-Splitting

Sources