Tsinghua's DBS learns content-adaptive frequency band splits for audio source separation
thu-ml · hf · 2026-10-16
A Tsinghua University team introduces DBS (Dynamic Band Splitting), an end-to-end framework for audio source separation.
- Existing band-split architectures use identical fixed frequency boundaries for every mixture, merging discriminative spectral structure or wasting tokens
- DBS predicts content-conditioned boundary scores within frequency blocks and selects a prescribed number of boundaries via hard Top-K, with content-dependent pooling
- Improves separation quality while maintaining a fixed band-token budget
More from Research
- Square lattice connective constant remains unsolved — ripe for AI, argues mathematician — tak3sh8 · 2026-10-16
- Quiz: Double the Residual Stream from 4096 to 8192 — How Does Param Count Change? — Sauers_ · 2026-10-16
- Researcher credits GPT for multiple accepted papers, incl. disproving Wegner's 1965 conjecture — ctjlewis · 2026-10-16
- AutoAgent0: VLM-powered agentic runtime lifts long-tail driving success by 28.3pp — yining_hong · 2026-10-16
- Shuffled labels hit AUC 0.59: selection bias, not signal, drives prosody benchmarks — maier_ak · 2026-10-16
- CODESTRUCT reframes codebases as AST action spaces to fix brittle LLM code agents — burkov · 2026-10-16