跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-10-10AI 評分39

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

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

AI 導讀

動態頻帶切分(DBS)是一種端到端音源分離框架,能在固定頻帶 token 預算下,為每個混合音訊學習頻率軸的切分位置。使用 128 個頻帶時,DnR v2 上 BSRNN 的平均 SI-SDR 由 7.69 提升至 9.07 dB,TIGER 則由 8.74 提升至 10.13 dB。程式碼、預訓練模型及音訊範例已在項目頁提供。

正文

We introduce Dynamic Band Splitting (DBS), an end-to-end framework that learns where to split the frequency axis for audio source separation, while keeping a fixed band-token budget.

DBS adapts band boundaries to each mixture and works with BSRNN, TIGER, BS-RoFormer, and MelRoFormer. With 128 bands in both models, average SI-SDR on DnR v2 improves from 7.69 to 9.07 dB for BSRNN and from 8.74 to 10.13 dB for TIGER. We also evaluate music separation on MUSDB18-HQ and speech separation on EchoSet.

🎧 Try the interactive demo to separate speech, music, and sound effects and watch the predicted frequency boundaries:
https://huggingface.co/spaces/JusperLee/DBS-Demo

Code, pretrained models, and audio examples are available on the project page.

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co