DTT-BSR+ restores music sources with two-stage generative-regression cascade

Music source restoration (MSR) is a challenging problem that requires both separating mixed audio tracks (unmixing) and reversing non-linear production effects like compression or reverb. Current methods struggle to balance precise signal reconstruction with preserving the natural sound of the source. To address this, a team of researchers from Youran Ni, Shihong Tan, Yuzhu Wang, and Gongping Huang developed DTT-BSR+, a two-stage cascade system. The first stage uses a generative DTT-BSR separator to produce stems that match the prior distribution of clean sources. The second stage refines these outputs using a modified Demucs network trained with time-domain and multi-resolution spectral losses.

The system outperforms both its single-stage predecessor and the state-of-the-art X-LANCE MSR system, achieving higher multi-mel signal-to-noise ratio (MMSNR) across all evaluated stems. The researchers also applied Fréchet Audio Distance (FAD) decomposition to analyze performance, revealing an inherent trade-off between signal reconstruction accuracy and semantic distribution fitting across different stems. This insight could guide future improvements in MSR. DTT-BSR+ has been accepted at Interspeech 2026, signaling broad interest in audio AI.