DTT-BSR+: A Generative-Regression Cascade for Music Source Restoration

2026-06-23

AI Summary

Music source restoration entails simultaneously addressing source separation and the inverse problem of nonlinear production effects, a challenge for which existing methods struggle to balance reconstruction fidelity and semantic consistency. This work proposes the first two-stage decoupled framework: in the first stage, a generative DTT-BSR separator models the semantic distribution of clean sources, while the second stage employs an enhanced Demucs network optimized with time-domain and multi-resolution spectral losses to refine signal reconstruction. The approach reveals an implicit trade-off between reconstruction accuracy and distributional alignment across different tracks. Experimental results demonstrate consistent superiority over the single-stage DTT-BSR baseline in terms of MMSNR across all tracks, and outperform the current state-of-the-art X-LANCE system on five individual tracks.

Abstract

Music source restoration (MSR) requires jointly addressing source unmixing and the inversion of non-linear production effects. Current methods struggle to achieve accurate target signal reconstruction while maintaining semantic consistency. To address this limitation, we propose DTT-BSR+, a two-stage cascade MSR system that decouples distribution fitting from signal reconstruction into separate stages. A generative DTT-BSR separator in the first stage produces stems matching the prior of clean sources, and a modified Demucs network in the second stage enhances the first stage output using time-domain and multi-resolution spectral losses. DTT-BSR+ improves multi-mel signal-to-noise ratio (MMSNR) over the single-stage DTT-BSR across all stems, and surpasses the state-of-the-art X-LANCE MSR system on five stems. We also reveal through Fréchet Audio Distance (FAD) decomposition an implicit trade-off between signal reconstruction accuracy and semantic distribution fitting across stems.

Problem

• Music Source Restoration
• Source Unmixing
• Non-linear Production Effects
• Signal Reconstruction
• Semantic Consistency

Innovation

• music source restoration
• generative-regression cascade
• distribution fitting
• signal reconstruction
• multi-resolution spectral loss