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

DTT-BSR+ is a newly proposed two-stage cascade system for music source restoration (MSR) that separates the processes of distribution fitting and signal reconstruction. The first stage utilizes a generative DTT-BSR separator to produce clean source stems, while the second stage employs a modified Demucs network to refine these outputs using time-domain and multi-resolution spectral losses. This approach achieves superior multi-mel signal-to-noise ratio (MMSNR) compared to the previous DTT-BSR model and outperforms the state-of-the-art X-LANCE system, highlighting a significant advancement in balancing signal reconstruction accuracy with semantic consistency in MSR tasks.