The SJTU X-LANCE Lab System for MSR Challenge 2025
This AI can perfectly separate and clean up 8 instruments from any messy audio track.
Deep Dive
Researchers from SJTU's X-LANCE Lab have open-sourced an AI system that won first place in the 2025 Music Source Restoration Challenge. Their sequential BS-RoFormer model separates, denoises, and removes reverb from audio, handling up to 8 instruments simultaneously. It achieved a top MMSNR score of 4.4623 and an FAD score of 0.1988, beating all other entries in three subjective and three objective evaluation categories. The code and checkpoints are now publicly available.
Why It Matters
This breakthrough could revolutionize audio production, making professional-grade music separation and cleanup accessible to everyone.