Experimental Eight-Stem Restoration

Models, parameters, and sources

XLANCE-MSR is exposed as an experimental eight-class source-restoration tool. MSR source restoration is not just another name for splitting stems: it means separation plus trying to recover cleaner dry sources. Normal six-stem separation mainly splits a finished song into vocals, drums, bass, guitar, piano, and other, and the stems often still sound like the record with reverb, compression, EQ, mastering, and codec artifacts. XLANCE-MSR targets finished or degraded audio, runs denoise first, then restores vocals, guitar, keys, synth, bass, drums, percussion, and orchestral while trying to remove degradation and some post-processing. It starts from a strong six-stem BS-RoFormer system and further splits drums and other; its results are read on the MSR source-restoration leaderboard, not merged into MVSep six-stem SDR rankings.

Paper and benchmark notes

Parameter guide

• Fixed model: XLANCE-MSR with public MIT checkpoints from chenxie95/xlance-msr-ckpt; the system starts from strong six-stem BS-RoFormer separation before eight-class restoration.
• Outputs: vocals, guitar, keys, synth, bass, drums, percussion, and orchestral WAV stems; effectively it further splits the ordinary six-stem other/drums areas.
• Score source: MSR Challenge 2025 ranked XLANCE-MSR first on objective and subjective source-restoration results; read MMSNR, Zimt, FAD, and MOS separation/restoration, not six-stem SDR.
• Evaluation scope: this is eight-class source restoration / eight-stem restoration, scored on the MSR restoration task rather than merged into MVSep six-stem SDR rankings.