MuScriptor: An Open Model for Multi-Instrument Music Transcription

MuScriptor addresses automatic transcription for complex, real-world music mixes containing multiple instruments. The authors describe a training strategy that combines synthetic-data pretraining, fine-tuning on real music audio, and reinforcement-learning post-training. The model also uses instrument-presence conditioning to customize which instruments are transcribed. The work releases open weights and positions MuScriptor for recordings spanning diverse musical genres. The supplied abstract does not provide benchmark scores, dataset sizes, compute details, or licensing terms.