Kyutai's MuScriptor Transcribes Full Band Mixes to MIDI

Why this matters

MuScriptor's own benchmark numbers make the case better than any marketing copy could. A model trained purely on synthetic MIDI hits 34.5 onset accuracy with a 0.85 false-positive rate; train it on real multi-instrument recordings instead, and onset accuracy jumps to 54.4 with false positives cut to 0.35. That gap tells us the synthetic-to-real problem in AMT was never about architecture, it was about training data nobody had bothered to collect at scale.

Kyutai and Mirelo releasing this open-weight suggests the field is finally treating multi-instrument transcription as a data problem worth solving properly, rather than squeezing more out of single-instrument pipelines. For developers building music tools, this is a usable baseline to fine-tune against rather than a research curiosity. For researchers, the synth-versus-real comparison is a clean ablation worth studying on its own.

Recall still sits at 0.78 on real audio, so full mixes with dense instrumentation will still trip it up. Watch whether Kyutai extends training data to genres beyond whatever they used here, and whether third parties fine-tune it toward specific instrument families.