SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision
SheetSage2 uses synthetic MIDI-to-audio supervision to turn music recordings into more coherent lead sheets.
The paper frames the problem as both a data shortage and a sequence-consistency problem. Its system combines synthetic training data, structured decoders for musical dependencies, and autoregressive distillation for inference. In the authors’ evaluation, one SheetSage2-AR model beats listed prior systems on 12 of 15 benchmark-metric pairs across eight collections. Model weights and inference code are public. HF Daily Papers' note
The paper frames the problem as both a data shortage and a sequence-consistency problem. Its system combines synthetic training data, structured decoders for musical dependencies, and autoregressive distillation for inference. In the authors’ evaluation, one SheetSage2-AR model beats listed prior systems on 12 of 15 benchmark-metric pairs across eight collections. Model weights and inference code are public. HF Daily Papers' note
score 4