AI research for musicians, from kobi.music.
copisteria reads a scanned page of printed music and writes MusicXML. Instead of hand-written rules, a 2M-parameter transformer reads every symbol in the context of the whole page, on top of a copista symbol detector in two sizes. It is made for real scans: authentic 1800s and 1900s editions, where other systems collapse.
| Model | Parameters | Error rate, string-quartet scans ↓ | Weights |
|---|---|---|---|
| copista-28m | 30.7M | 8.7 % | 🤗 kobimusic/copista-28m |
| copista-2m | 4.0M | 11.2 % | 🤗 kobimusic/copista-2m |
On the OMR-NED benchmark both sizes are ahead of every other system on every set: string quartets, Lieder and Polish piano scores, scanned and rendered. The figures and how they were measured are in copisteria's benchmark.
babymaker takes the same note from a few instruments and makes a new instrument from them. It is not a crossfade: the attack, the tone and the decay of the new note can each be part flute, part violin and part harp. It was made to give transcription models more timbres to learn from than the soundfonts there are. The blog post, with a demo.
copisteria and its weights are released under the Apache License 2.0. Please credit KobiMusic when you use them; each repository has a citation.