知Yizhi Daily
← Back to this edition

Conversations

Dettato audio-to-MIDI model achieves greater efficiency with 20000000 parameters

Edition 2026-11-21 · 08:00 Beijing time

A developer has released the open-source audio-to-MIDI model Dettato, saying it has about 20000000 parameters, roughly 65 times fewer than MuScriptor 1.3B, outperforms that open-source model on published benchmarks, and scores slightly above the closed-source Mirelo A2M v1.1 Pro. Training did not use copyright-protected recordings, and outputs can be used commercially.

According to the publisher, Dettato is an audio-to-MIDI transcription model with about 20000000 parameters, roughly 65 times smaller than MuScriptor 1.3B. The publisher says that on published benchmarks, Dettato is clearly better than MuScriptor 1.3B and also scores slightly higher than Mirelo's closed-source model, Mirelo A2M v1.1 Pro.

The publisher also cautions that benchmarks cannot replace testing in users' own scenarios. Dettato performs relatively poorly at correctly separating instruments and outputs beat-aligned performance MIDI; if users need sheet music that can be played directly, the publisher recommends tools such as Songscription or Klangio.

For training data, Dettato used synthesized instrument recordings, Creative Commons solo instruments mixed through the URMP dataset, and other instrument timbres fused from roughly 20 soundfonts using what the publisher calls the “soundfont babies” technique. The publisher emphasizes that no copyright-protected recordings were used in training, that the model is open source, and that its outputs may be used for commercial purposes.

The author says they started the project about 10 months ago and discarded the original approach in April. They initially treated automatic music transcription as a data problem, but later concluded that the bottleneck was architecture; their approach was to reframe audio music transcription as a local problem and use a pitch-lane model rather than relying on a large LLM component that can easily overfit to synthetic data.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader