kobimusic releases Dettato, a 20M-parameter open audio-to-MIDI model
The author behind kobimusic has introduced Dettato, an audio-to-MIDI transcription model that is open source and has just 20M parameters. The post calls it state of the art. According to the author, Dettato significantly beats MuScriptor 1.3B, which the post names as the next best open model, and scores a bit higher than Mirelo's closed-source model, 'Mirelo A2M v1.1 Pro', on Mirelo's published benchmarks. The author puts the size gap at roughly 65x: Dettato is roughly 65x smaller than MuScriptor. The post gives no benchmark names or scores; "significantly beats" and "a bit higher" are the author's wording.
The author also stresses the training data. Dettato was trained on "absolutely no copyrighted recordings". The data is a mix of synthetic instrument recordings and Creative Commons solo instruments mixed in using the URMP dataset. For the rest of the instruments, the author morphs together about 20 soundfonts with a technique called 'soundfont babies', which the author says they published earlier. The model is open source and its outputs can be used for commercial purposes. The stated driving goal of kobimusic is to help musicians, not replace them, and following copyright law is part of that.
The author adds a disclaimer about the benchmarks: it is hard to judge the model's accuracy for your own use case without trying it. One weak spot is named: the model performs poorly at splitting instruments correctly. The output is also only performance MIDI with a beat aligned. For something you can actually play, such as sheet music, the author recommends songscription or klangio, calling both excellent tools by great companies.
The post also explains how the model came about. The author began working on Dettato about 10 months ago but scrapped the work in about April, having treated automatic transcription (AMT) as a data problem when the bottleneck is architecture. The author draws a parallel with OMR, the first model they made. Both fields were first tackled with algorithms, and later both plugged in giant LLM components. Such models, the author argues, become overfit to synthetic data, because the global problem differs from the local one. Both can be reframed as local problems: a detector model for OMR and a pitch-lane model for AMT. The author closes by hoping that open and closed weight music retrieval companies learn from this, and says that avoiding copyrighted recordings does not hold you back and likely leaves you with a better model.
Key facts
- Dettato is an open-source audio-to-MIDI transcription model with 20M parameters, which the author says is roughly 65x smaller than MuScriptor.
- The author says it significantly beats MuScriptor 1.3B, the next best open model, and scores a bit higher than Mirelo's closed-source 'Mirelo A2M v1.1 Pro' on Mirelo's published benchmarks; no scores are given.
- Training used no copyrighted recordings: synthetic instrument recordings, CC solo instruments via the URMP dataset, and about 20 soundfonts morphed with the author's 'soundfont babies' technique.
- Outputs can be used commercially. The output is performance MIDI with a beat aligned, not playable material such as sheet music.
- The author admits the model performs poorly at splitting instruments correctly, and says its design came from treating transcription as an architecture problem rather than a data problem.
Why it matters
Dettato claims strong transcription quality from a very small model: 20M parameters against MuScriptor's 1.3B, a gap the author puts at roughly 65x. It also claims to have been built without any copyrighted recordings, which speaks to a live concern in music AI. The author's argument is that the copyright constraint pushes toward better design, and that the real bottleneck in transcription is architecture, not data. A pitch-lane model, treating the problem locally, is the proposed alternative to plugging in giant LLM components.
Who it affects
Musicians and developers who need to turn audio into MIDI are the direct audience. The author also addresses open and closed weight music retrieval companies, hoping they learn from the approach. Those who care about the provenance of training data may take note of the claim that no copyrighted recordings were used.
How to use it
The model is open source, and the author says its outputs can be used for commercial purposes. What it produces is performance MIDI with a beat aligned. If you need something you can play, such as sheet music, the author points to songscription or klangio instead. Since accuracy depends on the use case, the author advises trying the model yourself. No licence name is given in the post, only 'open source' and the commercial-use statement.
How solid is it
Everything here is the author's own account. The comparison with Mirelo's closed-source model uses Mirelo's published benchmarks, and the post names no benchmarks, metrics or numeric scores; "significantly beats" and "a bit higher" are the author's wording. No independent verification or third-party reaction is reported. The post says roughly 65x smaller, which is a parameter-count comparison; it gives no speed, compute or energy measurement.
Risks and caveats
The author warns that true accuracy is hard to gauge without trying the model, and says it performs poorly at splitting instruments correctly. The output is performance MIDI with a beat aligned, so it is not ready to play or read as notation. The claim that avoiding copyrighted recordings likely yields a better model is stated as a likelihood, not a certainty. The training-data claim rests on the author's description of the data mix.
“The constraint of not using copyrighted recordings doesn't hold you back, but actually encourages innovation, and it's likely you'll end up with a better model in the end.”
— kobimusic, introducing Dettato