---
title: "Dettato music transcription model beats MuScriptor with 20 million parameters, about 65 times smaller"
url: https://badiyaa.com/dettato-audio-midi-model-muscriptor/
published: 2026-11-21T13:52:00+00:00
modified: 2026-11-21T08:39:54+00:00
section: Tech
publisher: Badiyaa
---

# Dettato music transcription model beats MuScriptor with 20 million parameters, about 65 times smaller

Dettato, an open-source audio-to-MIDI model with 20 million parameters, claims to beat the 1.3-billion-parameter MuScriptor with no copyrighted data.

A solo developer has released Dettato, an open-source model that listens to a recording and writes out the notes as MIDI, and says it beats the best rival open model while being about 65 times smaller. Dettato has just 20 million parameters. MuScriptor, the model it claims to beat, has 1.3 billion.

The bigger headline for musicians may be what Dettato was not trained on. According to [the Dettato launch page from kobimusic](https://kobi.music/research/dettato), published on November 16, 2026, "we trained on absolutely no copyrighted recordings." The model is open source, and its outputs can be used commercially.

## What audio-to-MIDI actually means

Automatic music transcription, or AMT, takes a sound file and works out which notes were played, when they started and when they stopped. The output is MIDI, the instrument data format standardized in 1983 and maintained by the MIDI Manufacturers Association, as [Wikipedia's MIDI article](https://en.wikipedia.org/wiki/MIDI) explains. A MIDI file does not hold sound; it holds instructions such as pitch, timing and velocity, which makes it easy to edit, rearrange or feed into a synthesizer.

Turning a messy mix into clean notes is hard, especially when several instruments overlap. That is why the size of the claimed gap matters.

## The numbers behind the claim

The launch page says Dettato "significantly beats MuScriptor 1.3B - the next best open model" and scores slightly higher than a closed commercial model, Mirelo A2M v1.1 Pro, on that company's published benchmarks. The scoring uses three test sets, PHENICX, RWC classical and RWC royalty-free, with each set's F1 score taken as the average of onset, frame and offset accuracy.

The size ratio is easy to check. Divide 1.3 billion by 20 million and you get 65, so Dettato carries roughly one sixty-fifth of MuScriptor's parameters. When the developer first shared the project with a developer forum, the pitch rounded that to "roughly 60x smaller." Fewer parameters generally mean a model needs less memory and can run on more modest hardware, the same logic behind small local AI tools such as [Thinking Machines Lab's laptop-sized Inkling 2 Mini](https://badiyaa.com/inkling-2-mini-thinking-machines-laptop/).

## How it was trained without copyrighted music

Instead of commercial recordings, Dettato learned from synthetic instrument recordings plus Creative Commons solo instrument tracks mixed in using the URMP dataset. [The University of Rochester's URMP project page](https://labsites.rochester.edu/air/projects/URMP.html) describes that dataset as 44 classical pieces, from duets to quintets, assembled from separately recorded instrument parts, each supplied with a MIDI score and note annotations. The rest of Dettato's instruments come from blending around 20 soundfonts, using what the developer calls a "soundfont babies" technique.

The developer also explains a false start. Work began about 10 months ago, but the first version was scrapped around April because, in the developer's words, "I treated AMT as a data problem - but in reality, the bottleneck is architecture." The fix was a "pitch-lane model" that treats transcription as a local problem, similar to how the developer's earlier sheet-music reading model relied on a detector.

## The honest caveats

The launch page is unusually frank about limits. Dettato is weak at separating instruments correctly, and it produces performance MIDI aligned to a beat rather than tidy sheet music. For printable notation, the developer points readers to other tools. The demo track on the page is a public-domain recording of Sousa's The Stars and Stripes Forever by the United States Marine Band, a US government recording.

The stated mission of kobimusic is "to help musicians, not replace them." The developer's closing argument is that avoiding copyrighted recordings did not hold the model back but pushed it toward a better design. Anyone can test that claim directly: the launch page lets visitors upload their own audio and download the resulting MIDI. For more on how AI is landing on personal devices, see Badiyaa's report on [Pete Warden's bet on local AI with Moonshine Voice](https://badiyaa.com/pete-warden-moonshine-voice-local-ai/).

## Sources
- [Introducing Dettato](https://kobi.music/research/dettato), kobimusic, 2026
- [University of Rochester Multi-Modal Music Performance (URMP) Dataset](https://labsites.rochester.edu/air/projects/URMP.html), University of Rochester, 2018
- [MIDI](https://en.wikipedia.org/wiki/MIDI), Wikipedia, 2026

*Image: David J, CC BY 2.0 via Wikimedia Commons*

*This article was produced with AI assistance and reviewed by our Editorial Desk. [How we work](https://badiyaa.com/editorial-policy/).*
