⏱️ Reading time: 6 min

A model with just 20 million parameters has just outperformed an open-source rival 65 times its size at the task of converting audio into MIDI. It’s called Dettato, built by an independent developer who documented the whole process, the benchmarks and the architecture decisions at kobi.music.

📑 En este artículo
  1. What Audio-to-MIDI Transcription Is and What Dettato Solves
  2. How It Reaches State of the Art With 65 Times Fewer Parameters
  3. The Numbers Against Mirelo, MuScriptor and YourMT3+
  4. Trained Without a Single Copyrighted Recording
  5. What the Community Says: Limits Before Taking It to Production
  6. When Not to Use Dettato
  7. What This Means for a Developer
  8. References

Dettato tackles a problem known as automatic music transcription (AMT): taking an audio recording, say a song or a live performance, and turning it into MIDI events: which note sounded, exactly when it started and when it ended. It’s the step before turning audio into sheet music, into a file a synthesizer can play, or into data another tool can analyze (source).

What Audio-to-MIDI Transcription Is and What Dettato Solves

Think of audio transcription like separating a conversation recorded in a room full of people: you have to decide who spoke, when they started and when they stopped. With musical audio the problem is similar, except the “voices” are overlapping notes from several instruments playing at once, often on top of percussion.

According to its creator, Dettato reaches state-of-the-art among open models and lands just above Mirelo A2M v1.1 Pro, a closed model, using only 20 million parameters versus the 1,300 million of MuScriptor 1.3B: roughly 65 times fewer (source). The model is open source and its outputs can be used commercially.

How It Reaches State of the Art With 65 Times Fewer Parameters

The creator says he started working on Dettato about 10 months ago, but scrapped that first version in April 2026. The reason wasn’t a lack of data: it was the architecture. He had treated automatic transcription as a data problem, that is, training on more recordings, when the real bottleneck was how the model represents the problem (source).

He draws a parallel to optical music recognition (OMR), a field where he says he built his first model. Both fields started with classical algorithms and later added giant LLM-like components, which end up overfitting to synthetic data because they tackle the problem globally. Reformulated as a local problem, with one model per pitch lane (what the author calls a “pitch-lane model”), the size came down without losing accuracy (source).

Dettato uses 20M parameters, 65 times fewer than MuScriptor 1.3B. Foto de Michel Didier Joomun en Unsplash

The Numbers Against Mirelo, MuScriptor and YourMT3+

The main comparison uses three benchmarks (PHENICX, RWC Classical and RWC Royalty-Free) and averages three metrics into a composite score: onsets (whether the model detects the exact instant each note starts), frames (whether it tracks pitch correctly while the note sounds) and offsets (whether it detects when the note ends). There’s also a separate metric for percussion, drums (source).

ModelParametersCompositeOnsetsFramesOffsetsDrums
Dettato20M60.870.879.140.952.5
Mirelo A2M v1.1 Pro (closed)not published56.167.175.740.141.4
MuScriptor1,300M47.457.071.231.430.1
YourMT3+46M45.758.263.930.530.0

It’s not a clean sweep. Breaking it down by benchmark, Dettato reaches 91.3 on Bach10 and 64.8 on PHENICX-Anechoic, but on RWC Classical the closed model Mirelo A2M v1.1 Pro scores 65.8 against Dettato’s 64.4: the only published category where the closed model wins (source). Mirelo also didn’t report results on several of those benchmarks (Dagstuhl ChoirSet, RWC Popular, RWC Genre, RWC Jazz and Slakh2100), so the full comparison only exists for Dettato, MuScriptor and YourMT3+.

Trained Without a Single Copyrighted Recording

The least technical and most relevant part for anyone who wants to use Dettato in a product is how it was trained. The mix combines synthetic instrument recordings with solo parts under a Creative Commons license taken from the URMP dataset; the rest of the instruments are generated by combining about 20 soundfonts with a technique the author published before, called “soundfont babies”. No copyrighted recording went into training (source).

📌 Note: this isn’t the same as saying the model is perfect: it’s a guarantee about the origin of the training data, not about the accuracy of every transcription.

What the Community Says: Limits Before Taking It to Production

The launch was discussed on Hacker News under the title “Show HN: I made an audio-to-MIDI model 60x more efficient”. Two comments there add something the original announcement doesn’t mention.

A user identified as gpugreg pointed out a practical, not technical, problem: if the server is slow to respond, an impatient person might resubmit the same file multiple times and end up paying several times for the same conversion. That’s why he recommended implementing request deduplication on the site before scaling the service (source).

Another user, mycall, questioned how well the model generalizes to percussion sounds: training on the author’s own percussive sounds leaves out a much wider range, he estimates more than 10,000 possible timbres, not all of them percussion, than a small set of samples can cover (source). It’s exactly the kind of limitation that doesn’t show up in an aggregate benchmark, but that matters if the input audio has unconventional instruments.

When Not to Use Dettato

The creator himself sets a clear limit on the model’s scope: he acknowledges it’s hard to judge real-world accuracy for a specific use case without testing it, and admits that correctly separating instruments is one of the points where Dettato performs worse (source).

Also, the output is time-aligned performance MIDI (what’s called “performance MIDI”), not readable sheet music. Dettato doesn’t replace a sheet-music transcriber: it solves a different problem, that of raw MIDI events. For anyone who needs musical notation, the author himself recommends two tools built for that specific purpose, Songscription and Klangio, instead of forcing Dettato’s output into that use case (source).

Training combines the CC-licensed URMP dataset with custom soundfonts, without copyrighted recordings. Foto de Ivan Jermakov en Unsplash

What This Means for a Developer

For anyone building a transcription feature, an ear-training app, a DAW plugin, an instrument-practice education tool, Dettato’s size changes the cost equation. A 20-million-parameter model runs on a CPU or a modest GPU, while MuScriptor needs more memory and compute for a smaller accuracy gain (source). That matters especially for a small team that can’t afford large GPUs in production.

The open license and training without copyrighted recordings also simplify legal review before integrating it into a commercial product, something that’s rarely true of audio datasets where the origin of each recording isn’t entirely clear. If the project needs MIDI data for analysis, automatic accompaniment or music practice, Dettato is a direct candidate to try first for its size-to-accuracy ratio. If the end goal is generating readable sheet music for a musician, it’s worth evaluating Songscription or Klangio in parallel, because Dettato doesn’t solve that problem and was never meant to.

References

Featured image: Foto de Bruno Croci en Unsplash

📬 Get new articles by email

We only email about big articles (1-2 a month).

How each article is made (sources, testing and editorial criteria): How we work.

Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.

Leave a comment

El Solitario Editorial

Article prepared by the El Solitario editorial team from cited sources, under the editorial responsibility of Luis Navarro. More details in "How we work".

0 Comments

Leave a Reply

Avatar placeholder

Your email address will not be published. Required fields are marked *

You can include code inside <code>…</code> or, for several lines, <pre><code>…</code></pre>.

This site uses Akismet to reduce spam. Learn how your comment data is processed.