dettato uses roughly 21 million parameters for audio-to-MIDI conversion, making its model about 60 times smaller than muscriptor 1.3b. Its engineering appeal is clear, but its benchmark claims and data licensing still require independent scrutiny.
Converting music into editable MIDI poses more than a basic transcription challenge. The real obstacle is often the cost of running the model.
For a large model running in the cloud, a few seconds of latency and a few cents per inference may be acceptable. In a local workflow, however, users may need to process batches of recordings, connect the output to a DAW, or run an automated pipeline. Whether the model fits in GPU memory or can use a CPU or Apple Silicon determines whether it is practical.
That is what makes dettato worth examining.
The developer has released dettato, an open-weight audio-to-MIDI model, along with a research page. According to the project, dettato is about 60 times smaller than muscriptor 1.3b, delivering an estimated 60-fold efficiency gain. It also reportedly outperforms muscriptor 1.3b across every test set used by the project.
In other words, the project shifts the debate from which model is largest to how much usable output each parameter and unit of compute can produce.
60 Times Smaller Means More Than Lower GPU Usage
Start with deployment requirements.
muscriptor 1.3b is a 1.3B-parameter model, while dettato has about 21 million parameters, a difference of nearly 60 times. That is more than a matter of moving weights from storage into GPU memory; it changes the engineering options:
- Local processing becomes more practical without depending on a cloud API.
- Latency and inference costs are easier to control during batch transcription.
- The model has a better chance of running on CPUs, edge devices, or Apple Silicon.
- Developers can keep audio on the local machine rather than uploading it to a third-party service.
Using a rough FP16 estimate, 1.3B parameters alone require about 2.6 GB for the weights, compared with approximately 42 MB for 21 million parameters. Actual deployment also involves the runtime, caching, and input and output handling, so file size is not the whole picture. Even so, the difference in scale is substantial.
The cost of audio-to-MIDI work often extends beyond the model itself. Music files can be lengthy, and a job may involve dozens of tracks at once. With fewer parameters, teams can use larger batch sizes, greater concurrency, and more aggressive retry strategies. For an individual developer or small team, completing the workload matters more than improving a leaderboard score by half a point.
dettato is therefore pursuing more than a place at the top of an accuracy chart. It aims to turn the model from an expensive cloud component into a lightweight module that can be embedded in a local toolchain.
Beating muscriptor 1.3b
According to the project summary, dettato outperforms muscriptor 1.3b on every test set. The project also claims a roughly 60-fold reduction in model size and describes the result as approximately 60 times more efficient.
The combination of those claims is what makes the release interesting.
When a small model surpasses a larger one, the first question is whether the evaluation was fair. Were the training data different? Was preprocessing different? Could the test sets contain data leakage? Do the metrics reflect real-world use?
dettato provides a research entry point but has not yet answered all of these questions. The project’s conclusions are a useful signal, not a final verdict. In audio tasks, pitch, timing, pedal control, dynamics, and track separation can all materially affect the results. A single aggregate score may also conceal different types of errors.
The release reached the Hacker News front page, where it received 149 points and 26 comments. The developer has also made the model weights and research page available. For a specialist model, this is more useful than a page containing only a paper abstract. Users can download and run it, then determine whether the reported results hold up.
A sufficiently complete independent benchmark is still missing. At minimum, both models need to process the same set of real recordings on the same hardware, with identical input formats and post-processing conditions. Only then can a more accurate model be distinguished from an easier-to-use system.
No Copyrighted Recordings Used for Training
The project says dettato was not trained on copyrighted recordings.
That distinction matters for music AI.
Audio training datasets are generally harder to assemble than text corpora. Rights in a recording may involve performers, songwriters, recording producers, and distributors. Even if a model never directly copies a passage of audio, it may still learn melodies, performances, or timbral characteristics from protected works.
Unclear data provenance directly affects whether a model can be used in a commercial product. It also determines whether researchers can inspect the data and reproduce the training process.
Avoiding copyrighted recordings reduces one category of direct licensing risk, but two concepts still need to be separated:
- No copyrighted recordings were used.
- The project has complete and clear rights to all training data.
The first does not imply the second. Training material may also include MIDI annotations, synthesized audio, existing labeled datasets, data-generation code, and third-party tools or assets. Open model weights do not automatically eliminate risk throughout the data pipeline.
Compared with music recordings of uncertain provenance, a public and auditable data process is better suited to research and reuse. Developers can continue asking what data was used, how it was cleaned, where the labels came from, and whether particular works were excluded.
What Open Weights Actually Unlock
The value of open weights is not simply that the model is free.
They turn the model from an opaque API into a local component that developers control. Users can adjust inference settings, replace post-processing, connect it to an existing DAW, and decide whether data leaves their machine. For music-production workflows that often include unreleased work, local execution is not a bonus feature but part of the product design.
dettato could fit into several specific workflows:
- Convert humming, piano recordings, or live performances into MIDI for further editing in a DAW.
- Extract melodies and harmonies from a media library in batches for search or tagging.
- Produce editable snippets quickly for video scores, podcasts, and game prototypes.
- Join an automated transcription pipeline before passing the result to a human editor for correction.
Open weights also make community validation easier. When developers find edge cases, they can inspect model behavior, adjust quantization, or add data for particular instruments without waiting for a vendor to expose an API.
Research page: https://kobi.music/research/dettato
The Small Model’s Win May Not Be about the Model
dettato’s advantage may come primarily from data cleaning, training objectives, and evaluation design rather than its network architecture.
Audio-to-MIDI transcription contains several distinct subtasks: rhythm recognition, pitch tracking, chord inference, pedal handling, and note and velocity annotation. More parameters do not guarantee better performance on every task. A smaller model trained specifically for MIDI, with cleaner data and more suitable post-processing, may outperform a general model with more parameters but less focused training objectives.
A model that is 60 times smaller does not necessarily run 60 times faster. Smaller weights can reduce loading time and memory use, but real-world speed also depends on operators, sequence length, hardware, batch size, and input decoding.
The project describes its efficiency gain as approximately 60 times. Without public end-to-end measurements, however, a size ratio should not be presented as a speed ratio.
Three sets of results would help establish the case:
- End-to-end runtime on the same machine with the same audio.
- Separate metrics for pitch, timing, dynamics, and other dimensions rather than only an aggregate score.
- Failure cases involving drums, piano, melody, polyphony, vocals, and recordings with pedal data.
If those results hold, dettato’s significance would not be that “21M beat 1.3B.” It would be that a specialist model found a more sensible allocation of resources.
Editorial Perspective
dettato matters not because a 21-million-parameter model defeated a 1.3B model, but because it reframes specialist AI in terms of output per unit of resource. Earlier discussions of audio-to-MIDI tended to focus on model size, parameter count, and isolated accuracy scores. Real deployments must also account for local latency, batch-processing costs, privacy, data provenance, and time spent on manual corrections. A smaller model with clearer licensing that can run offline may be more useful for a product even if it still makes errors on complex material.
What distinguishes this release is the combination of open weights, a public research page, and community testing on Hacker News. It gives the claim of “60 times more efficient” a downloadable and runnable target rather than leaving it solely in the announcement. The likely direction for music tools is more small, open models paired with specialist data, rather than blanket reliance on general-purpose large-model APIs.
The 60-fold claim is still primarily the project’s own. Whether it refers to file size, compute, or efficiency across the full pipeline requires an independent benchmark, while copyright-friendly training must be demonstrated through specific datasets and license files. dettato is worth trying, but the broader conclusions should wait for evidence.