TinyMyo: a compact EMG model runs on a microcontroller
Researchers report a 3.6-million-parameter EMG encoder that handles several tasks and, they say, is the first EMG foundation model deployed on an ultra-low-power microcontroller.
In plain words
Muscle-signal systems often need large models that cannot run on small wearable hardware. TinyMyo is a much smaller model trained on public muscle recordings to handle several jobs, such as recognizing hand gestures. Its authors report it can run directly on a low-power chip, so no cloud connection is needed.
What was announced
The model, described in an arXiv paper, is a Transformer encoder pre-trained with self-supervised masked reconstruction on public EMG datasets, then adapted to different tasks with minimal changes to the prediction head. The authors evaluated it on hand gesture classification, hand kinematic regression, speech production, and speech recognition. For gesture classification they report accuracies of 87.98% on NinaPro DB5, 97.1% on UCI-EMG, and 96.6% on EPN-612, which they state outperforms previously reported EMG foundation models.
On the GAP9 ultra-low-power microcontroller, the authors measured 0.785 s inference time and 44.91 mJ per inference at an average power of 57.18 mW. A smaller 1.9-million-parameter variant processed 5 s and 1 s input windows in 0.496 s and 0.089 s respectively. The team says this is, to their knowledge, the first deployment of an EMG foundation model on an ultra-low-power microcontroller, a step toward on-device motor intent decoding and human-machine interaction. These are self-reported results from the paper's abstract and have not been independently confirmed.
What it does not show
- Evidence is limited to the paper's abstract; full methods, baselines, and evaluation details were not available for verification.
- Performance and power figures are the authors' own measurements and have not been independently confirmed.
For security and IT teams
- Scope
- The reported deployment targets the GAP9 ultra-low-power microcontroller, with results covering gesture, kinematic, and speech-related EMG tasks rather than a single application.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.