Every wrist-worn gesture interface runs into the same wall. More electrodes give you a better read on what the hand is doing. More electrodes also mean more data to move over a radio, and the radio is the thing that drains the battery and sets the size of the band.
This paper picks the compression side and tests whether anything was lost.
EMG Tokenization Preserves Movement Information at a Fraction of the Bitrate was posted to arXiv on 23 October 2026 (arXiv:2610.32674) by Raul C. Sîmpetru, Amin Olamazadeh, Devon Rohlf, Vlad Cnejevici and Alessandro Del Vecchio. The Comments field gives 11 pages, 7 figures, a 12-page supplement, and Supplementary Data 1–9 as CSV ancillary files — but names no venue, so this is an unreviewed preprint.
The ancillary CSVs are worth a sentence on their own. Shipping nine machine-readable data files with a preprint means the numbers below can be checked by someone who did not run the study, which is rarer than it should be.
The constraint, stated plainly
Wireless electromyography (EMG) interfaces must balance electrode coverage against transmission bandwidth.
That is the engineering problem in one line. Everything else is a response to it.
What was built
Two causal tokenizers, both fitted without pose labels:
- An encoder pretrained to reconstruct EMG spectra — a learned representation, trained only to rebuild the signal.
- One based on SPAR, described as an engineered spectral-spatial representation of neighbouring electrodes — hand-designed rather than learned.
Causal is the load-bearing word. A causal tokenizer uses only past samples, so it can run on a live stream. An acausal one needs to see ahead, which is fine for offline analysis and useless for an interface you are wearing.
And fitted without pose labels means neither tokenizer was told what the hand was doing. They compress EMG as EMG. Whether hand-pose information survives that compression is the question, not an objective they were optimised for — which makes a positive result considerably more interesting.
The experiment and the numbers
320-channel EMG from nine participants, personalised and within-session, comparing tokenized against continuous features using matched decoder architectures and training procedures.
Averaged across decoders, tokenization changed mean 3D hand-pose error by:
| Representation | Change in mean error | 95% CI |
|---|---|---|
| SPAR | +0.01 mm | −0.12 to 0.14 |
| Learned | +0.45 mm | 0.10 to 0.82 |
Both met a 1 mm non-inferiority criterion.
And the payoff:
- 41× smaller stored token arrays relative to raw EMG
- 36× smaller relative to continuous SPAR features
They also checked generalisation beyond their own recordings: on sampled windows of the public emg2pose benchmark’s test partition, absolute differences in mean error between learned tokens and continuous features were reported, with the arXiv record truncating before the figure.
Why the statistics here deserve credit
Three things this paper does that similar work usually does not.
It declared a non-inferiority criterion. “1 mm” was set as the bar, and the results are reported against it. That is the correct statistical frame for a compression claim: you are not trying to prove the compressed version is better, you are trying to establish it is not meaningfully worse. Testing for a difference and failing to find one is not the same thing, and conflating them is the standard error in this genre.
It gives confidence intervals, not just means. The SPAR interval spans zero (−0.12 to 0.14), so that result is consistent with no change at all. The learned interval does not (0.10 to 0.82), so the learned tokenizer is reliably slightly worse — still inside the bound, but the degradation is real rather than noise. Reporting that distinction instead of rounding both to “no difference” is the honest version.
The engineered representation won. SPAR, hand-designed from neighbouring-electrode structure, came out indistinguishable from continuous features while the learned encoder took a small measurable hit. The authors had every incentive to find the opposite and reported what they got.
The caveats that bound it
Nine participants, personalised, within-session. Each of those words removes a claim. Nine is small. Personalised means the decoder was fitted per person, so this is not cross-user generalisation. Within-session means no electrode reapplication — and electrode shift between sessions is the single largest practical problem in wearable EMG, since moving a band a centimetre changes which muscles each channel sees. None of this is hidden, but it means the result is about information content under compression, not about a product that works tomorrow morning.
320 channels is a laboratory array, not a watch strap. The relevance to consumer devices is indirect: it shows that dense coverage can be compressed, which is an argument for building denser arrays, not a demonstration of one in a wristband.
“Stored token arrays” is about payload size, not measured radio power. Transmission energy tracks payload closely enough that a 41× reduction is clearly significant, but the paper’s figure is a data-volume ratio and should be quoted as one.
What it changes
The useful reframing is that bandwidth may be the wrong thing to economise on by reducing electrodes. The intuitive trade — fewer channels to fit the radio budget — throws away spatial information that cannot be recovered. Compressing a dense array instead throws away redundancy, and the measurements here suggest the hand-pose content is in the part that survives.
For anyone building gestural instruments or accessible input devices, the practical reading is: if you are considering a sparse array because of a wireless budget, the budget may be solvable at the representation layer instead. And if you want to reproduce this, the nine CSVs are sitting on the arXiv record.
Sources
- EMG Tokenization Preserves Movement Information at a Fraction of the Bitrate — arXiv:2610.32674
- arXiv PDF
- emg2pose benchmark — Meta Research
- arXiv listing for cs.HC