Myovox: Reading Speech from the Muscles of the Face

# Myovox: Reading Speech from the Muscles of the Face
Source: [https://arxiv.org/html/2609.17548](https://arxiv.org/html/2609.17548)
###### Abstract

Myovox, from*myo*\(muscle\) and*vox*\(voice\), decodesopen\-vocabularyEnglish text from 31\-channelsurface electromyography \(sEMG\)recorded from the muscles of the face duringvocalized speech\. It takes the single\-subject*emg2speech*General Corpusfrom a published 51\.17% word error rate to 18\.53%, in three separable moves, each measured in isolation\. First, I recover the open\-vocabulary decode settings missing from the public release and reach a faithful40\.63%WER/ 39\.02%PERbaseline whose phone error rate matches the published one to within 0\.8 points, so theacoustic modelis reproduced faithfully\. Second, I replace thecausalencoder with abidirectionalConformertrained by a four\-termcross\-modal distillationagainst the parallel audio’sWavLM\-Largelayer\-9 features, reaching26\.14%WER/ 22\.34%PERfrom theelectromyographyalone\. Third, Iensembletwo acoustic models, union their multi\-scalenn\-bestlists, andrerankwith aQLoRA\-fine\-tuned 7Blanguage model, reaching18\.53%WER, the best result reported on this corpus, though not the best reported for sEMG\-to\-text on other corpora \(Section[2](https://arxiv.org/html/2609.17548#S2)\)\. I then report the negative result that bounds the whole approach: reranking is exhausted at 18\.5% because the binding constraint is the electromyographic*acoustic*phone error rate \(∼\\sim20\.9%\), not the language model\. The correct words are simply absent from the acoustic posteriors, so no reranker can reach the 9\.30%nn\-bestoracle\. All test numbers are on the 400\-sentence held\-outtest setunder the authors’ official 8,500 / 760 / 400sequential split; every hyperparameter is tuned once onvalidationand applied once to test\.

## 1Introduction

When you speak, the decision to say a word reaches the muscles of your face, jaw, and throat long before any sound leaves your mouth\. Those muscles fire, and the firing shows up as faint electrical voltages on the skin\. Myovox reads that electricity\. A grid of 31 sensors records the voltages a few thousand times a second, and a chain of models turns the recording into a line of English text\. The name is literal:*myo*for the muscle,*vox*for the voice the muscle was reaching for\.

The long\-term goal ofelectromyographicspeech decoding is to help people who cannot speak because of conditions such as ALS or a laryngectomy\. This report does not demonstrate that\. Myovox was trained entirely on a healthy speaker who vocalized normally, so I cannot claim that it works for people who have lost their voice\. What I can say is that it demonstrates a different use case: allowing a healthy person to communicate without producing audible speech, where the intention to speak is decoded from facial muscle activity instead of sound\. I view this as a stepping stone toward truesilent\-speechsystems rather than the destination\.

The report is organized around a single discipline: start from a strong published result, and measure the effect of each change in isolation before applying the next\. That is what makes the final number decomposable into a decode fix, a modeling gain, and an extraction gain, rather than a single undifferentiated delta\. The starting point is the*emg2speech*work of Gowda et al\.\[Gowdaet al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib1)\], whose Appendix D\.4 \(“emg2text”\) reports 51\.17%WERand 38\.19%PERon the General Corpus\. The three sections that follow trace three moves: a decode correction that recovers most of the published gap while matching the original phone error rate \(Section[4](https://arxiv.org/html/2609.17548#S4)\); a full\-context acoustic model trained to imitate the parallel audio \(Section[5](https://arxiv.org/html/2609.17548#S5)\); and a decode\-time stack of ensembling,nn\-best union, and language\-model reranking \(Section[6](https://arxiv.org/html/2609.17548#S6)\)\. Table[1](https://arxiv.org/html/2609.17548#S1.T1)is the whole story in one place\. Section[7](https://arxiv.org/html/2609.17548#S7)carries the report’s main finding: why the system stops improving at 18\.5%, and where the next gain has to come from\. Terms of art are collected in the Glossary \(Section[10](https://arxiv.org/html/2609.17548#S10)\), and every occurrence of a glossary term in the text links to its entry; the objections the project actually drew are answered in Section[11](https://arxiv.org/html/2609.17548#S11)\.

Two numbers run through everything\.*Word error rate*\(WER\) is what a reader cares about: the fraction of words wrong after substitutions, insertions, and deletions\.*Phone error rate*\(PER\) is the honest gauge of the acoustic model: the fraction ofphonemeswrong, measured bygreedy decodingwith nolexiconand no language model\.PERcannot be improved by a better decoder or a better language model\. As it turns out,PERis what limits the entire system\.

Table 1:The whole report in one table\. All test numbers are on the 400\-sentence held\-out set under the authors’ 8,500 / 760 / 400 sequential split\.PERis the decoder\-independent greedy CTC phone error rate\. Bold marks each stage’s headline word error rate\.†Phone error rate of the acousticensemblebeneath the pipeline\. Reranking operates on words and cannot changePER, which is precisely the point of Section[7](https://arxiv.org/html/2609.17548#S7)\.

## 2Related work

Two non\-invasive routes lead from a person’s intention to speak to a line of text, and they differ in where along the motor chain they tap\.*Surface electromyography*reads the muscles of the face and neck as they articulate, which is the last link before sound\.*Electro\- and magnetoencephalography*read cortex, which is many links earlier\. Myovox belongs to the first family\. Table[2](https://arxiv.org/html/2609.17548#S2.T2)places it against both\. Invasive intracortical decoders are excluded throughout: they are the performance ceiling, but they are not an alternative Myovox competes with\.

### Surface EMG to text: the Gaddy line\.

The reference corpus for open\-vocabulary sEMG speech decoding is the one released by Gaddy and Klein\[Gaddy and Klein,[2020](https://arxiv.org/html/2609.17548#bib.bib16)\]: eight facial channels at 1 kHz, roughly nineteen hours from a single English speaker, recorded in bothsilentandvocalizedmodes with audio available for the vocalized half\. Their first system synthesized speech from silent EMG and was scored by transcribing that speech; a Transformer front\-end and an auxiliary phoneme loss then cut the open\-vocabulary error substantially\[Gaddy and Klein,[2021](https://arxiv.org/html/2609.17548#bib.bib17)\]\. Gaddy’s thesis\[Gaddy,[2022](https://arxiv.org/html/2609.17548#bib.bib18)\]moved from synthesis to direct EMG\-to\-text and set the prior state of the art at 28\.8%WERon silent EMG and 23\.3% on vocalized EMG\. MONA LISA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]then combined cross\-modal alignment \(a contrastive objective binding EMG and audio latents, plus training on audio\-only LibriSpeech\) with an LLM reranker, and reached12\.2%WERon silent EMG and 3\.7% on vocalized EMG\. Both halves of that system have direct descendants here: thecross\-modal distillationof Section[5](https://arxiv.org/html/2609.17548#S5)and theLIFT\-style reranker of Section[6](https://arxiv.org/html/2609.17548#S6)\. A recent line asks the harder question of what is achievable from unvoiced EMG with no paired voiced recordings or audio at all\[Mohapatraet al\.,[2025](https://arxiv.org/html/2609.17548#bib.bib20)\], which is the setting a patient would actually present\.

### Myovox is not the best sEMG\-to\-text system; it is the best one on this corpus\.

This distinction matters enough to state before the table rather than after it\. MONA LISA’s 3\.7% on vocalized EMG is 15 points better than the 18\.53% reported here, and it was obtained two years earlier\. The two numbers are not measured on the same recordings and are not directly comparable: the Gaddy corpus is eight channels and about nineteen hours with both speaking modes from one speaker, while the*emg2speech*General Corpusis 31 channels and 9,660 sentences from a different speaker in one mode; MONA additionally trains against an audio\-only corpus that Myovox does not use\. What can be said cleanly is narrower: on the*emg2speech*General Corpus, 18\.53% is the best result I am aware of, against a published 51\.17%\. Whether the residual gap to MONA is a property of the corpus, of thecovariance front\-end, or of the model is precisely what a single corpus cannot resolve;[11](https://arxiv.org/html/2609.17548#S11)takes the question up directly\.

### Non\-invasive brain to text\.

The parallel effort decodes cortex rather than muscle\. Brain2Qwerty\[Lévyet al\.,[2025](https://arxiv.org/html/2609.17548#bib.bib21)\]decoded typed sentences from 35 volunteers, reaching a 32%CERfrom MEG and 67% from EEG\. Its successor\[Zhanget al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib22)\]scaled the data to roughly 22,000 sentences across nine subjects and reached 39%WERfrom MEG\. Two things about it are worth a reader’s attention here\. The first is architectural convergence: independently of this work, that system also settles on aConformerencoder trained withCTCand aLoRA\-fine\-tuned LLM on top: the same two components as Sections[5](https://arxiv.org/html/2609.17548#S5)and[6](https://arxiv.org/html/2609.17548#S6), arrived at from a different signal\. The second is that it independently reproduces the finding of Section[7](https://arxiv.org/html/2609.17548#S7)\. Their encoder alone reaches 55%WER; annn\-gramlanguage modeltakes it to 43%; the fine\-tuned LLM takes it to 39%\. There, the LLM*worsens*the character\-level metric \(31%CERagainst the encoder’s 28%\) even as it improves the word\-level one, and the authors conclude that final performance is dominated by upstream encoder quality\. What transfers to Section[7](https://arxiv.org/html/2609.17548#S7)is that conclusion, not the mechanism, and the difference is worth stating: their LLM generates the sentence autoregressively and is conditioned on MEG\-derived word embeddings as well as on the encoder’s characters, so it can and does degrade the character metric, whereas the reranker of Section[6](https://arxiv.org/html/2609.17548#S6)sits downstream of the greedy CTC output and cannot movePERat all\. Their own text\-only ablation, which leaves the LLM nothing but the encoder’s characters, is the closer analogue, and it is worth about sixWERpoints over the encoder alone\. Either way a language model bolted onto a weak encoder buys a bounded, single\-digit number ofWERpoints\. That is the same wall, found on a different signal\.

Table 2:Non\-invasive speech and language decoding, grouped by corpus\. Numbers are as reported by each work and are*not*comparable across groups: corpora, electrode counts, speaking modes, subject counts, and metrics all differ\. “Aud\.” marks whether parallel or external audio is used as a training signal \(never at inference\)\. Invasive intracortical systems are omitted by design\.SystemModeVocabAud\.WER↓\\downarrowPER↓\\downarrowCER↓\\downarrow*Facial sEMG→\\rightarrowtext: Gaddy corpus*\(8 ch, 1 kHz,∼\\sim19 h, 1 speaker\)\[Gaddy and Klein,[2020](https://arxiv.org/html/2609.17548#bib.bib16)\]Gaddy & Klein 2020\[Gaddy and Klein,[2020](https://arxiv.org/html/2609.17548#bib.bib16)\]∗silentopen✓68\.0––Gaddy & Klein 2021\[Gaddy and Klein,[2021](https://arxiv.org/html/2609.17548#bib.bib17)\]∗silentopen✓42\.2––Gaddy 2022\[Gaddy,[2022](https://arxiv.org/html/2609.17548#bib.bib18)\]silentopen✓28\.8––Gaddy 2022\[Gaddy,[2022](https://arxiv.org/html/2609.17548#bib.bib18)\]vocalizedopen✓23\.3––MONA LISA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]silentopen✓12\.2––MONA LISA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]vocalizedopen✓3\.7––*Facial sEMG→\\rightarrowtext: emg2speech General Corpus*\(31 ch, 5 kHz, 9,660 sentences, 1 speaker\)\[Gowdaet al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib1)\]Gowda et al\. 2026, App\. D\.4\[Gowdaet al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib1)\]vocalizedopen✓51\.1738\.19–Myovox, corrected decode \(§[4](https://arxiv.org/html/2609.17548#S4)\)vocalizedopen✓40\.6339\.02–Myovox, Conformer \+ distillation \(§[5](https://arxiv.org/html/2609.17548#S5)\)vocalizedopen✓26\.1422\.34–Myovox, Conformer, EMG\-only \(§[5](https://arxiv.org/html/2609.17548#S5)\)vocalizedopen–26\.1023\.71–Myovox, full pipeline\(§[6](https://arxiv.org/html/2609.17548#S6)\)vocalizedopen✓18\.5320\.90†–*Non\-invasive brain→\\rightarrowtext*\(typed sentences, healthy volunteers\)Brain2Qwerty v1, EEG\[Lévyet al\.,[2025](https://arxiv.org/html/2609.17548#bib.bib21)\]typingopen–––67Brain2Qwerty v1, MEG\[Lévyet al\.,[2025](https://arxiv.org/html/2609.17548#bib.bib21)\]typingopen–––32Brain2Qwerty v2, MEG\[Zhanget al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib22)\]typingopen–39–31
∗These two systems*synthesize speech*and are scored by transcribing the synthesized audio, so theirWERis an intelligibility measure rather than a direct text\-decoding score\. Brain2Qwerty v1 reports onlyCER; v2 reports both\. Brain2Qwerty v1 averages 35 subjects \(19%CERfor the best\); v2 averages 9 subjects \(22%WERfor the best\)\. †ThePERof the acousticensemblebeneath the full pipeline, not of the pipeline: reranking operates on words and cannot changePER\.

### What is new here\.

Against that background, this report contributes three things, none of which is a new architecture\. It supplies the open\-vocabularydecode settingsmissing from the*emg2speech*release, without which that release decodes at roughly 75%WERrather than the 40\.63% reported here, and shows by a matchedphone error ratethat the acoustic model is unchanged \(Section[4](https://arxiv.org/html/2609.17548#S4)\)\. It prints a control that undercuts its own favoured component: an encoder trained onelectromyographyalone matches the audio\-distilled one at the word level, so the parallel audio is not a hidden crutch beneath the headline number \(Section[5](https://arxiv.org/html/2609.17548#S5)\)\. And it locates, with three independent lines of evidence, the point at which decode\-time machinery stops paying \(Section[7](https://arxiv.org/html/2609.17548#S7)\)\.

## 3The data

The recordings are the healthy\-subject*General Corpus*from the*emg2speech*release\. Its properties are summarized in Table[3](https://arxiv.org/html/2609.17548#S3.T3)\. Three facts about it shape the rest of the report\.

### It is one person\.

The entire system is trained and evaluated on recordings from a single healthy speaker\. I therefore do not know how well it generalizes to other people, anatomies, or speaking styles\. The results in this report should be read as a demonstration of what is possible for one speaker, not as evidence that the same performance will transfer to others\.

### It has parallel audio\.

While the subject spoke, a microphone recorded the voice at the same time; per\-sentence electromyography and audio durations correlate at 1\.000\. That audio is the training signal I lean on hardest in Section[5](https://arxiv.org/html/2609.17548#S5)\. It is also a crutch that will not exist atsilent\-speechinference, so I treat it as a training\-time\-only teacher and, at the end, check how much the system depends on it\.

### It is open\-vocabulary\.

The decoder works over a 34,546\-wordLibriSpeech\-derivedlexicon\[Panayotovet al\.,[2015](https://arxiv.org/html/2609.17548#bib.bib11)\], roughly five times the corpus vocabulary, so the system can emit words it never saw in training\. Twelve of the 2,429 test tokens falloutside the lexicon, a 0\.49%WERfloor that no amount of modeling can remove\.

Table 3:The*emg2speech*General Corpus\.Thesplitis the authors’ own, taken in recording order\. I keep it exactly\. Every knob in this report is tuned on the 760validationsentences and then applied, once, to the 400testsentences\. That discipline is the only thing that makes the final number mean anything\.

## 4Baseline: reproduction and a decode correction

### The acoustic model, unchanged\.

I keep the authors’acoustic modelas released\. The features are ashrinkage covarianceof the 31 channels \(thevec​\(E\)\\mathrm\{vec\}\(E\)representation whose geometry Gowda et al\. develop in their articulation\-decoding work\[Gowdaet al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib3), Gowda and Miller,[2025](https://arxiv.org/html/2609.17548#bib.bib2)\]\), taken over a 25 mswindow with a 20 ms hop, giving 50 frames per second\. The encoder is the releasedDualHeadTDSCTC: acausaltime\-depth\-separable convolution\[Hannunet al\.,[2019](https://arxiv.org/html/2609.17548#bib.bib5)\]with twoCTCheads\[Graveset al\.,[2006](https://arxiv.org/html/2609.17548#bib.bib7)\], one predicting 100HuBERTunits\[Hsuet al\.,[2021](https://arxiv.org/html/2609.17548#bib.bib8)\]and one predicting 40phonemes, coupled by a fixedP​\(phone∣unit\)P\(\\text\{phone\}\\mid\\text\{unit\}\)consistency table\. The training loss is0\.8⋅CTCunit\+0\.1⋅CTCphone\+0\.1⋅consistency0\.8\\cdot\\mathrm\{CTC\}\_\{\\text\{unit\}\}\+0\.1\\cdot\\mathrm\{CTC\}\_\{\\text\{phone\}\}\+0\.1\\cdot\\text\{consistency\}, adual\-CTCobjective\. Phoneme posteriors are turned into words by an open\-vocabularyweighted finite\-state transducer,HLG=H∘L∘G\\hyperlink\{gl:hlg\}\{\{\\color\[rgb\]\{0,0,0\}\\mathrm\{HLG\}\}\}=H\\circ L\\circ G, built withk2 and icefall\[Povey and others,[2023](https://arxiv.org/html/2609.17548#bib.bib12)\]\. It is the same graph the authors use, which is why my phone error rate can be compared to theirs directly\.

### What was missing\.

The public notebooks reproduce the model but not the decode\. They omit the handful of decode\-time settings that separate good posteriors from good words, and without them the released configuration decodes at roughly 75%WER\. Four things were missing, in rough order of importance\.

1. 1\.Blank penalty\.CTC posteriors are dominated by the blank symbol \(peak≈0\.92\\approx 0\.92\), and the release applies no penalty at all\. On the first 200 validation sentences at scale 1\.0, sweeping the penalty from 0 to 2 dropsWERfrom 77\.6% to 60\.6%\. This is the single largest lever in the entire baseline\.
2. 2\.Checkpoint selectionby validationPER, rather than by validation CTC loss, which moves testPERfrom 42\.9% to 39\.0%\.
3. 3\.A missingwords\.txt, which I regenerated fromlexicon\.txt\.
4. 4\.Acoustic scale\.The weight on the acoustic scores relative to the language model is unspecified in the release, which decodes at the default of 1\.0\. Tuning it on validation returns 1\.0 for this baseline, so it buys nothing here\. It matters for the Conformer of Section[5](https://arxiv.org/html/2609.17548#S5), where validation selects 0\.25\. I list it because the right value is not the same for both encoders, so a reader reproducing the decode has to set it deliberately rather than inherit it\.

I tuned the\(blank,scale\)\(\\text\{blank\},\\text\{scale\}\)pair jointly on validation and applied\(2\.0,1\.0\)\(2\.0,1\.0\)once to test\.

### Result\.

Table[4](https://arxiv.org/html/2609.17548#S4.T4)is the outcome\. The corrected decode reaches 40\.63%WER/ 39\.02%PERon test: 10\.5WERpoints below the published number, with the phone error rate matched to within 0\.8 points \(39\.0 against 38\.2\)\. That matchedPERis the point of this section\. Phone error rate is decoder\-independent, so if it lands on the published value then the acoustic model has been reproduced faithfully, and the entireWERgain is attributable to a correctly specified open\-vocabulary decode rather than to any change in the model\. Note that the published 51\.17% is not itself recoverable from the public release, which decodes at roughly 75%; the comparison rests on the matched phone error rate, not on a reproducedWER\. This checkpoint is thewarm startfor the encoder in the next section\.

Table 4:Baseline reproduction\. TheWERgain over the published number comes entirely from the decode; the phone error rate is matched, not beaten, which is the credibility argument\. Bold marks this section’s headline number\.

## 5The acoustic model: a full\-context Conformer taught by the audio

Two changes take the baseline from 40\.63 / 39\.02 to 26\.14 / 22\.34: giving the encoder full context, and teaching it to imitate the parallel audio\.

### Full context\.

The releasedtime\-depth\-separableencoder iscausal: it left\-pads bykernel−1\\text\{kernel\}\-1and never looks ahead\. That is the right choice forstreamingand the wrong one for offline transcription, where the whole sentence is available at once\. I replace it with abidirectionalConformer\[Gulatiet al\.,[2020](https://arxiv.org/html/2609.17548#bib.bib6)\]: four layers ofmulti\-head self\-attentioninterleaved withdepthwise convolution\(four heads, feed\-forward width 1024, convolution kernel 31\)\. The covariance front\-end, the two CTC heads, and the audio projectionwarm\-startfrom the baseline checkpoint of Section[4](https://arxiv.org/html/2609.17548#S4); the Conformer itself trains from scratch\.

### A four\-term cross\-modal objective\.

Only a better acoustic model lowers phone error rate, so this is where I spent the effort\. I pull the electromyographic encoder toward the parallel audio’sWavLM\-Largelayer\-9 features\[Chenet al\.,[2022](https://arxiv.org/html/2609.17548#bib.bib9)\], precomputed once for all 9,660 sentences, through a single linear projection into WavLM’s 1024\-dimensional space\. The construction is inspired by thecross\-modalapproach of Benster et al\.\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]\. On top of the three acoustic terms from the baseline, the objective adds threedistillationterms:

ℒ=0\.8​CTCunit\+0\.1​CTCphone\+0\.1​cons\.⏟acoustic \(as in the baseline\)\+0\.5​ℒL2\+0\.5​ℒInfoNCE\+1\.0​ℒrecCTC⏟cross\-modal distillation\\mathcal\{L\}=\\underbrace\{0\.8\\,\\mathrm\{CTC\}\_\{\\text\{unit\}\}\+0\.1\\,\\mathrm\{CTC\}\_\{\\text\{phone\}\}\+0\.1\\,\\text\{cons\.\}\}\_\{\\text\{acoustic \(as in the baseline\)\}\}\\;\+\\;\\underbrace\{0\.5\\,\\mathcal\{L\}\_\{\\text\{L2\}\}\+0\.5\\,\\mathcal\{L\}\_\{\\text\{InfoNCE\}\}\+1\.0\\,\\mathcal\{L\}^\{\\text\{CTC\}\}\_\{\\text\{rec\}\}\}\_\{\\text\{cross\-modal distillation\}\}\(1\)The three new terms are: \(i\) a masked, frame\-resampledL2L\_\{2\}regression of the projection onto WavLM layer\-9; \(ii\) a frame\-synchronousInfoNCEcontrast\[van den Oordet al\.,[2018](https://arxiv.org/html/2609.17548#bib.bib10)\]\(τ=0\.1\\tau=0\.1\) that sharpens each frame toward its own audio moment and away from the others; and \(iii\) a CTC loss through a small,*frozen*WavLM\-to\-phoneme recognizer \(LayerNorm, two 1\-D convolutions, a linear layer to 41 classes; trained to∼\\sim10%PERand then frozen\)\.

Term \(iii\) is the one that earns its place relative to a plain feature\-matching setup\. TheL2L\_\{2\}and contrastive terms pull the projection*close*to WavLM, but close is not the same as useful: a projection can be smooth and audio\-like and still not be phoneme\-decodable\. Forcing the projection through a frozen recognizer that already reads phonemes out of real audio makes “close” mean “decodable into the right phonemes\.” It guards the regression against smooth\-but\-empty blur\.

### Result, and a control that surprised me\.

Table[5](https://arxiv.org/html/2609.17548#S5.T5)gives the outcome, with validation\-tunedscale0\.25\. The full\-context, distilled Conformer reaches 26\.14%WER/ 22\.34%PERon test, acoustic\-only: a gain of 14\.49WERand 16\.68PERpoints over the baseline, present in the greedy phone error rate with no language model involved, and significant under apaired bootstrap\. The control is the interesting part\. An otherwise identical Conformer trained on electromyography alone, with no audio distillation at all, reaches 26\.10%WER, indistinguishable from 26\.14%\. At the word level the audio teacher buys essentially nothing\. The distillation does help phone error rate \(22\.34 against 23\.71\), but it is the full\-context encoder, not the audio crutch, that movesWER\. That is good news for the silent\-speech setting, where the parallel audio will not be available\.

Table 5:The acoustic model\. The full\-context encoder is what movesWER\. The audio distillation helpsPER, but at the word level an encoder trained on electromyography alone matches it\. Bold marks this section’s headline number\.

## 6The final pipeline: ensemble, union, and a language\-model reranker

The last stretch is all decode\-time\. The acoustic model does not change; what changes is how much is squeezed out of it\.

### An acoustic ensemble\.

Iaverage the per\-frame phone log\-probabilitiesof two encoders: the distilled Conformer of Section[5](https://arxiv.org/html/2609.17548#S5), and a second Conformer trained with heavier anti\-overfitting defenses \(stronger jitter and dropout, plus aBiLSTMaudio\-teacher frame\-level KL\)\. Averaging the two alone takes testWERfrom 26\.14 to 23\.47 at the best of theacoustic scalesswept, the ensemble decoding between 23\.5% and 25\.1% across them\. Worth noting for later: the augmented member’s phone error rate is no better than the distilled member’s, and averaging the two lowers greedyPERto 20\.9%, 1\.4 points below either member\. The gain is modest either way: 2\.7WERpoints at best, against the 14\.49 that full context bought in Section[5](https://arxiv.org/html/2609.17548#S5)\. How much of it follows from the improved argmax and how much from decode\-level diversity is taken up in[11](https://arxiv.org/html/2609.17548#S11)\.

### Multi\-scalenn\-best union\.

From the ensemble’slatticeI extractnn\-best listsat severalacoustic scalesand take their union\. The union lowers thenn\-best oracleWER, the score a perfect judge would obtain by always picking the best candidate in the pool, from 11\.94% for the best single\-scale list to 9\.30%, while the union’s single\-best candidate sits at 23\.26%\. The gap between 23\.26 and 9\.30 is the headroom arerankercan, in principle, recover: the reranker below takes 4\.7 points of it, and Section[7](https://arxiv.org/html/2609.17548#S7)explains why the remaining nine are out of reach\.

### The LIFT reranker\.

Ifine\-tuneQwen2\.5\-7B\-Instruct\[Qwen Team,[2024](https://arxiv.org/html/2609.17548#bib.bib15)\]withQLoRA\[Dettmerset al\.,[2023](https://arxiv.org/html/2609.17548#bib.bib14)\]\(4\-bit NF4 plus a rank\-16LoRAadapter\[Huet al\.,[2022](https://arxiv.org/html/2609.17548#bib.bib13)\]\) to map*\(candidate list \+ detected phonemes\)*to the reference, in theDCoND\-LIFTstyle\[Liet al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib4), Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]\. Two variants are chosen on validation\. The*free*variant writes its own correction, which lets it recover a candidate that is close but wrong, at the cost of being able to hallucinate\. The*constrained*variant must pick an existing candidate, so it cannot hallucinate, but neither can it rescue an answer that was never proposed\.

Leakageis the obvious risk with a language model on a single\-subject corpus, so I built three controls before trusting the number\. The reranker is trained only on the training split, and its training candidates are produced by two\-foldcross\-decoding: each half of the training data is decoded by a model that did not train on it, so the candidates look like realistic errors rather than the memorized∼\\sim1%WERthe model achieves on its own training set\. Six of the 400 test sentences are exact\-text duplicates of training phrases, so I report the score both with and without them \(18\.53 against 18\.75\)\. And averbatim\-recall auditmeasures how often free generation produces a reference that was*absent*from the candidate set: the count is zero, so no training text is leaking through the language model\.

Table 6:The final pipeline\. Both bootstrap confidence intervals exclude zero, so the rerank gain is significant\. The nine\-point gap between the oracle and the achievedWERis the subject of Section[7](https://arxiv.org/html/2609.17548#S7)\.Table[6](https://arxiv.org/html/2609.17548#S6.T6)collects the result: 18\.53%WER, with the reranker worth−4\.7\-4\.7WERpoints over the union’s single\-best output of 23\.26%, and both bootstrap intervals well clear of zero\.

## 7Why it stops at 18\.5%

This is the report’s central finding\. It is a negative one, and it tells the next person where to spend effort\. Reranking is exhausted at 18\.5%, and the limit is acoustic, not linguistic\. Three pieces of evidence point the same way\.

First, phone error rate responds to the interventions that moved word error rate only in small increments, and expensively\. The anti\-overfitting augmentation of the second ensemble member leaves it untouched\. A∼\\sim10%\-PERaudio teacher buys 1\.4 points \(23\.71 to 22\.34, Section[5](https://arxiv.org/html/2609.17548#S5)\), which the decoder then fails to convert into words at all\. A second full model, averaged in, buys 1\.4 more \(22\.34 to 20\.90, Section[6](https://arxiv.org/html/2609.17548#S6)\)\. Everything downstream of the full\-context encoder movesPERby 1\.44 points in total, against the 16\.68 that full context bought on its own\.

Second, the audio teacher cannot transfer its own quality\. The electromyography\-only Conformer matches the distilled one at the word level \(26\.10 against 26\.14, Section[5](https://arxiv.org/html/2609.17548#S5)\)\. A teacher that is itself only a∼\\sim10%PERrecognizer does not hand its phonetics to the muscle encoder\.

Third, and most concretely, there is a nine\-pointoraclegap that no reranker closes\. The union oracle is 9\.30%; the reranker reaches 18\.53%\. The residual is not a language problem\. For those utterances the reference words are*absent from theacoustic posteriors*: there is no candidate to select, and nothing for the constrained reranker to move toward\. Free generation that “fixed” them would be hallucination, and the audit shows it does not happen \(recall zero\)\.

The conclusion is that∼\\sim20\.9% acoustic phone error rate is the binding constraint\. Getting below roughly 10%WERon this task requires a better electromyographic acoustic model \(more data, multiple subjects, or a stronger front\-end that reads phonemes out of the raw signal\), not a bigger or smarter language model\. The entire decode\-time stack in Section[6](https://arxiv.org/html/2609.17548#S6)is, in effect, a careful way of extracting the most words possible from a fixed amount of phonetic information\. The same conclusion is reached independently, on an entirely different non\-invasive signal, by Brain2Qwerty v2\[Zhanget al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib22)\], whose fine\-tuned LLM improvesWERwhile*worsening*CERand whose authors identify upstream encoder quality as the dominant term \(Section[2](https://arxiv.org/html/2609.17548#S2)\)\.

## 8Limitations

- •One subject\.Every number comes from a single healthy speaker\. The encoder memorizes the 8,500 training sentences \(trainingPERis far below the∼\\sim27% validationPER\), so cross\-subject robustness is untested\.
- •Not the best sEMG\-to\-text system\.18\.53% is the best result on the*emg2speech*General Corpus, but MONA LISA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]reports 3\.7% on vocalized EMG on the Gaddy corpus\. The two are not measured on the same recordings \(Section[2](https://arxiv.org/html/2609.17548#S2)\), and I have not run Myovox on the Gaddy corpus, so I cannot say whether the gap is the corpus or the model\.
- •Validation is harder than test\.Both systems score better on test than on validation, which is a roughly nine\-point property of this fixedsequential splitrather than evidence of test\-set overfitting\. It also means the headline number is measured on the easier segment of the corpus\. I report both columns rather than the flattering one\.
- •The audio is a training\-time crutch\.WavLMdistillationuses the parallel audio, which does not exist at silent\-speech inference\. At the word level the electromyography\-only encoder matches it \(Section[5](https://arxiv.org/html/2609.17548#S5)\), which is reassuring for the eventual application\.
- •The ensemble’s two contributions are not separated\.Averaging the two encoders improvesPERby 1\.4 points andWERby 2\.7\. How much of the word\-level gain follows from the sharpened argmax and how much from the reshaped posterior is not measured here \([11](https://arxiv.org/html/2609.17548#S11)\)\. The question does not bear on Section[7](https://arxiv.org/html/2609.17548#S7), which holds under either answer, but the report should not be read as having settled it\.
- •Test duplicates\.Six of the 400 test sentences duplicate training text, so the score is reported with and without them \(18\.53 against 18\.75\)\.
- •Reranking saturates\.The headline 18\.53% is bounded by acoustic phone error rate, not by the language model\. This is a negative result about language\-model reranking for electromyography, not a claim that reranking is useless: it is worth−4\.7\-4\.7WER\.
- •Vocalized, not silent\.The subject spoke aloud, so this validates the pipeline on the easier case\. It is a stepping stone towardsilent speech, not a demonstration of it\.

## 9The project website

Myovox is documented in two ways: this report, which is terse and aimed at readers who already know the field, and a website, which is aimed at a curious reader who is new to speech decoding and to electromyography\. The site is a cinematic, scroll\-driven explainer of surface\-EMG speech decoding, showing how the electrical signals of the facial muscles are turned into text at an 18\.53% word error rate\. It walks each stage of the pipeline with animation rather than equations, which is why this report carries no architecture diagrams: the moving explanation lives at[varshith\-0\.github\.io/myovox](https://varshith-0.github.io/myovox/)\. The project is open\-sourced under the MIT license, and the site is paired with this technical report and a reproducible pipeline so that every claim on it is auditable\.

## 10Glossary

Every term below is a jump target: clicking any occurrence of the term in the body of the report brings the reader here\. Definitions are given as the term is used in this report, not in full generality\.

Metrics

WER\(word error rate\)The fraction of reference words that are wrong after the best alignment of hypothesis to reference, counting substitutions, insertions, and deletions:\(S\+I\+D\)/N\(S\+I\+D\)/N\. It is the reader\-facing number, and it depends on the acoustic model, the decoder, and the language model together\. Lower is better; 18\.53% is the headline result here\.

PER\(phone error rate\)The same edit\-distance quantity computed overphonemesrather than words, measured bygreedy decodingof the CTC posteriors with nolexiconand nolanguage model\. Because nothing downstream of the encoder is involved,PERis*decoder\-independent*: it isolates the quality of theacoustic model\. In this report it serves as both the credibility check \(Section[4](https://arxiv.org/html/2609.17548#S4)\) and the binding constraint \(Section[7](https://arxiv.org/html/2609.17548#S7)\)\.

CER\(character error rate\)The same edit distance computed over characters\. Not used for any Myovox result, sincePERis the sharper instrument for an acoustic model that emits phonemes\. It appears only in Section[2](https://arxiv.org/html/2609.17548#S2), because the non\-invasive brain\-to\-text literature reports it: those systems decode keystrokes, for which the character is the natural unit\. The usual ladder of granularity runs character, phone, word\.

Oracle WER\(nn\-best oracle\)TheWERthat would be obtained if a perfect judge always selected the single best candidate from thenn\-best pool\. It is a lower bound on what anyrerankerover that pool can achieve, and therefore a measure of how much information the pool contains\. Here the union pool has a 9\.30% oracle against an 18\.53% achievedWER, and the remaining nine points are words the acoustics never proposed at all\.

Acoustic posteriorThe per\-frame probability distribution over output symbols \(phonemes, units, or blank\) emitted by the encoder\. If the correct word’s phonemes never receive appreciable probability mass, no decoder or reranker can recover it\. That is the failure mode diagnosed in Section[7](https://arxiv.org/html/2609.17548#S7)\.

Models and architectures

Acoustic modelThe neural network that maps the input signal, heresEMGfeatures, to per\-frame symbol posteriors\. It is the only component whose qualityPERmeasures\.

Language modelA model of which word sequences are plausible in English, used either inside the decoding graph \(theGGofHLG\) or after decoding, as areranker\. It supplies word\-level priors, not phonetic evidence\.

TDS\(time\-depth separable convolution\)The convolutional encoder used in the released baseline\. It factorizes a convolution into a time\-only and a channel\-only \(depthwise and pointwise\) part, giving cheap large receptive fields\. The released variant iscausal\.

ConformerAn encoder block that interleavesmulti\-head self\-attentionwithdepthwise convolution, so it can model both long\-range and local structure\. ThebidirectionalConformer replaces the causal TDS encoder in Section[5](https://arxiv.org/html/2609.17548#S5), and is responsible for most of theWERgain\.

Multi\-head self\-attentionThe mechanism by which every frame in a sentence attends to every other frame, with several independent “heads” attending to different things\. It is what gives the Conformer full\-sentence context\.

Depthwise convolutionA convolution applied independently per channel\. It is cheap in parameters, and is used inside the Conformer block to capture local, short\-time structure\.

Causal encoderAn encoder that at each frame sees only past and present frames, never the future \(implemented here by left\-padding\)\. Required forstreaming, but a handicap for offline transcription\.

Bidirectional \(full\-context\) encoderAn encoder that may look at the entire utterance, future frames included\. Legitimate whenever the whole sentence is recorded before decoding begins, as it is here\.

StreamingProducing output as the signal arrives, in real time, without waiting for the sentence to finish\. Not required by this report, which decodes offline\.

HuBERTA self\-supervised speech model whose discretized representations give the 100unitsthat the first CTC head predicts\.

HuBERT unitOne of 100 discrete, self\-supervised speech tokens obtained by clustering HuBERT features\. Units are a finer\-grained, learned alternative to phonemes, and they carry the dominant \(0\.8\-weighted\) CTC loss\.

WavLM\(WavLM\-Large, layer 9\)A large self\-supervised speech representation model\. Its layer\-9 features, extracted from the parallel audio, are the teacher signal for thecross\-modal distillation\. Layer 9 is used because the mid\-stack layers of such models carry the most phonetic information\.

Qwen2\.5\-7B\-InstructThe 7\-billion\-parameter open\-weight instruction\-tuned language model used as theLIFTreranker\.

LoRA\(low\-rank adaptation\)A fine\-tuning method that freezes the base model and trains only a small pair of low\-rank matrices injected into each weight \(here, rank 16\)\. It makes adapting a 7B model cheap, and makes its changes easy to isolate\.

QLoRALoRA applied on top of a base model quantized to 4 bits \(NF4\), so that a 7B model can be fine\-tuned on a single consumer GPU\. Used for the reranker\.

BiLSTMA bidirectional long short\-term memory recurrent network\. Used here as the audio teacher for the frame\-level KL term of the second ensemble member\.

Frozen modelA model whose weights are held fixed while it participates in another model’s training\. The small WavLM\-to\-phoneme recognizer is frozen so that it acts as a fixed judge of whether the projected features are phoneme\-decodable\.

Speech and electromyography

Electromyography \(EMG\)The measurement of the electrical activity produced by muscle fibres when they contract\.

sEMG\(surface electromyography\)EMG measured non\-invasively, by electrodes placed on the skin rather than by needles inserted into muscle\. Here: a 31\-channel array on the face, sampled at 5 kHz\.

Non\-invasive BCI\(EEG, MEG\)The other non\-invasive route to text, tapping cortex rather than muscle\. EEG measures electrical fields at the scalp and is cheap and portable; MEG measures the magnetic fields of neuronal activity and is far cleaner but requires a room\-sized, cryogenically cooled instrument\. Both are far upstream of articulation, and both currently trail sEMG on error rate \(Section[2](https://arxiv.org/html/2609.17548#S2)\)\. Not used in this report; included only for comparison\.

Vocalized speechSpeech produced aloud, with the vocal folds active and audible sound emitted\. All Myovox recordings are vocalized, which is why the parallel audio exists at all\.

Silent speechArticulating without producing sound, and in the strictest case without moving air, so that only muscle activity is available\. The eventual target application, and*not*what this report evaluates\.

Phoneme \(phone\)The smallest sound unit that distinguishes words in a language\. The model predicts a 40\-phoneme inventory \(41 classes once blank is included\), andPERis measured over it\.

Shrinkage covariance, thevec​\(E\)\\mathrm\{vec\}\(E\)front\-endThe input representation\. Rather than raw waveforms, each analysis window is summarized by the regularized \(“shrunk”\) covariance matrix across the 31 channels, vectorized\. It captures how the channels co\-activate, which is the form in which the geometry of facial articulation shows up\.

Frame, window, hopThe signal is chopped into 25 ms*windows*advanced by a 20 ms*hop*, yielding one feature*frame*every 20 ms, that is, 50 frames per second\.

Training

CTC\(connectionist temporal classification\)A loss that trains a frame\-level classifier against an unaligned target sequence by summing over all alignments, using a specialblanksymbol to mean “no output here\.” It removes the need for frame\-by\-frame labels, which do not exist for EMG\.

Dual\-CTCThe baseline objective: two CTC heads on one encoder, one predicting HuBERTunitsand one predicting phonemes, tied together by aconsistencyterm\.

Consistency tableA fixed conditional tableP​\(phone∣unit\)P\(\\text\{phone\}\\mid\\text\{unit\}\)used to penalize disagreement between the two CTC heads, so that the unit head’s richer signal informs the phone head\.

DistillationTraining one model \(the student\) to imitate the internal representations or outputs of another \(the teacher\), rather than only to fit the labels\.

Cross\-modal distillationDistillation in which teacher and student read*different*signals of the same event: here the teacher reads audio \(WavLM\) and the student reads muscle activity \(sEMG\)\. The teacher exists only at training time; at inference the student runs on EMG alone\. The idea is adapted from MONA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]\.

InfoNCEA contrastive loss that pulls a representation toward its matching partner, the same moment in the parallel audio, and pushes it away from all the non\-matching ones in the batch, sharpened by a temperatureτ\\tau\(0\.1 here\)\.

Warm startInitializing part of a new model from the weights of a previously trained one, rather than from random\. The covariance front\-end, the CTC heads, and the audio projection all warm\-start from the baseline checkpoint\.

Checkpoint selectionChoosing which training snapshot to keep, by a criterion evaluated on validation\. Selecting by validationPERrather than by validation CTC loss is worth roughly 4PERpoints here\.

Fine\-tuningContinuing to train an already\-trained model on a new, usually narrower task\. Here: adapting Qwen2\.5 to the reranking task with QLoRA\.

Paired bootstrapA significance test\. Resample the test utterances with replacement many times, recompute the*difference*between two systems on the same resampled set each time, and read off a confidence interval\. If the interval excludes zero, the improvement is unlikely to be sampling noise\.

Cross\-decoding \(two\-fold\)Generating the reranker’s training candidates by splitting the training data in half and decoding each half with a model trained only on the other half\. Without it, the candidates would reflect the model’s memorized∼\\sim1% trainingWER, and the reranker would learn to correct errors it will never see at test time\.

LeakageAny path by which information about the evaluation data reaches the model during training, inflating the score\. On a single\-subject corpus the main risk is a language model that has memorized training sentences which recur at test\.

Verbatim\-recall auditThe control that counts how often the free reranker emits a reference sentence that was*not*present anywhere in its candidate list, that is, produced from memory rather than from evidence\. The count here is zero\.

Decoding

Greedy decodingTaking the highest\-probability symbol at each frame independently, then collapsing repeats and blanks\. No lexicon, no language model, no search\. This is howPERis measured, which is whyPERreflects the acoustic model and nothing else\.

WFST\(weighted finite\-state transducer\)A finite\-state machine whose transitions carry input symbols, output symbols, and weights\. Composing several of them gives a single graph that maps frame\-level posteriors to word sequences while accumulating scores\.

HLGThe composed decoding graphH∘L∘GH\\circ L\\circ G\.HHmaps CTC frames to phonemes,LLis thelexiconmapping phoneme strings to words, andGGis thelanguage modelover word sequences\. Decoding is a search through this graph\.

LexiconThe pronunciation dictionary: for each word, its phoneme sequence or sequences\. It defines exactly which words the decoder is capable of producing\.

k2 and icefallThe open\-source finite\-state and speech\-recognition toolkits used to construct and search the HLG graph, matching the original authors’ setup\.

LibriSpeechA large public read\-speech corpus\. Its pronunciation lexicon, 34,546 words here, is what makes the decodeopen\-vocabulary\.

Open vocabularyDecoding over a lexicon far larger than the corpus vocabulary \(roughly five times larger here\), so that the system can output words it never saw in training\. The alternative would be a closed\-set classifier over a fixed sentence or word list\. Word error rates are not comparable across this boundary: a closed\-vocabulary system with a small phrase list can post a much lowerWERwhile solving a much easier problem\.

OOV\(out of vocabulary\)A reference word absent from the lexicon, and hence impossible to emit\. Twelve of the 2,429 test tokens are OOV, a hard 0\.49%WERfloor\.

Acoustic scaleThe weight given to the acoustic scores relative to the language\-model scores during graph search\. Set it too low and the language model overwhelms the evidence; set it too high and the language model stops helping\. Tuned on validation: 1\.0 for the baseline, which is also the released default, so tuning changes nothing there; 0\.25 for the Conformer, where it matters\.

Blank symbol, and blank penaltyCTC’s “emit nothing here” symbol, which dominates the posteriors \(peak≈0\.92\\approx 0\.92\)\. The blank*penalty*subtracts a constant from its score at decode time so that real symbols can compete\. It is the single largest lever in the baseline decode\.

LatticeThe compact graph of high\-scoring alternative paths retained during decoding, from whichnn\-best lists are extracted\.

nn\-best listThe topnncomplete hypotheses from the lattice, ranked by score\. Extractingnn\-best lists at several acoustic scales and taking their*union*widens the pool a reranker can choose from, lowering theoracleWERwithout changing the acoustic model\.

EnsembleAveraging the per\-frame phone log\-probabilities of two independently trained encoders before decoding\. Worth 2\.7WERpoints here at best \(26\.14 to 23\.47\), alongside a 1\.4\-pointPERimprovement \(22\.34 to 20\.90\)\. How much of the word\-level gain follows from the sharpened argmax and how much from the reshaped posterior is not separated in this report; see[11](https://arxiv.org/html/2609.17548#S11)\.

RerankingRescoring or rewriting thenn\-best candidates with a stronger model, usually a language model, so as to pick a better one than the acoustic score alone would\. Bounded from below by the oracleWER\.

LIFT\(DCoND\-LIFT style\)The reranking recipe adopted here: prompt a fine\-tuned LLM with the candidate list*and*the detected phoneme sequence, and have it produce the final transcript\. The*constrained*variant must return one of the candidates; the*free*variant may write its own correction\.

Data

General CorpusThe single\-subject, healthy\-speaker portion of the*emg2speech*release, used throughout: 9,660 sentences of 31\-channel sEMG with simultaneously recorded audio\.

Gaddy corpusThe reference open\-vocabulary sEMG corpus of Gaddy and Klein\[Gaddy and Klein,[2020](https://arxiv.org/html/2609.17548#bib.bib16)\]: eight facial channels at 1 kHz, roughly nineteen hours from a single English speaker, recorded in both silent and vocalized modes\. It is the benchmark on which MONA LISA and most other sEMG\-to\-text systems report, and it is*not*the corpus used here\.

Sequential split \(8,500 / 760 / 400\)The authors’ own division of the corpus, taken in recording order rather than at random, and kept unchanged here\. Because it is sequential, validation and test are not statistically interchangeable: validation is systematically harder, by roughly nine points\.

Validation setThe 760 sentences on which every hyperparameter is tuned: blank penalty, acoustic scale, checkpoint, and reranker variant\. Nothing is ever tuned on test\.

Test setThe 400 held\-out sentences on which every reported test number is computed, once, with the settings already fixed on validation\. Six of them duplicate training text, so results are reported both with and without them\.

## 11Frequently asked questions

These are the questions the project actually drew, in roughly the order they were asked\. I have tried to answer the sharp form of each rather than the polite one, and where the honest answer is “I do not know,” it says so\.

Reproduction and the headline result

Q1You report a 32\-point improvement over the published number\. Did you improve their model, or only their decoder?

Both, in separable amounts, and Section[4](https://arxiv.org/html/2609.17548#S4)exists to keep them separable\. The first 10\.5 points are a decoder fix and nothing else\. The evidence is the phone error rate: my baseline lands at 39\.02% against the published 38\.19%, a gap of 0\.8 points \(Table[4](https://arxiv.org/html/2609.17548#S4.T4)\)\. BecausePERis measured bygreedy decodingwith nolexiconand nolanguage model, a matchedPERestablishes that theacoustic modelis the same acoustic model, so the entireWERgain must come from supplying the decode settings the public release omits, of which theblank penaltyis much the largest\. The remaining 22 points split into a genuinely better acoustic model \(Section[5](https://arxiv.org/html/2609.17548#S5),PER39\.02→22\.3439\.02\\rightarrow 22\.34, worth 14\.5WERpoints\) and a decode\-time stack worth 7\.6WERpoints that takesPERonly 1\.4 further, from 22\.34 to 20\.90, and whose largest single component, the reranker, cannot movePERat all \(Section[6](https://arxiv.org/html/2609.17548#S6)\)\. So: roughly a third of the total from decoding, a little under half from modeling, and the balance from extraction\. I would rather publish that breakdown than a single undifferentiated delta\. One caveat worth stating plainly: I never reproduce 51\.17% itself, because the released configuration decodes at roughly 75%\. The claim to have reproduced their*model*rests on the matched phone error rate, not on a matched word error rate\.

Q2Is 18\.53%WERgood? What does a sentence at that error rate look like?

Approximately one word in five is wrong, whether substituted, inserted, or dropped\. The errors are not distributed evenly: most sentences read cleanly and a minority are badly mangled\. For calibration, a modern recognizer on clean audio sits near 5%, and the published starting point for this task was 51\.17%\. The practical reading is that 18\.53% is useful to a human who can tolerate correcting it, and not yet trustworthy unattended\.

Q3MONA LISA reported 3\.7%WERon vocalized EMG in 2024\. Why is your number five times worse?

Because it is a different corpus, and I do not know how much of the gap that explains\. MONA LISA\[Bensteret al\.,[2024](https://arxiv.org/html/2609.17548#bib.bib19)\]evaluates on theGaddy corpus: eight channels, about nineteen hours, one speaker, recorded in both silent and vocalized modes\. Myovox evaluates on the*emg2speech*General Corpus: 31 channels, 9,660 sentences, a different speaker, one mode\. MONA additionally trains against audio\-only LibriSpeech, which I do not\. So the honest claim is the narrow one: 18\.53% is the best number on*this*corpus, against a published 51\.17%, and Myovox is not the state of the art for sEMG\-to\-text in general\. There is a real scientific question hiding in the gap, and it is the sharpest one the project raises\. MONA reaches 3\.7% with a cross\-modal audio objective; my control shows my cross\-modal audio objective buying nothing at the word level \(26\.10 against 26\.14\)\. Either their contrastive formulation carries phonetic information that mine does not, or thecovariance front\-enddiscards what their raw\-signal front\-end keeps, or this corpus is simply harder\. Running Myovox on the Gaddy corpus would separate those, and I have not done it\.

Q4Your test score is*better*than your validation score\. That usually indicates a problem\.

It usually does, so it deserves a direct answer rather than a footnote\. Thesplitis*sequential*: the corpus is cut in recording order, not at random, so validation and test are not exchangeable samples of one distribution\. Test is the easier segment, by roughly nine points, and the gap is stable across every system I trained, including the baseline, on which no tuning was performed\. Overfitting to test would have*widened*the gap as I tuned; it did not\. The protection here is procedural rather than statistical: every hyperparameter is selected on the 760 validation sentences and applied once to the 400 test sentences\. Both columns are reported throughout, rather than only the flattering one\. The corollary, which I would rather state than have pointed out to me, is that 18\.53% is the error rate on the easier half of this corpus, not on a random sentence drawn from it\.

Scope: vocalized versus silent speech

Q5Does this read thoughts?

No\. It reads muscles\. Myovox decodes the electrical activity of facial muscles as they contract to articulate, which is the terminal link of the motor chain, long after intention has become movement\. Absent articulation, there is nothing at the electrodes to decode\. The systems that do read cortex, non\-invasively, are theEEG and MEGdecoders of Section[2](https://arxiv.org/html/2609.17548#S2), and they are currently far behind\.

Q6The subject spoke aloud\. Doesn’t the presence of parallel audio make the result vacuous?

It makes the claim weaker than a silent\-speech demonstration would be, which is why the limitation is stated plainly rather than buried\. It does not make it vacuous\. The audio exists in the*recording*; it does not exist in the*model’s input*\. Every reported number, 40\.63, 26\.14, and 18\.53, is produced by a system whose inference\-time input is 31 channels ofsEMGand nothing else\. Audio enters only as a training\-time teacher \(Section[5](https://arxiv.org/html/2609.17548#S5)\), and[11](https://arxiv.org/html/2609.17548#S11)below shows that even this proves dispensable at the word level\. Whatvocalizationbuys is a stronger, more consistent signal than silent articulation produces, which is precisely why this is a stepping stone rather than the destination\.

Q7Would this work for someone who cannot speak, an ALS patient, say?

I do not know, and I would rather say so than gesture at the application and let the reader supply the optimism\. The model is trained on one healthy speaker articulating normally\. A patient with ALS presents different muscle recruitment, different signal amplitude, frequently altered anatomy, and no parallel audio from which to distil\. Nothing in this report constitutes evidence that a model trained under my conditions transfers to theirs\. Closing that gap requires data from the affected population, not extrapolation from mine\.

Q8Audio and EMG were recorded simultaneously\. How do you know the model is not exploiting acoustic leakage into the electrode array?

The suspicion is well placed\. The structural answer is that the input is ashrinkage covariance across the 31 channels, a representation of how muscle groups co\-activate over a 25 ms window rather than a waveform, so there is no spectral path by which speech could survive into the model’s input in decodable form\. For leakage to explain the result, it would have to carry enough phonetic detail to sustain a 22\.34%PERthrough a skin\-surface array band\-limited to muscle frequencies, and then survive the covariance summary\. I did not run a dedicated acoustic\-shielding ablation, so I offer this as a design argument, not as a measured control\.

Modeling choices

Q9Your own control shows the EMG\-only encoder matching the distilled one onWER\. Was the WavLM distillation therefore wasted?

At the word level, effectively yes: 26\.10 against 26\.14 is a coin flip, and I chose to print that control in the same table as the result rather than relegate it to a footnote\. Two things survive\. The distillation does improvePER\(22\.34 against 23\.71\), so the encoder is phonetically better even though the decoder cannot convert the improvement into words, which is itself a preview of the ceiling diagnosed in Section[7](https://arxiv.org/html/2609.17548#S7)\. More importantly, the negative result is load\-bearing: it demonstrates that the parallel audio is not a hidden crutch supporting the headline number, which is exactly what one needs to know before extending any of this to silent speech, where audio will not exist\. A control that invalidates a favored component is worth more than the component was\.

Q10Why retain the frozen WavLM\-to\-phoneme recognizer? AreL2L\_\{2\}andInfoNCEnot sufficient to match the teacher’s features?

L2L\_\{2\}and InfoNCE make the projection*close*to WavLM\. Close is not the same as useful\. A projection can be smooth, correctly scaled, and well correlated frame\-by\-frame with the teacher, and still be phonetically empty: regression toward a high\-dimensional target will happily settle on a blurred average that satisfies the loss while discarding the distinctions that separate one phoneme from the next\. Thefrozenrecognizer changes what “close” is permitted to mean\. It is a small network that already reads phonemes out of real audio, and its weights never move, so pushing the EMG projection through it and demanding a low CTC loss forces the projection to be*decodable into the correct phonemes*rather than merely audio\-shaped\. It is a guard against smooth\-but\-empty blur\.

The language\-model reranker

Q11A 7B language model on a single\-subject corpus invites memorization\. How do you exclude it?

By assuming it would occur, and building three controls before trusting the number\. First, the reranker sees only the training split, and its training candidates are produced by two\-foldcross\-decoding, each half decoded by a model that never trained on it, so it learns to repair realistic errors rather than the memorized∼\\sim1%WERthe acoustic model achieves on its own training data\. Second, six of the 400 test sentences duplicate training text exactly, so the score is reported both with and without them \(18\.53 against 18\.75\) and the reader may take whichever they find credible\. Third, theverbatim\-recall auditcounts how often free generation emits a reference sentence*absent*from the candidate list, that is, recalled from memory rather than selected from evidence\. The count is zero\.

Q12The free reranker may write anything it likes\. Is 18\.53% simply fluent hallucination?

That is the failure mode the verbatim\-recall audit was constructed to detect, and it returns zero: the free variant never produces a correct reference that the acoustics did not propose\. Its gains come from repairing candidates that were close but wrong, not from inventing candidates that were absent\. The structural argument is the one that also bounds the system \(Section[7](https://arxiv.org/html/2609.17548#S7)\)\. A model hallucinating its way to the answer would be closing the nine\-point gap to theoracle\. It is not\. It halts precisely where the acoustic evidence halts, which is what a well\-behaved reranker does and what a hallucinating one would not\.

The acoustic ceiling

Q13Distillation bought 1\.4 points ofPERand noWERat all\. Theensemblebought 1\.4 points ofPERand 2\.7WER\. Why the difference?

BecausePERandWERinterrogate different objects, and the ensemble moves both\.PERis agreedyargmax over frames: it asks only whether the top symbol is correct\. TheWFSTdecoder never takes the top symbol; it searches over*paths*, weighing whole sequences against thelexiconand thelanguage model\. Averaging two encoders’ log\-probabilities sharpens the argmax, which is the 1\.4 points ofPER, but it also reshapes the tails of the distribution, and the tails are where the search actually operates\. Paths on which the two models disagree get suppressed; correct paths that were never top\-ranked survive to be found\.

I did not separate the two contributions, and this report should not claim to have\. The distillation control offers a rough calibration: there, 1\.37 points ofPER\(23\.71 to 22\.34\) bought−0\.04\-0\.04WER\. If that exchange rate held at 21–22%PER, the ensemble’s 1\.44 points would likewise buy nothing and the 2\.67WERpoints would be the reshaped posterior alone\. But that assumes the marginalWERvalue of aPERpoint is constant across the range, which it probably is not: a word needs all of its phonemes, so the curve is likely convex and the lower\-PERpoints worth more\. Some of the 2\.67 may be phonetic after all\.

Either answer leaves Section[7](https://arxiv.org/html/2609.17548#S7)standing, which is why the question can be left open here\. If the gain is phonetic, it is direct evidence thatPERis the quantity that binds\. If it is diversity, the ensemble redistributes information rather than adding any\. It cannot rescue the system either way: 2\.7WERpoints against the 14\.49 that full context bought, and a phone error rate still at 20\.9%\.

Q14Is 20\.9%PERa ceiling for EMG, or only for*your*EMG model?

Only for mine\. A single\-subject study cannot support the stronger claim\. What the report does establish is narrower and, I think, more useful: for this acoustic model on this corpus, 20\.9%PERis the binding constraint, and no amount of decoder or language\-model sophistication passes it\. Section[7](https://arxiv.org/html/2609.17548#S7)gives three independent lines of evidence that the ceiling is acoustic\.PERmoves only in small, expensive increments under interventions that moveWER; the audio teacher cannot transfer its own phonetic quality; and the nine\-point oracle gap consists of words the posteriors never proposed\. Whether the ceiling is a property of surface EMG itself, of the covariance front\-end, or of possessing 8,500 sentences from one person, this study cannot distinguish\. That is the next experiment\. MONA LISA’s 3\.7% on a different corpus \([11](https://arxiv.org/html/2609.17548#S11)\) is direct evidence that it is not a property of surface EMG itself\.

Q15Is there any independent support for the claim that the reranker, not the encoder, is the exhausted component?

Yes, and it arrives from a different signal entirely, which is the best kind of corroboration\. Brain2Qwerty v2\[Zhanget al\.,[2026](https://arxiv.org/html/2609.17548#bib.bib22)\]decodes typed sentences from non\-invasive MEG with aConformerCTCencoder and aLoRA\-fine\-tuned LLM above it, which is structurally the same stack as Sections[5](https://arxiv.org/html/2609.17548#S5)and[6](https://arxiv.org/html/2609.17548#S6)\. Their encoder alone reaches 55%WER; annn\-gram language model takes it to 43%; the fine\-tuned 4B LLM takes it to 39%\. The LLM is worth four points over thenn\-gram, and it buys them while*worsening*the character error rate: it trades local accuracy for fluency\. The analogy is in the conclusion rather than the mechanism, and the difference matters\. Their LLM rewrites the sentence autoregressively and reads MEG\-derived word embeddings directly, so it can degradeCER; my reranker operates on words downstream of the greedy CTC output and cannot movePERat all, which is whyPERhere stays pinned at 20\.9% rather than worsening\. Their own text\-only ablation, which leaves the LLM nothing but the encoder’s characters, is the closer analogue to what I built, and it is worth about sixWERpoints\. Their conclusion is nonetheless mine: final performance is dominated by upstream encoder quality, and improving the encoder is the priority\. Two systems, two signals, one wall\.

Q16What is the single highest\-value next step toward sub\-10%WER?

Not the decoder\. Everything downstream of the encoder is exhausted, which is the finding, and the reason Section[7](https://arxiv.org/html/2609.17548#S7)was written\. The gain has to come from the acoustic model: more data, multiple subjects, or a front\-end that recovers phonemes from the raw 31\-channel signal rather than from a 25 ms covariance summary\. If forced to choose one, I would choose additional speakers, because it attacks the ceiling and the generalization limitation at the same time, and because a single\-subject corpus is not a constraint one can model around\.

Reproducibility and deployment

Q17What did this cost to train, and can it be reproduced?

The code, the configurations, and, critically, the decode settings are open\-sourced under the MIT license at[github\.com/Varshith\-0/myovox](https://github.com/Varshith-0/myovox)\. That includes theblank penalty, thecheckpoint\-selection criterion, the regeneratedwords\.txt, and theacoustic scale: the settings whose combined absence is why the released configuration decodes at roughly 75%WERrather than the 40\.63% this baseline reaches\. The blank penalty is much the largest of the four; I did not separate them further\. The data is the*emg2speech*General Corpus, and must be obtained from its authors\. Training ran on a single NVIDIA RTX 3080 Ti with 12 GB of VRAM, and took roughly 15 to 25 hours end to end depending on the experiment \(baseline reproduction, Conformer training, or reranker fine\-tuning\)\. TheQLoRAconfiguration was chosen precisely so that the 7B reranker fits on that same 12 GB consumer card\.

Q18Does it run in real time?

No, and this is a deliberate trade rather than an oversight\. The central architectural decision of Section[5](https://arxiv.org/html/2609.17548#S5), replacing thecausalencoder with abidirectionalConformer, lets the model attend to the entire utterance, future frames included, so no word can be emitted until the sentence ends\. Theensemble, the multi\-scalenn\-best union, and the 7Brerankerabove it are all offline by construction\. Astreamingvariant is certainly buildable, and the released causal TDS encoder is one, but it would surrender most of the 14\.49WERpoints that full context purchased\. The objective of this report was to locate the ceiling, not to ship an interface\.

## Acknowledgments

Throughout the project I made heavy use of Anthropic’s Claude: for working through the decode\-time failures, structuring the distillation objective and the leakage controls, sifting the literature, and drafting this report\. The design decisions, the experiments, and any errors that remain are mine\.

## References

- T\. Benster, G\. Wilson, R\. Elisha, F\. R\. Willett, and S\. Druckmann \(2024\)A cross\-modal approach to silent speech with LLM\-enhanced recognition\.arXiv preprint arXiv:2403\.05583\.Note:MONA LISAExternal Links:2403\.05583Cited by:[https://arxiv.org/html/2609.17548#S10.I4.ix5.p1.1](https://arxiv.org/html/2609.17548#S10.I4.ix5.p1.1)item Cross\-modal distillation,[§11](https://arxiv.org/html/2609.17548#S11.p8.1),[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px1.p1.1.1.1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.13.3.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.14.4.1),[§5](https://arxiv.org/html/2609.17548#S5.SS0.SSS0.Px2.p1.4.1.1),[§6](https://arxiv.org/html/2609.17548#S6.SS0.SSS0.Px3.p1.1.1.1.1.1.2),[2nd item](https://arxiv.org/html/2609.17548#S8.I1.i2.p1.1)\.
- S\. Chen, C\. Wang, Z\. Chen, Y\. Wu, S\. Liu, Z\. Chen, J\. Li, N\. Kanda, T\. Yoshioka, X\. Xiao, J\. Wu, L\. Zhou, S\. Ren, Y\. Qian, Y\. Qian, M\. Zeng, X\. Yu, and F\. Wei \(2022\)WavLM: large\-scale self\-supervised pre\-training for full stack speech processing\.IEEE Journal of Selected Topics in Signal Processing16\(6\),pp\. 1505–1518\.Note:arXiv:2110\.13900Cited by:[§5](https://arxiv.org/html/2609.17548#S5.SS0.SSS0.Px2.p1.4.1)\.
- T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. Zettlemoyer \(2023\)QLoRA: efficient finetuning of quantized llms\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2305\.14314Cited by:[§6](https://arxiv.org/html/2609.17548#S6.SS0.SSS0.Px3.p1.1.1.1.1)\.
- D\. Gaddy and D\. Klein \(2020\)Digital voicing of silent speech\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 5521–5530\.Note:Introduces the 8\-channel, single\-speaker silent/vocalized facial EMG corpusExternal Links:2010\.02960Cited by:[https://arxiv.org/html/2609.17548#S10.I6.ix2.p1.1](https://arxiv.org/html/2609.17548#S10.I6.ix2.p1.1)item Gaddy corpus,[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px1.p1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.5.5.2),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.6.6.1)\.
- D\. Gaddy and D\. Klein \(2021\)An improved model for voicing silent speech\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 2: Short Papers\),External Links:2106\.01933Cited by:[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px1.p1.1.1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.7.7.1)\.
- D\. Gaddy \(2022\)Voicing silent speech\.Ph\.D\. Thesis,University of California, Berkeley\.Cited by:[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px1.p1.1.1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.11.1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.12.2.1)\.
- H\. T\. Gowda, D\. C\. Comstock, and L\. M\. Miller \(2026\)Emg2speech: synthesizing speech from electromyography using self\-supervised speech models\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(ACL\),Note:arXiv:2510\.23969\. Baseline reproduced here \(Appendix D\.4, “emg2text”\)\.Cited by:[Table 1](https://arxiv.org/html/2609.17548#S1.T1.4.7.3.2),[§1](https://arxiv.org/html/2609.17548#S1.p3.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.15.5.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.8.8.1),[Table 4](https://arxiv.org/html/2609.17548#S4.T4.5.3.1.1)\.
- H\. T\. Gowda, Z\. D\. McNaughton, and L\. M\. Miller \(2024\)Geometry of orofacial neuromuscular signals: speech articulation decoding using surface electromyography\.Journal of Neural Engineering\.Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4)\.
- H\. T\. Gowda and L\. M\. Miller \(2025\)Non\-invasive electromyographic speech neuroprosthesis: a geometric perspective\.arXiv preprint arXiv:2502\.05762\.Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4)\.
- A\. Graves, S\. Fernández, F\. Gomez, and J\. Schmidhuber \(2006\)Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks\.InProceedings of the 23rd International Conference on Machine Learning \(ICML\),pp\. 369–376\.Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4.3.3.3.3)\.
- A\. Gulati, J\. Qin, C\. Chiu, N\. Parmar, Y\. Zhang, J\. Yu, W\. Han, S\. Wang, Z\. Zhang, Y\. Wu, and R\. Pang \(2020\)Conformer: convolution\-augmented transformer for speech recognition\.InInterspeech,Note:arXiv:2005\.08100Cited by:[§5](https://arxiv.org/html/2609.17548#S5.SS0.SSS0.Px1.p1.1.1.1.1.1.1)\.
- A\. Hannun, A\. Lee, Q\. Xu, and R\. Collobert \(2019\)Sequence\-to\-sequence speech recognition with time\-depth separable convolutions\.InInterspeech,Note:arXiv:1904\.02619Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4.3.3.3)\.
- W\. Hsu, B\. Bolte, Y\. H\. Tsai, K\. Lakhotia, R\. Salakhutdinov, and A\. Mohamed \(2021\)HuBERT: self\-supervised speech representation learning by masked prediction of hidden units\.IEEE/ACM Transactions on Audio, Speech, and Language Processing29,pp\. 3451–3460\.Note:arXiv:2106\.07447Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4.3.3.3.3.3.3)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2106\.09685Cited by:[§6](https://arxiv.org/html/2609.17548#S6.SS0.SSS0.Px3.p1.1.1.1.1.1)\.
- J\. Lévy, M\. Zhang, S\. Pinet, J\. Rapin, H\. Banville, S\. d’Ascoli, and J\. King \(2025\)Brain\-to\-text decoding: a non\-invasive approach via typing\.Nature Neuroscience\.Note:Brain2Qwerty v1\. arXiv:2502\.17480External Links:2502\.17480Cited by:[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px3.p1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.19.9.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.20.10.1)\.
- J\. Li, T\. Le, C\. Fan, M\. Chen, and E\. Shlizerman \(2024\)Brain\-to\-text decoding with context\-aware neural representations and large language models\.arXiv preprint arXiv:2411\.10657\.Note:DCoND\-LIFT\. Journal of Neural Engineering, 2025\.Cited by:[§6](https://arxiv.org/html/2609.17548#S6.SS0.SSS0.Px3.p1.1.1.1.1.1.2)\.
- P\. Mohapatra, A\. Pandey, X\. Zhang, and Q\. Zhu \(2025\)Can LLMs understand unvoiced speech? exploring EMG\-to\-text conversion with LLMs\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),Vienna, Austria,pp\. 703–712\.External Links:2506\.00304Cited by:[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px1.p1.1.1.1.1.2.1)\.
- V\. Panayotov, G\. Chen, D\. Povey, and S\. Khudanpur \(2015\)Librispeech: an asr corpus based on public domain audio books\.InIEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 5206–5210\.Cited by:[§3](https://arxiv.org/html/2609.17548#S3.SS0.SSS0.Px3.p1.1.1.1)\.
- D\. Poveyet al\.\(2023\)K2 and icefall: fsa/fst algorithms and asr recipes\.Note:[https://github\.com/k2\-fsa/k2](https://github.com/k2-fsa/k2),[https://github\.com/k2\-fsa/icefall](https://github.com/k2-fsa/icefall)Cited by:[§4](https://arxiv.org/html/2609.17548#S4.SS0.SSS0.Px1.p1.4.4.4.3.3.3.3.3.3.3.2.1.1.1)\.
- Qwen Team \(2024\)Qwen2\.5 technical report\.arXiv preprint arXiv:2412\.15115\.Cited by:[§6](https://arxiv.org/html/2609.17548#S6.SS0.SSS0.Px3.p1.1.1.1)\.
- A\. van den Oord, Y\. Li, and O\. Vinyals \(2018\)Representation learning with contrastive predictive coding\.arXiv preprint arXiv:1807\.03748\.Cited by:[§5](https://arxiv.org/html/2609.17548#S5.SS0.SSS0.Px2.p1.3.2)\.
- M\. Zhang, J\. Lévy, C\. Rommel, J\. Rapin, C\. Bel, J\. Bonnaire, D\. Nieto, P\. Bourdillon, S\. Pinet, S\. d’Ascoli, T\. Moreau, and J\. King \(2026\)Accurate decoding of natural sentences from non\-invasive brain recordings\.Meta AI technical report\.Note:Brain2Qwerty v2, 29 June 2026\.[https://github\.com/facebookresearch/brain2qwerty](https://github.com/facebookresearch/brain2qwerty)Cited by:[§11](https://arxiv.org/html/2609.17548#S11.p38.2),[§2](https://arxiv.org/html/2609.17548#S2.SS0.SSS0.Px3.p1.1.1),[Table 2](https://arxiv.org/html/2609.17548#S2.T2.10.21.11.1),[§7](https://arxiv.org/html/2609.17548#S7.p5.1.1)\.