Myovox: Reading Speech from the Muscles of the Face
cs.CL
Quick verdict
Claims 51.17%→18.53% WER on single-subject sEMG speech decoding via decode fixes, a bidirectional Conformer, and LLM reranking; rigorous but narrow single-speaker scope.
Summary
Myovox decodes open-vocabulary English text from 31-channel facial surface EMG recorded during vocalized speech, on the single-subject emg2speech General Corpus (8,500/760/400 sequential split). It reports three isolated improvements over the published Gowda et al. baseline: (1) recovering missing decode settings (blank penalty, PER-based checkpoint selection) to reach 40.63% WER with PER matched to the published 38.19%, supporting faithful acoustic reproduction; (2) replacing the causal TDS encoder with a bidirectional Conformer trained with a four-term objective including WavLM-Large layer-9 distillation, reaching 26.14% WER / 22.34% PER — though an EMG-only control matches this at the word level (26.10%); (3) ensembling two Conformers, multi-scale n-best union, and QLoRA-tuned Qwen2.5-7B reranking, reaching 18.53% WER. The central negative finding is that reranking saturates: the ~20.9% acoustic phone error rate binds, correct words are absent from the posteriors, and the 9.30% n-best oracle is unreachable. Limitations: single healthy speaker, vocalized speech only, no measured acoustic-leakage control, untested generalization.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The published 51.17% WER baseline on the emg2speech General Corpus is reproduced faithfully (matched greedy PER 39.02 vs 38.19) and improved to 40.63% WER purely by recovering missing open-vocabulary decode settings, chiefly a blank penalty.
- A bidirectional Conformer trained with a four-term WavLM-Layer-9 cross-modal distillation objective cuts acoustic error to 26.14% WER / 22.34% PER, and an EMG-only control matches it at the word level (26.10%), so parallel audio is not required for the word-level gain.
- Ensembling two Conformers, multi-scale n-best union (oracle 11.94%→9.30%), and QLoRA-fine-tuned Qwen2.5-7B reranking reach 18.53% WER — the best reported on this corpus — with leakage controls and bootstrap-significant gains.
- Reranking is exhausted at ~18.5% because the ~20.9% acoustic phone error rate binds: the correct words are absent from the acoustic posteriors, so the 9.30% n-best oracle is unreachable and further gains must come from the acoustic model, not the language model.
Strengths
Strict isolation of every intervention
Each stage is measured separately (Tables 1, 4-6), decomposing the 32.6-point headline into ~10.5 points of decode fix, ~14.5 of modeling, and ~7.6 of extraction — an unusually clean ablation structure for a single-author report.
Matched-PER reproduction argument
Section 4 anchors baseline credibility on greedy PER (39.02 vs published 38.19), which is decoder-independent; this is the right metric given that the public release cannot regenerate the published 51.17% WER (~75% without the missing settings).
A control that undercuts the favored component
Leakage controls for the 7B reranker
A load-bearing negative result
Weaknesses
Single subject, single mode
No measured acoustic-leakage control
Headline improvement anchored to an unreproducible number
Distillation objective not ablated per term
Ensemble and audio-free headline unresolved
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What word error rate does Myovox achieve?
18.53% WER on the 400-sentence test set of the single-subject emg2speech General Corpus, down from the published 51.17%. Roughly one word in five is still wrong, so the output is useful with human correction but not yet reliable unattended. The author also reports 18.75% after excluding six test sentences that duplicate training text.
Can Myovox read thoughts or work for silent speech?
No. It decodes facial muscle activity (surface EMG) during vocalized — aloud — speech, not brain activity and not silent articulation. An EMG-only control model matches the audio-distilled one at the word level, which is encouraging for future silent-speech use, but silent speech itself was never tested, nor was any patient population.
Is Myovox the state of the art for EMG-to-text?
Only on its specific corpus. MONA LISA reported 3.7% WER on the separate Gaddy corpus (vocalized EMG), far lower, but the corpora, electrode counts, and speakers differ, so the numbers are not directly comparable. 18.53% is the best figure the author is aware of on the emg2speech General Corpus, against a published 51.17%.
Is the code and data for Myovox available?
Code, configurations, and the recovered decode settings are MIT-licensed at github.com/Varshith-0/myovox, and training runs on a single 12 GB consumer GPU in roughly 15-25 hours per experiment. The emg2speech General Corpus is not public and must be obtained from its original authors.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The abstract in the extracted text is truncated mid-sentence; the full body covers the missing content (all test numbers are on the 400-sentence held-out test set).
- No figures are present in the extract — the report states architecture diagrams live on its project website — and Equation 1 is reconstructed from inline text.
- Table formatting and the reference list may be imperfectly extracted; page-level layout and any supplementary material were not verifiable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.