← SSI archive · Review rubric

2026 · arXiv · Field expert review · confidence high

EMG-to-Speech with Fewer Channels

Injune Hwang, Jaejun Lee, Kyogu Lee

arXiv

@misc{emg-to-speech-with-fewer-channels,
  title = {EMG-to-Speech with Fewer Channels},
  author = {Injune Hwang and Jaejun Lee and Kyogu Lee},
  year = {2026},
  note = {arXiv},
  eprint = {2602.06460},
  archivePrefix = {arXiv},
  url = {http://arxiv.org/abs/2602.06460v1},
}

Exhaustive subset search shows useful channel complementarity, and full-channel pretraining helps smaller EMG inputs. Single-person evaluation, unclear selection independence and a dropout text/figure conflict limit layout recommendations.

Reading guidance

Verdict

full-text draft · priority high · confidence high

Why it matters

Makes channel interactions and task-dependent sensor utility explicit, rather than equating individually important channels with the best low-channel system.

What to trust

Basis: full text + summary. Coverage: high. 10 evidence records back the review.

What is weak

Automatic WER combines encoder, vocoder and recognizer errors. Anatomical assignments are interpretations of broad surface recordings. Paper describes resizing the first convolution for reduced inputs but does not fully specify pretrained weight transfer. Exact checkpoint, optimization schedule and result split are not provided in the paper. One subject, no explicit sample counts or independent split for choosing among seventy subsets versus reporting their final WER. No repeated-run intervals, statistical tests or human listening study. Frame/category phoneme errors are not interchangeable with synthesized-speech WER. Figure 3 conflicts with Section 5.4 about the best dropout setting at four channels. Retrospective channel removal from a fixed eight-channel recording setup; no physically redesigned four-channel device, latency, power, comfort or clinical communication trial. Existing one-person open-vocabulary dataset and offline synthesized-speech evaluation; no new clinical, multi-person, mobile or real-world recording study. Overclaim risk: Moderate-high for practical sensor-layout generalization or anatomical causality; moderate for the narrower empirical transfer/subset conclusions..

Read before

SSI review rubric

Read next

SSI archive

Axes

Task

speech-reconstruction

Modality

Surface electromyography from selected channels of an eight-channel face/neck array; voiced acoustic targets and phoneme labels are training supervision.

Hardware

Eight existing surface-EMG channels distributed across cheeks, chin and throat, with retrospective subsets. Figure 1 lists positions including left cheek, left chin, below chin, throat and right facial sites; physical reduced-channel acquisition is not tested.

Body site

face;neck

Output

speech-audio

Vocabulary

continuous open-vocabulary corpus speech

Metrics

Table 1 four-channel WER: subset 1356 47.2%, 2357 47.3%, 1346 47.7%; at least ten combinations outperform greedy 1234. Table 2 inclusion-average WER: channel 3 51.4%, channel 2 52.3%, channel 1 52.6%, channel 5 52.8%. Table 3 total phoneme error 16.0% with eight channels and 17.1% after removing channel 8; silence 4.5% versus 5.8%. Figure 3 places the full-channel WER reference at approximately 35%; exact fine-tuned numeric values are not tabulated. Author-reported, not independently reproduced.

Evaluation mode

Offline channel-ablation study with greedy backward elimination, exhaustive four-channel search, direct phoneme-head category errors, and HiFi-GAN synthesis transcribed by Whisper medium.

Review confidence

high

Overclaim risk

Moderate-high for practical sensor-layout generalization or anatomical causality; moderate for the narrower empirical transfer/subset conclusions.

Expert take

The useful result is that sensor selection is a joint prediction problem: channel 6 is removed first by greedy elimination yet belongs to the best four-channel subset, while channel 3 appears in nine of the ten highest-ranked subsets. Evaluating all seventy combinations is a meaningful improvement over a partial search, and the paper shows why a ranking of individual channel importance should not be treated as an optimal device layout. The best from-scratch four-channel combination, 1/3/5/6, reaches 47.2% recognizer-derived WER; full-channel pretraining reduces the penalty of using four to six channels, but the plotted reduced-channel results remain above the roughly 35% eight-channel reference. Transfer learning and channel dropout must also be separated: plain pretraining already helps, and the paper does not show that added dropout always improves it. In fact, Section 5.4 says no dropout is best with four channels, whereas Figure 3 places the 0.125-dropout curve lowest there. That discrepancy needs source results before selecting a training policy. The phoneme analysis is valuable as a complementary diagnostic, but category errors and surface-electrode locations do not isolate causal contributions of particular muscles; silence is included in the overall phoneme error, and channel importance differs from the synthesis-WER ranking. Finally, all results come from one existing participant and virtually removed channels. The work offers a credible training and sensor-search baseline, while independent subset validation and a genuinely reduced wearable recording study remain necessary before claiming a practical sensor layout.

True value

Makes channel interactions and task-dependent sensor utility explicit, rather than equating individually important channels with the best low-channel system.

What changed

Canon before

Gaddy-style EMG-to-speech already maps facial/neck muscle signals to acoustic features with phoneme supervision and a vocoder. Reduced-channel subsets and dropout are established ideas; channel information depends on interactions and the downstream objective.

Delta from canon

Completes the full four-of-eight subset search and compares random initialization with full-channel pretraining at dropout probabilities 0, 0.125 and 0.25.

Position in field

Single-participant EMG-to-speech benchmark analysis aimed at fewer facial/neck sensors.

Evidence

Limits

Technical limits

Automatic WER combines encoder, vocoder and recognizer errors. Anatomical assignments are interpretations of broad surface recordings. Paper describes resizing the first convolution for reduced inputs but does not fully specify pretrained weight transfer. Exact checkpoint, optimization schedule and result split are not provided in the paper.

Evaluation limits

One subject, no explicit sample counts or independent split for choosing among seventy subsets versus reporting their final WER. No repeated-run intervals, statistical tests or human listening study. Frame/category phoneme errors are not interchangeable with synthesized-speech WER. Figure 3 conflicts with Section 5.4 about the best dropout setting at four channels.

Deployment limits

Retrospective channel removal from a fixed eight-channel recording setup; no physically redesigned four-channel device, latency, power, comfort or clinical communication trial.

Scope limits

Existing one-person open-vocabulary dataset and offline synthesized-speech evaluation; no new clinical, multi-person, mobile or real-world recording study.

Manual check: Release exact Figure 3 values and resolve the four-channel dropout contradiction. State the train/development/test utterance counts and isolate subset/dropout selection from final evaluation. Pin the released code and identify which split produced each result. Explain adapting the first convolution or input mask for pretrained reduced-channel models. Repeat across seeds, sessions and participants, and physically record the reduced array.