2026 · arXiv · Field expert review · confidence high

Silent Speech Interfaces in the Era of Large Language Models: A Comprehensive Taxonomy and Systematic Review

Kele Xu, Yifan Wang, Ming Feng, Qisheng Xu, Wuyang Chen, Yutao Dou, Cheng Yang, Huaimin Wang

BibTeX

@misc{silent-speech-interfaces-in-the-era-of-large-language-models-a-comprehensive-taxonomy-and-system,
  title = {Silent Speech Interfaces in the Era of Large Language Models: A Comprehensive Taxonomy and Systematic Review},
  author = {Kele Xu and Yifan Wang and Ming Feng and Qisheng Xu and Wuyang Chen and Yutao Dou and Cheng Yang and Huaimin Wang},
  year = {2026},
  note = {arXiv},
  eprint = {2603.11877},
  archivePrefix = {arXiv},
  url = {http://arxiv.org/abs/2603.11877v1},
}

A broad and useful SSI taxonomy, but missing review-selection detail and concrete citation errors make its benchmark tables unsuitable for unverified rankings or deployment claims.

Verdict: full-text draftPriority: highConfidence: highBasis: full text + summaryCoverage: high

Reading guidance

Verdict: full-text draft · priority high · confidence high
Why it matters: A cross-modal reading map that brings language-model priors and generative synthesis into the broader SSI landscape, with numerical claims requiring primary-source verification.
What to trust: Basis: full text + summary. Coverage: high. 9 evidence records back the review.
What is weak: Reference mismatches, mixed metric scales, duplicate rows and underspecified cohort/split conditions limit the tables. Statements about inherent noise immunity, recovered intent and universal latency requirements exceed what a heterogeneous narrative review can establish. No explicit reproducible search query/date, screening counts, eligibility rules, extraction procedure or study-quality assessment is reported. Tables mix recognition, reconstruction, perception, typing and articulator-analysis tasks with incompatible metrics and sparse split details. Citation mismatches and duplicated entries require correction. Wearable, clinical and noisy-environment examples are taken from different cited studies; no integrated system or deployment experiment is performed by this review. Narrative synthesis of heterogeneous SSI and adjacent resources. Its tables include listening EEG and wrist typing, which do not directly establish silent speech-production decoding. Overclaim risk: High for comprehensive systematic coverage, cross-modal rankings, universal usability thresholds and field-wide parity with acoustic ASR; lower for the descriptive taxonomy..
Read before: SSI review rubric
Read next: SSI archive

Axes

Task: survey
Modality: Survey of neural, EMG, magnetic, ultrasound, optical, acoustic and RF sensing, including adjacent speech resources.
Hardware: No new hardware; surveys EEG/ECoG, sEMG, EMA/PMA, ultrasound, MRI, cameras/depth, earables, contact sensors and radar.
Body site: brain; face; jaw; lip; oral-cavity; throat; tongue
Output: Literature taxonomy, benchmark/dataset tables and research roadmap
Vocabulary: Mixed closed-set, open-vocabulary, reconstruction and adjacent tasks
Metrics: No new system performance measurement. Tables II-IV reproduce accuracy, WER, CER, PER, MCD, MOS and other values from heterogeneous studies. Some rates appear as proportions and others as percentages. The review repeatedly proposes 15% WER and 50 ms latency thresholds, but supplies no common user study establishing them as universal criteria.
Evaluation mode: Narrative taxonomy and tabulation of published results; no matched benchmark, meta-analysis or new user evaluation.
Review confidence: high
Overclaim risk: High for comprehensive systematic coverage, cross-modal rankings, universal usability thresholds and field-wide parity with acoustic ASR; lower for the descriptive taxonomy.

Expert take

This review is useful for orienting readers across neural, muscular, articulatory, imaging and active-sensing approaches to silent speech. Its separation of recognition, direct synthesis and language-model integration helps explain where linguistic priors enter a pipeline and why sensor information, user calibration and computational cost remain important. Its quantitative synthesis needs substantially more caution. Although the title calls it systematic, the paper does not provide a reproducible study-selection and quality-assessment procedure; a Web of Science publication-count figure is not a documented search-and-screening protocol for the review itself. Tables combine small-command accuracy, open-vocabulary error rates, acoustic distortion, auditory-perception decoding and even wrist typing, without the split and calibration details needed for fair comparisons. There are also concrete attribution errors: Table IV assigns a 125,000-word intracortical ALS result with 2.5% WER to reference 132, which the bibliography identifies as Silentspeller, an electropalatography text-entry paper. Table II places references 83 and 88 under ultrasound, although their listed titles concern electro-optical stomatography. These are not merely stylistic issues; they break the path from a claimed result to its evidence. The repeated 15% WER usability threshold and 50 ms latency requirement should therefore be treated as contextual claims, not general acceptance standards demonstrated by this survey. The paper can guide literature discovery and research questions, but its numerical tables should not be imported into a leaderboard or used to infer clinical readiness without checking the original studies.

True value

A cross-modal reading map that brings language-model priors and generative synthesis into the broader SSI landscape, with numerical claims requiring primary-source verification.

What changed

Canon before

Earlier SSI reviews organize biosignal sensors and recognition/synthesis pipelines. This review adds emphasis on modern generative models, representation alignment and wearable form factors.

Delta from canon

Connects sensing positions along the speech-production chain with representation learning, direct synthesis and LLM-based correction or embedding alignment.

Position in field

Broad SSI synthesis spanning restorative communication, consumer interaction, sensing and generative decoding.

Evidence

“ The review organizes SSI sensing along neural, neuromuscular and articulatory stages, then relates recognition, synthesis and language-model methods to these modalities. ”

actual_novelty · Sections II-III; Tables I-IV; PDF pp. 3-10 · confidence 0.99

“ Figure 1 describes Web of Science keyword publication counts for 2011-2025. No explicit study-screening flow, inclusion/exclusion protocol or quality-assessment procedure accompanies the review. ”

validation_scope · Figure 1; Introduction and review structure; PDF pp. 1-3 and 10-14 · confidence 0.99

“ Table IV attributes an intracortical ALS result with 125,000-word vocabulary and 2.5% WER to reference 132; the bibliography identifies reference 132 as Silentspeller, an electropalatography text-entry paper. ”

limitation · Table IV and Reference 132; PDF pp. 9 and 18 · confidence 0.99

“ Table II classifies references 83 and 88 as UTI, while their bibliography titles explicitly concern electro-optical stomatography. This is a modality-to-reference mismatch. ”

limitation · Table II and References 83/88; PDF pp. 6 and 16-17 · confidence 0.99

“ Tables II-IV combine command accuracy, word/character/phoneme errors, distortion, top-10 perception accuracy and other measures. Evaluation splits and calibration conditions are not systematically tabulated, preventing fair cross-system ranking. ”

limitation · Tables II-IV; PDF pp. 6, 8 and 9 · confidence 0.99

“ Table V includes wrist EMG typing, passive audiobook EEG and auditory-evoked EEG alongside speech-production datasets. Reviewer assessment: these are adjacent resources, not interchangeable silent speech-decoding evidence. ”

validation_scope · Section IV-B and Table V; PDF pp. 11-12 · confidence 0.99

“ The text repeatedly describes 15% WER as a usability threshold and 50 ms as an end-to-end latency requirement, without a common user evaluation establishing universal applicability across the surveyed tasks. ”

limitation · Sections III-D and IV-A; Section VI-B; PDF pp. 10-11 and 14 · confidence 0.99

“ The language-model discussion distinguishes post-processing, candidate rescoring and direct biosignal-to-embedding alignment, and acknowledges latency costs and over-correction that can diverge from user intent. ”

actual_novelty · Section III-D; PDF pp. 9-10 · confidence 0.98

“ The roadmap identifies user/session variation, missing parallel targets, continual learning, multimodal alignment and privacy as unresolved directions; these are proposed research priorities rather than tested solutions. ”

validation_scope · Sections IV-C and VI; PDF pp. 11 and 13-14 · confidence 0.99

Limits

Technical limits

Reference mismatches, mixed metric scales, duplicate rows and underspecified cohort/split conditions limit the tables. Statements about inherent noise immunity, recovered intent and universal latency requirements exceed what a heterogeneous narrative review can establish.

Evaluation limits

No explicit reproducible search query/date, screening counts, eligibility rules, extraction procedure or study-quality assessment is reported. Tables mix recognition, reconstruction, perception, typing and articulator-analysis tasks with incompatible metrics and sparse split details. Citation mismatches and duplicated entries require correction.

Deployment limits

Wearable, clinical and noisy-environment examples are taken from different cited studies; no integrated system or deployment experiment is performed by this review.

Scope limits

Narrative synthesis of heterogeneous SSI and adjacent resources. Its tables include listening EEG and wrist typing, which do not directly establish silent speech-production decoding.