Idea mining: silent speech & neural interfaces (2026-08-08)¶

Private (design/ is gitignored). The public output of this work is the new "Silent speech: when the muscle does carry the words" section in docs/research/muscle-brain-control.md, refs 13–28.

PDFs of cited papers are kept in a local, gitignored cache and are never published. Only citations and our own summaries go public; no paper text is redistributed.

Corpus — 20 sources, every one verified before use¶

Verified via Crossref (journal DOIs) and DataCite (10.48550/arXiv.<id>). export.arxiv.org hangs from this sandbox — Crossref and DataCite both work, so use DataCite for arXiv, not the arXiv API. The machine-readable dump of the verified metadata is kept in the same local, gitignored cache.

Verification caught two errors before they reached the site:

LipLearner's first author is Zixiong Su, not Naoki Kimura. A search-results grouping implied Kimura; the record says otherwise.
Chenyu Tang is first author of the Nature Sensors 2026 SSI review — he is a principal in this field, not a junior co-author as the earlier reading suggested.

The ideas, ranked by how soon they could matter¶

1. Widen the activation seam from onset to intent ⭐ highest leverage¶

EMGBackend implements HotkeyBackend — it can say start and stop, and nothing else. Every system in the corpus produces a word or an intent, and none of them can express it through our seam. A protocol that admits an intent-carrying source is the smallest change that makes SilentWear, the Cambridge headphones, emg2speech and any BCI decoder pluggable.

Blocking design question, now public as open question 6: a source with a ~3% error rate can inject text the user never reviewed. What is the confirmation model? ADR-v2-010's needs_confirm (already used for gaze deixis) is the precedent.

2. Silent commands + voice dictation — the hybrid nobody ships¶

The corpus is unanimous: 96–97% on 10–30 words (Tang 2025, Kurotaki 2026), ~68% WER open-vocabulary (Gaddy & Klein 2020). So closed vocabularies are solved and dictation is not.

That maps onto our command grammar, not the dictation path. Silently mouthing undo / new line / send while the voice carries prose is buildable today and is a genuine product position. It also inverts the usual pitch — we are not claiming EMG will replace typing, we are claiming it should replace the modifier key.

3. The shared latent space is the cheap integration point¶

emg2speech: S3 (self-supervised speech) representations linearly predict EMG power at r = 0.85, and EMG mapped into S3 space synthesises audio with no vocoder training. Consequence: an EMG decoder targeting the representation a speech encoder already computes can reuse our whole downstream stack — LM, vocabulary priming, disfluency filter — rather than rebuilding it. Integration at the representation layer, not the text layer.

4. Cross-user domain shift is a problem we already own¶

emg2qwerty (108 users, 346 h) names domain shift across users and sessions as the central obstacle. We already carry per-user calibration and an encrypted on-device learning corpus for exactly that failure in the voice path. Same machinery, different signal — worth saying out loud to any collaborator.

5. A degradation ladder across activation sources¶

Cross-modal masking (del Blanco 2026) fuses sEMG + lipreading so the system survives losing one modality. We have several activation sources (key, EMG, gaze) treated as alternatives. Treating them as a ladder that degrades gracefully is the same idea and costs no new dependency.

6. A device-neutral activation contract¶

Inoue 2025 trains one model across heterogeneous electrode configurations. That is the sensor-side analogue of the contract/ vectors built for the Android port. An activation contract — "these are the events a source may emit, here are the cases it must pass" — would let a lab prove its adapter works without owning our pipeline. Reuses a pattern already proven here.

7. Evidence for the accessibility case (no code, immediate use)¶

emg2speech demonstrated with an ALS participant silently articulating.
Card et al. 2026 (Nature Medicine): long-term independent home use of an intracortical BCI for speech and cursor control.
Willett 2023: 62 WPM, 23.8% WER at 125k vocabulary; Card 2024: 2.5% WER.

The honest framing this supports, now on the public page: the fast, accurate, open-vocabulary paths still require surgery; every non-invasive path buys reliability by shrinking the vocabulary. That sentence is defensible and nobody else assembles the table behind it.

8. Power budget: an always-on wearable trigger is proven feasible¶

SilentWear: 14 channels, 20.5 mW, >27 h on 150 mAh, 15k-parameter CNN on-device. This retires "would a wearable trigger even last a day" as an objection.

Two publishable gaps we are positioned to fill¶

Nobody publishes false-activation rates over a real workday. Every paper reports session accuracy on 3–4 subjects. FP/hour across a working day, with electrodes re-seated between sessions, is the number that decides whether a trigger is livable — and we have the daemon, the logging and the users to measure it. Already open question 1 on the public page.
There is no open closed-loop testbed for non-keyboard text input. Every group rebuilds a demo harness. An Apache-2.0, cross-platform pipeline with a documented activation seam is that testbed. This is both the collaboration pitch (see an internal planning note) and a plausible systems paper.

Next¶

Ideas 1 and 2 are roadmap candidates; neither is filed as an issue yet.
Ideas 7 and 8 are pure citation work — the accessibility use-case page could reuse them today.
Same treatment could be applied to docs/research/eye-control.md and voice-control.md; only muscle-brain-control.md was swept in this pass.