Inner speech in motor cortex and implications for speech neuroprostheses

Cell 188(17), 21 August 2025, doi 10.1016/j.cell.2025.06.015; PMC12360486. The PMC text is the "unedited manuscript that has been accepted for publication". Read by the main session for research direction R3, 4 October 2026. Provenance: papers/amadeus/kunz2025_inner_speech.provenance.json.

What was read

• Read in full: the Europe PMC full-text XML, converted to text:
  • summary, introduction, results, discussion, limitations;
  • all STAR Methods (participants, signal processing, every task and analysis);
  • captions of Figures 1–7 and S1–S6, the key resources table, declarations and references.
• Not read:
  • the supplementary tables (S1–S7: behaviours, words, sessions, trial counts, windows, statistics, training data);
  • Video S1; the figures as images; the Dryad dataset and the code.
  • The 23 authors above are taken from the XML's contributor list. An earlier version of this line was not, and was wrong; corrected 4 October.

What they did

• Participants. Four people in the BrainGate2 trial, each with 2–6 intracortical microelectrode arrays (64 channels each) along the left precentral gyrus: areas 6v, 55b, 4, 6d and PEF, placed by individual HCP parcellation.
  • T12: ALS, severely dysarthric.
  • T15: ALS.
  • T16: pontine stroke, dysarthric.
  • T17: ALS, anarthric and ventilator-dependent; communicates only with eye movements.
  • Data come from post-implant days 88–995.
• Tasks:
  • Seven single-syllable words under attempted (vocalised or mimed) speech, three inner-speech strategies (motoric, 1st-person auditory, 3rd-person auditory), listening, and silent reading.
  • Real-time decoding of self-paced inner speech of whole sentences: an RNN gives phoneme probabilities, an n-gram language model gives words. Training used up to 1 hour of cued inner speech plus earlier attempted-speech data, with a separate input layer per session.
  • Uninstructed inner speech:
    • serial recall of arrow sequences, with control tasks designed not to elicit speech;
    • silent counting of coloured shapes;
    • prompted thinking ("think about the lyrics of the first song that comes to mind", "think about your daily morning routine", "clear your mind").
  • Privacy mechanisms: "imagery-silenced" training (inner speech labelled as silence) and a keyword unlock ("chittychittybangbang").

Main results (verified)

• Inner speech, listening and reading are represented in ventral and mid precentral gyrus (areas i6v, 55b) as scaled-down versions of attempted speech.
  • Same word geometry, smaller modulation: T12's motoric inner speech is about 52% of attempted speech.
  • Seven-word decoding: T12 inner speech 72.6% (attempted 97.9%); T16 listening 92.1% (chance 14.3%).
• Real-time inner-speech BCI.
  • 50-word vocabulary: word error rate 24% (T12), 14% (T15), 33% (T16).
  • 125,000-word vocabulary: 26–54% (T15, T16).
  • A decoder trained only on attempted speech decodes inner speech above chance.
  • All participants preferred inner speech for its lower effort.
• Uninstructed inner speech is partly decodable.
  • Serial recall: in T12's i6v, all three positions of a 3-arrow sequence were decoded from the delay period. The single-arrow and line-drawing controls were not. Instructing a verbal strategy raised decoding at every position. T16 replicated it more weakly (go period only).
  • Counting: decoded numbers rose over the trial (slopes 0.48 and 0.33, p ≈ 1e-9) under a unigram number-only language model. Instructed sentences gave no slope.
  • Prompted thinking: more words were decoded for verbal-thought prompts than for "clear your mind": T15 22.3 vs 5.8; T16 9.3 vs 4.6. Autobiographical prompts gave 6.7 (T15) and 7.2 (T16).
  • "Most decoded sentences were largely gibberish with occasional plausible phrases". The outputs were withheld for privacy and uncertain representativeness.
• A "motor-intent" dimension separates attempted from inner speech. It is as large as, or larger than, word modulation, and survives in anarthric T17. Removing it leaves word accuracy but drops behaviour accuracy.
• Preventing unintended decoding.
  • Imagery-silenced training keeps attempted-speech performance and suppresses output on inner speech (offline).
  • The keyword unlock was detected correctly in 98.75% of real-time trials (95% CI 93.2–99.97%). Word error rate on unlocked trials was 43.45%.

Limits

• The authors' own:
  • Four participants; individual variation in how inner speech is used.
  • "It was not possible to accurately decode complete, intelligible sentences during free-form thinking."
  • Decoding may be limited to concrete verbal strategies (verbal memory, counting, lyrics). "Private inner monologue likely differs between individuals and may not unravel concretely, which could make it difficult or impossible to decode from motor cortex."
• Online retraining in evaluation. T15's evaluation blocks used online retraining, including on the sentences being evaluated. T16's 125k-vocabulary decoder retrained on cued sentences after decoding them. Those headline numbers come from adaptive decoders.
• Manuscript slips:
  • a session named "t15.2025.12.15" where "t15.2024.12.15" is meant;
  • the counting null histogram cited as "Figure 6G–H" (it is 5G–H);
  • Fig. 3D says chance used 100 shuffles where the methods say 10,000;
  • placeholder "[DOI]" and "[repo]" in the availability statement.
  • None affects the main claims.
• Evidence weight. Uninstructed-speech effects rest on 1–2 participants per task, few prompted-thinking trials (each prompt shown once), and counts of decoded words, not content accuracy.
• Population. Motor cortex speech areas only, in people with paralysis who had years of practice with attempted-speech BCIs.

What it means for Amadeus (inference)

• The best current answer to "can recording give access to someone's thinking": partly, for the verbal part, from speech motor cortex, with implants. Instructed inner speech decodes at word error rates of 14–54%. Uninstructed inner speech leaves decodable traces when the task forces verbal working memory (counting, rehearsing a sequence). Free-form thought decodes to gibberish. That bounds the month-of-thinking idea: even with intracortical arrays, today's decoders read verbal thought when it is concrete. Daydreaming, imagery and non-verbal deliberation are not reached here.
• Thinking is not all inner speech. Inner monologue varies between people, and the authors cite the debate over whether language is used for thought at all (Fedorenko et al. 2024, not read). Some of a person's thinking may never pass through speech areas. A speech-area recorder would capture a person's verbal habits of thought and miss others. That argues for multiple anchors (imagery, affect, choices), not one.
• The adapter pattern again. The decoder uses a separate input layer per recording session around a shared RNN, the LFADS stitching idea at clinical scale. Each person's decoder is trained on their own data. Nothing is shared across people.
• The intended/unintended distinction is measurable. The motor-intent dimension separates "saying" from "thinking". For Amadeus's provenance fields that is a measured variable of what the person meant to express versus what passed through their mind. It is a candidate neural correlate of "said" versus "thought", relevant to the distinction between what someone states and what they hold. It does not detect lying.
• Consent can be built into the decoder (keyword unlock, imagery silencing). If a month of recording ever happens, a person-controlled gate on what gets decoded is technically available. It belongs in the ethics design (architecture doc, "Ethics").
• Not an E3 dataset. The data are on Dryad, but the autobiographical prompts are few, single trials, with no later behaviour to predict, and the decoder outputs are withheld.