# Neuralink, Terry and speech restoration — sources and methods Reviewed September 23, 2026 for Managing Expectations Health. This is a source-based technology review, not a systematic review, medical advice, eligibility assessment or independent test of Neuralink hardware. ## Source chain User-supplied URL, preserved exactly: https://youtube.com/shorts/c7_YH-gbc8c?is=vv2XFGRDWveucpag Deeplearn HQ’s 59-second Short was uploaded September 23, 2026. Its MP4 was downloaded, checked for audio/video with ffprobe and transcribed with faster-whisper. Sampled frames show secondary narration intercut with participant footage and the company’s speech-restoration page. No platform captions were available for the Short; the transcript is automatic and names were checked externally. Original company film: https://www.youtube.com/watch?v=_j806JHhCRo Title: Speaking With The Mind | Neuralink. Uploader: Neuralink. Upload date: September 18, 2026. Runtime: 416 seconds. Official metadata, description and English captions were retrieved. Full-film MP4 download returned HTTP 403; no claim is made that the complete film was downloaded or independently tested. Full captions and raw media/metadata are kept privately; the public article embeds the official video and includes direct watch links. The official description says Terry was diagnosed with bulbar-onset ALS in 2024, retains mouth/body movement, and has speech that is hard to understand. It describes progression from miming speech to thinking words. The film identifies him as the third VOICE participant at approximately 2:12, not the third Neuralink recipient overall. Approximate caption anchors: 3:39 decoder training; 4:22 button-to-audio workflow; 6:26 family communication. Caption times are navigation aids, not laboratory latency measurements. September 18 is a publication date, not an independently established implant date. The primary company X post credits Grok Voice: https://x.com/neuralink/status/2100993363403063404 A successful primary extraction and secondary embeds preserved the wording. Later accesses were inconsistent. The company says output resembles Terry’s natural voice; this is synthetic/computer-generated speech, not proof of restored biological vocal function. The reviewed sources do not establish the precise voice-cloning data, complete software architecture or that Grok itself decodes neural signals. ## Company demonstration versus measured outcomes The official video, post, speech-restoration page and registry did not supply Terry-specific word-error rate, tested vocabulary, correct words per minute, latency distribution, test-set denominator, training-data quantity, longitudinal reliability or adverse-event rate. This is a scoped source limitation, not proof that data do not exist elsewhere. The film mentions about 3,000 channels, but the article does not treat that hardware claim as an efficacy measurement or replace it with older N1 specifications. A shown button-to-speaker workflow is not proof of an always-on, fully hands-free interface. The company’s deliberate imagined-speech claim is preserved without turning it into unrestricted mind reading or independently proving the absence of all motor activity. Independent reporting was read for context, not clinical replication: https://ndtv.com/health/news/elon-musks-neuralink-brain-implant-helps-speech-impaired-man-speak-first-words-12068823 https://teslanorth.com/2026/09/19/neuralink-terry-voice-trial-grok-voice ## VOICE primary registry https://clinicaltrials.gov/study/NCT07224256 https://clinicaltrials.gov/api/v2/studies/NCT07224256 https://neuralink.com/trials/speech-restoration/ The full ClinicalTrials.gov API record was checked independently by the main editor and a research review. Neuralink Corp; single-group, open-label early feasibility study; six estimated participants, not six completed implants or successful outcomes. The served record lists recruiting, no posted results, actual start October 3, 2025, estimated primary completion October 2028 and completion October 2031. Last update was posted November 4, 2025; a current retrieval timestamp does not make the sponsor’s status verification current. Primary outcomes are device- and procedure-related adverse events at 12 months. Secondary outcomes include longer safety follow-up to 48 months and monthly maximum correct words per minute. A monthly maximum is not typical everyday conversational performance. Registry inclusion and an FDA-regulated-device field are not marketing authorization. The record specifies ages 22–75; no healthy volunteers; severe speech impairment AND impaired upper-limb function; ALS, PLS, stroke or spinal cord injury; English communication and caregiver requirements, plus other criteria. The public webpage is an abbreviated overview, not a complete eligibility checklist. The official webpage required browser rendering; a plain HTTP/extractor response was only a shell. Its visible text was inspected successfully. No registry form was submitted and no patient information was entered. ## Independent peer-reviewed studies The three selected papers concern two distinct participants; Card and Wairagkar report different systems/analyses in the same ALS participant. These are independent of Neuralink, not three independent patient cohorts. Their metrics must not be attributed to Terry. 1. Card et al., NEJM (2024), DOI 10.1056/NEJMoa2314132, PMID 39141853. One ALS participant. Attempted speech decoded to text, then personalized text-to-speech. Final reported approximately 97.5% word accuracy over 8.4 months; roughly 32 words per minute in self-paced conversation; more than 248 cumulative hours. The 99.6% initial result used a 50-word vocabulary. Word-level performance is not the proportion of perfectly correct sentences. Researcher setup and residual motor/speech function limit generalization. https://pubmed.ncbi.nlm.nih.gov/39141853/ https://doi.org/10.1056/NEJMoa2314132 2. Littlejohn, Cho et al., Nature Neuroscience (2025), DOI 10.1038/s41593-025-01905-6, PMID 40164740. One post-stroke participant, cortical-surface recordings and 80-ms decoding increments. The published figure identifies a 50-PHRASE task, not the “50 words” shorthand in an NIH news summary. Full task, latency-reference and accuracy differences matter. Online cue-based tests and offline continuous-decoding demonstrations are kept separate. Some full publisher text was subscription-restricted; abstract, figures, supplementary descriptions and the data-availability section were accessible. The public article does not reproduce unsupported full-text statistics. https://www.nature.com/articles/s41593-025-01905-6 https://www.nature.com/articles/s41593-025-01905-6/figures/2 3. Wairagkar et al., Nature (2025), DOI 10.1038/s41586-025-09127-3, PMID 40506548. The same ALS participant as Card, with direct acoustic synthesis, intonation and short melodies. A six-choice sentence-matching task yielded 94.34% mean accuracy; open-transcription median word-error rate was 43.75%. These are different tests, not interchangeable accuracy claims. Published text says the output was often, but not consistently, intelligible. https://pmc.ncbi.nlm.nih.gov/articles/PMC12369848/ https://doi.org/10.1038/s41586-025-09127-3 Published article records/full-text sources are linked. Attempts to download final publisher/PMC PDFs were blocked or returned non-PDF content. A legitimate public draft press-proof PDF of the Wairagkar paper was obtained privately, but is not republished or labelled the final article. No complete final-study PDF download is claimed. ## Canada is a separate protocol https://clinicaltrials.gov/study/NCT06700304 https://www.uhn.ca/corporate/News/PressReleases/Pages/uhn-selected-canadian-site-for-can-prime-study-clinical-trial-involving-neuralink.aspx UHN’s November 20, 2024 release describes Health Canada authorization to recruit for CAN-PRIME at Toronto Western Hospital, focusing on external-device control. The CAN-PRIME API currently lists recruiting, with a last posted update May 31, 2025. Its function, endpoints and protocol are not VOICE’s speech-restoration study. The reviewed VOICE registry names UT Southwestern in Dallas. No Canadian VOICE access or routine speech-product availability is established. A broad unquoted CAN-PRIME API search returned unrelated results and was rejected. A sponsor-restricted Neuralink search identified the correct NCT06700304 record; its identity and protocol were inspected, rather than accepting a superficially similar search result. ## Regulatory, privacy and practical-care context https://www.fda.gov/medical-devices/how-study-and-market-your-device/breakthrough-devices-program FDA’s public program explanation separates expedited development/review from the safety/effectiveness requirements for marketing authorization. No case-specific designation date is asserted. A partial extraction of the implanted-BCI guidance introduction was available, but the direct full PDF request was blocked; it is not represented as a complete downloaded guidance document or a source for Terry-specific adverse events. https://neuralink.com/patient-registry-privacy-notice/ The full notice was read via browser rendering, last updated March 27, 2025. It says registry enrollment is neither trial enrollment nor a prerequisite; a separate research consent is required. It describes information collection, research-team/service-provider sharing, anonymized-information uses and withdrawal/deletion routes, with limits. It is not the complete implanted-trial neural-data contract. Questions about secondary AI training, voice-model control, cancellation, withdrawal and support are editorial questions—not misconduct findings. https://www.als.org/navigating-als/living-with-als/therapies-care/augmentative-alternative-communication https://alsnetwork.org/understanding-als/living-with-als/guides-for-daily-living/als-communication-devices These sources support discussion of speech-language pathology, communication boards, speech-generating devices, eye-gaze/head/switch access and voice/message banking. No individual device, trial enrollment or treatment change is recommended. Regional funding is not inferred from U.S. information. ## Publication scope and safeguards Fresh live search entries contained no Neuralink or matching Short article, so this is a new expanded Health feature, not a duplicate. Only ManagingExpectations.net is in scope. No PBC post, My Cancer PDF append, clinical enrollment, outreach or commerce transaction is performed. The public package contains original editorial notes, sanitized metadata, selected registry fields and an evidence ledger. Raw captions, signed media URLs, full upstream captures and draft proofs remain private. Deployment is limited to the explicit manifest with backups, current shared-file hash checks, exact public-file read-back and browser/mobile/search verification.