A married man named Terry, living with bulbar-onset ALS, said “I, I love you” to his wife on a video Neuralink published this week. The sentence was decoded from his brain by an implanted brain-computer interface and voiced through an AI clone of his own pre-ALS recordings. It is Neuralink’s most detailed publicly disclosed thought-to-speech demonstration to date, the first to route through xAI’s Grok Voice engine for personalised voice synthesis, and comes from the third participant in the company’s VOICE early-feasibility trial.
What Neuralink demonstrated
Terry produced two demonstration phrases in an at-home setting: “Hello, how are you doing today?” and “I, I love you” spoken to his wife, who is present on camera. His wife’s on-camera line: “To just be able to have a conversation with my husband, this means the world to us.”
Terry’s synthesised voice is produced by Grok Voice, xAI’s speech engine, using a clone trained on his own pre-ALS speech recordings. His training progression moved from attempted mimed speech to imagined (thought-only) speech. The demonstration crossed 5.4 million views within 48 hours across Neuralink’s X and YouTube channels.
The VOICE trial
VOICE is a Neuralink-sponsored early feasibility study registered on ClinicalTrials.gov as NCT07224256 under the formal title “VOICE: An Early Feasibility Study of a Precise Robotically Implanted Brain-Computer Interface for Communication Restoration.” The study evaluates the N1 implant plus the R1 surgical robot in participants with severe irreversible speech impairment. The N1 implant carries an FDA Breakthrough Device Designation. It is not FDA-approved and not commercially available.
VOICE sits alongside Neuralink’s three motor-decoding studies: PRIME (NCT06429735) in the United States, CAN-PRIME (NCT06700304) in Canada, and GB-PRIME (NCT07127172) in the United Kingdom. It is a distinct protocol focused on communication restoration, not motor cursor control.
Terry in the Neuralink implant sequence
Terry is the third participant enrolled in VOICE specifically. He is not Neuralink’s third overall patient. The prior implantee sequence on the motor-decoding PRIME arm includes Noland Arbaugh (P1, January 2024, cervical spinal cord injury), Alex (P2, 2024), and Brad Smith (PRIME first ALS participant, publicly announced May 2025). Another VOICE participant, Kenneth Shock, implanted in January 2026, appeared in a Neuralink clip posted on 16 September 2026 saying “I love you” to his wife. Terry’s disclosure came two days later and expands the visible pipeline detail. Additional PRIME and VOICE participants have been disclosed at intervals.
Brad Smith’s earlier speech restoration relied on cursor-typed text produced through the PRIME motor decoder, then read aloud through a cloned voice. Terry’s demonstration is a direct neural-to-audio decoding, where the intended speech is decoded from cortical activity and voiced without an intermediate typing step.
What Neuralink has not disclosed
Neuralink has not published words-per-minute, word-error-rate, latency, session count, or vocabulary size for Terry’s speech decoding. There is no peer-reviewed methods paper. Terry’s implant date, implant site, implanting surgeon, and cortical target region are not publicly stated. His last name, age, and city of residence are not publicly disclosed.
Where this lands in the invasive speech-decoding field
Two other invasive speech-decoding lines have published peer-reviewed benchmarks this cycle.
The UCSF Chang lab published Brosler et al. in Nature Neuroscience on 14 September 2026, reporting simultaneous speech and gesture decoding from a single 253-channel high-density ECoG surface array on two participants in the BRAVO trial (Bravo-1r with brainstem stroke, Bravo-6 with ALS). Bravo-1r reached median 100 percent decoding accuracy on both speech and gesture across three real-time conversation blocks.
The BrainGate consortium, in a UC Davis-led paper by Nicholas Card and colleagues, published a speech-decoding neuroprosthesis in the New England Journal of Medicine in 2024, using Utah microelectrode arrays implanted in Casey Harrell, an ALS participant. The BrainGate result set the accurate and rapidly calibrating speech neuroprosthesis benchmark for that year.
Platform contrast matters. Neuralink’s N1 is a skull-mounted implant connected to flexible polymer threads carrying penetrating microelectrodes. UCSF’s platform is a 253-channel high-density ECoG surface implant on the pial surface. BrainGate’s Utah arrays are rigid penetrating microelectrodes. Each platform trades surgical footprint against signal characteristics differently. Terry’s demonstration is the first Neuralink disclosure to route through xAI’s Grok Voice for personalised voice synthesis, arriving in a commercial invasive BCI operator field where Paradromics reported real-time speech in a University of Michigan participant nine days earlier.
Grok Voice as the speech-synthesis layer
The Grok Voice attribution is one of the more strategically interesting details in the disclosure. xAI’s speech engine sits on the output side of the pipeline, converting decoded intended-speech representations into audible cloned-voice audio. If Grok Voice becomes Neuralink’s default speech-synthesis layer across future VOICE participants, the pattern is a Musk-affiliated end-to-end pipeline from cortical acquisition through to voice output. Whether other invasive BCI operators adopt xAI or route to alternative speech-synthesis vendors (ElevenLabs, WellSaid, Play.ht, or open-source variants) is an integration question the field has not yet settled.
Not disclosed
The exact implant date, implant site, cortical target region, surgical team, and Terry’s demographic details beyond first name and marital status are not publicly disclosed. No peer-reviewed decoder-performance metrics have been published. No FDA IDE endpoint reporting is in the public domain for VOICE at this stage. Any downstream secondary coverage citing specific words-per-minute or word-error-rate figures is not sourced to Neuralink.
What to watch
Whether Neuralink publishes a peer-reviewed methods paper describing the N1-based speech decoder. The comparable milestone from BrainGate (Card et al. NEJM 2024) and from UCSF (Brosler et al. Nature Neuroscience 2026) both landed as peer-reviewed papers with quantitative endpoints. Neuralink’s field-competitive framing improves substantially if it matches that publication expectation.
Whether Neuralink discloses words-per-minute, word-error-rate, and vocabulary-size benchmarks for Terry’s decoding, in a form that can be compared like for like against BrainGate and UCSF numbers. Absent published numbers, Terry’s disclosure functions as a demonstration and a category signal rather than a benchmarkable performance claim.
Whether the fourth VOICE participant surfaces on a similar release cadence in the coming months. Neuralink’s demonstration cadence is now a leading indicator of both trial pace and disclosure discipline.
Whether Grok Voice becomes the standing speech-synthesis layer for VOICE participants, and whether other invasive BCI operators (Precision Neuroscience, Paradromics, Synchron) integrate with xAI or route to alternative speech-synthesis providers.
Whether the FDA converts the Breakthrough Device Designation to formal marketing authorization on any timeline visible from public FDA CDRH activity. VOICE remains an early feasibility study.
Sources
Primary and near-primary coverage:
- Neuralink Speech Restoration trial page
- ClinicalTrials.gov NCT07224256
- Neuralink on X, VOICE announcement thread
- ABC News on X, Terry “I love you” clip
- BASENOR, Neuralink Gives ALS Patient Terry His Voice Back, September 2026
- UNILAD Tech, ALS patient speaks for the first time in years just by thinking, 21 September 2026
- Tesla North, Grok Voice pipeline details
- Central Oregon Daily / ABC affiliate, third VOICE participant clip
Field-comparison references:
Cross-reference to prior InsideBCI coverage: