# Speech Neuroprosthesis **Entity class:** Technology and device class **Domain:** Brain-Computer Interfaces / Speech Restoration **Doc Type:** Technology Router **Maturity:** Human-translational **Collection:** [[collections/Neurotech|Neurotech]] **Primary Sources:** [[research/Research Inferences|Research Inferences]] (INF-0002, INF-0007) ## Definition **A speech neuroprosthesis decodes intended speech from neural activity and emits it as text or synthesized voice.** It is the device class where the corpus's read channel reaches its highest bandwidth and its clearest clinical necessity: for people who are locked in, the interface is the only channel. Three programs now hold human results, and the recent gains belong to the decoder rather than the sensor, which means they compound across every implant already in a head. ## The 2025 and 2026 record | Program | Record | Date | |---|---|---| | [[wiki/BrainGate2 home BCI\|BrainGate2]] with [[wiki/T15\|T15]] | 3,800-plus unsupervised hours of home use; 183,060 sentences totaling 1,960,163 words at about 56 words per minute; 99.2% word accuracy against a 125,000-word vocabulary in structured testing | June 2026, *Nature Medicine* | | Stanford and Emory, [[wiki/Inner Speech Decoding\|inner speech]] | silently imagined sentences decoded at up to 74% accuracy from a 125,000-word vocabulary, behind a thought-password gate recognized above 98% | 14 August 2025, *Cell* | | [[wiki/Paradromics Connexus\|Paradromics Connexus]] | first Connect-One participant in real-time conversation, including a live phone call; distinct patterns for attempted and imagined speech | 14 September 2026 | | [[wiki/Neuralink VOICE Trial\|Neuralink VOICE]] | third participant progressing from mimed speech to thinking words and hearing them in a synthesized voice | 18 September 2026 | ## Forward Capability The commercial floor sits at parity with talking: [[wiki/Paradromics|Paradromics]] frames Connexus against transcribed human speech at roughly 40 bits per second. The interesting regime opens above it, where decoded output exceeds speech rate and the interface stops substituting for a lost channel. Two consequences follow. The restored person speaks in a voice generated by a model, which makes the identity object part biological and part trained artifact. And when the model supplies most of an utterance's structure, authorship becomes a measurable ratio rather than a philosophical question. ## Open Questions - Peer-reviewed latency and accuracy for open-vocabulary inner speech outside the laboratory. - How decoders hold up across years of electrode drift and recording-population change. - Whether an access-control gate becomes a standard feature or stays a research demonstration. ## Research Inferences These entries translate the forward-looking register in [[research/Research Inferences|Research Inferences]] into ordinary wiki prose. The tier labels apply to the inference, not automatically to every factual anchor inside it. Proposed IDs above INF-0500 come from the 2026-10-23 cycle: [[research/Research Inferences Integration Report - 2026-10-23|Research Inferences Integration Report — 2026-10-23]] · [[research/Research Inferences Integration Audit - 2026-10-23|Research Inferences Integration Audit — 2026-10-23]]. - **INF-0002 — Plausible.** Paradromics frames Connexus against transcribed human speech at roughly 40 bps, a comparison that sets the commercial floor at parity with talking. The interesting regime opens above it: once decoded output exceeds speech rate, the interface stops substituting for a lost channel and becomes a channel humans never had. - **INF-0007 — Strongly indicated.** Academic speech BCIs reached 62–78 words per minute with implanted electrodes and model-side decoding, which means the recent gains are attributable to the decoder more than the sensor. Decoder improvement is free of surgical risk and compounds across every existing implant already in a head. - **INF-0022 — Strongly indicated.** Locked-in patients are the population for whom the interface is the only channel, which makes them the group that will use the highest-bandwidth configurations first and longest. Clinical necessity, not enthusiasm, will produce the first humans who have lived years mediated by a decoder. - **INF-0083 — Strongly indicated.** Work through 2026 has moved from decoding externally supplied captions toward recovering participants' own inner speech via subject-specific neural-semantic alignment. Inner speech is the first genuinely private content to become instrumentable, and the technique requires no change to the underlying language model. - **INF-0096 — Strongly indicated.** Restoring speech with a synthesized version of the person's own pre-morbid voice is already technically routine. The restored person therefore speaks in a voice generated from a model, and the resulting identity object is part biological, part trained artifact. ## Relationships - **implementations:** [[wiki/Paradromics Connexus|Paradromics Connexus]] · [[wiki/Neuralink VOICE Trial|Neuralink VOICE Trial]] · [[wiki/BrainGate2 home BCI|BrainGate2 home BCI]] · [[wiki/BrainGate2 long-term home BCI study|BrainGate2 long-term home BCI study]]. - **method:** [[wiki/Inner Speech Decoding|Inner Speech Decoding]] · [[wiki/Semantic Neural Decoding|Semantic Neural Decoding]] · [[wiki/Speech Restoration|Speech Restoration]] · [[wiki/Neural Decoding|Neural Decoding]]. - **researchers:** [[wiki/Edward Chang|Edward Chang]] · [[wiki/Sergey Stavisky|Sergey Stavisky]] · [[wiki/Matthew Willsey|Matthew Willsey]]. - **population:** [[wiki/Locked-In Syndrome|Locked-In Syndrome]] · [[wiki/Amyotrophic Lateral Sclerosis|Amyotrophic Lateral Sclerosis]]. - **participant record:** [[wiki/T15|T15]]. - **cluster_home:** [[wiki/Neural Interfaces and Continuity Architecture|Neural Interfaces and Continuity Architecture]] · [[wiki/Cognitive Interface Layer|Cognitive Interface Layer]] · [[wiki/Brain-Computer Interfaces|Brain-Computer Interfaces]]. - **priority_context:** [[wiki/Med-X BCI-2035 Perspective|Med-X BCI-2035 Perspective]] ranks stroke and depression above speech loss by disability-adjusted life years. - **ecosystem:** [[collections/Neurotech|Neurotech]]. ## Sources - [Study of promising speech-enabling interface — Stanford Medicine, August 2025](https://med.stanford.edu/news/all-news/2025/08/brain-computer-interface.html) - [Paradromics Achieves Real-Time Speech with Connexus — Business Wire, 14 September 2026](https://www.businesswire.com/news/home/20260914434545/en/Paradromics-Achieves-Real-Time-Speech-with-Connexus-Brain-Computer-Interface) - [Neuralink brain implant helps man with ALS speak using his thoughts — MobiHealthNews, 2026](https://www.mobihealthnews.com/news/neuralink-brain-implant-helps-man-als-speak-using-his-thoughts) - [articles/Technologies for Consciousness Mapping and Transfer](articles/Technologies%20for%20Consciousness%20Mapping%20and%20Transfer.md) — section VII, the read channel, for the BrainGate2 home-use figures. As of 23 October 2026.