Overview

A speech BCI records from the part of the motor cortex that controls the muscles of speech and decodes attempted speech into phonemes, words, text or a synthesised voice. The main hardware routes are penetrating microelectrode arrays such as the Utah array and high-density surface grids (ECoG). Published 2023 results put the Stanford BrainGate system at about 62 words per minute with a 23.8 per cent word-error rate on a 125,000-word vocabulary, and UCSF’s avatar system at about 78 words per minute.

Academic programmes

At UC Davis, Casey Harrell, who has ALS, has used a 256-electrode intracortical system at home for more than 3,800 hours over 22 months without researchers present; on a prompted word-copy task with a 125,000-word vocabulary it reached 99.2 per cent accuracy. At UCSF, Edward Chang’s group decoded speech and gestures together from a 253-channel ECoG array to drive a virtual avatar, published in Nature Neuroscience in September 2026. A Stanford-led BrainGate study found two speech-preferential areas in the precentral gyrus, with implications for where future arrays are placed.

Company trials

In September 2026 the first participant implanted with Paradromics’ Connexus spoke her own words in real time three months after implantation. In Neuralink’s VOICE trial, participant Terry spoke through a clone of his own pre-ALS voice using the Neuralink N1. Neither company has published words per minute, error rates or vocabulary size. In China, NeuroXess disclosed in May 2026 a figure of 300 Mandarin characters a minute, decoded through a language-model layer in a woman with epilepsy, ahead of a paper.

Why it matters

Speech restoration gives implanted BCIs a defined patient group, people with ALS or brainstem stroke, and outcomes that can be measured in words per minute and error rates. The commercial question is whether company systems match published academic performance and work reliably at home without research staff.

Coverage on InsideBCI

• 23 September 2026: Neuralink discloses direct thought-to-audio speech decoding in ALS participant Terry, the VOICE trial’s third implantee
• 18 September 2026: UCSF brain-computer interface decodes speech and body gestures simultaneously from a single cortical implant
• 16 September 2026: Paradromics’s first Connexus participant used a brain implant to speak her own words
• 28 June 2026: Stanford maps the human precentral gyrus at single-neuron resolution across eight paralysed BCI participants
• 16 June 2026: A man with ALS has used his speech brain implant at home for nearly two years, mostly on his own