The Pulse
Neuralink Shows Thought-to-Speech Voice Reconstruction in VOICE Trial
Neuralink has shown participant Terry using an implanted brain-computer interface to produce spoken words from intended speech. The company says the VOICE trial combines neural decoding with Grok Voice to recreate a participant’s natural vo
AI.info Team ·
Neuralink has released a new demonstration from its VOICE clinical trial showing a participant named Terry producing spoken words through a brain implant and a reconstructed version of his own voice. The company published the demonstration on September 18, 2026, describing it as part of an effort to restore communication for people who can no longer speak.
In the video, Terry first helps train the system by miming speech as accurately as he can. Neuralink then shows him progressing to silently intending words and hearing them spoken aloud. The company says the system uses its brain-computer interface to decode intended speech, while Grok Voice provides the audible output.
Neuralink’s original post does not publish clinical performance figures, such as word-error rate, response latency, vocabulary size or sustained conversational speed. The demonstration therefore shows a working trial interaction, not evidence that the system performs consistently across users or ordinary conversations.
Terry’s Training Moves From Mimicry to Silent Speech
Neuralink says Terry’s training began with attempted or mimed speech. That stage gives the decoding system neural activity associated with the words he is trying to form, even when his physical speech is impaired. The later step asks him to think the words without visibly speaking them, with the system turning the resulting neural signals into audible speech.
Voice reconstruction adds a second layer to the problem. A text-only interface can convey the words a person selects, but it does not reproduce the sound of that person’s voice. Neuralink’s demonstration instead presents speech intended to resemble Terry’s natural voice, allowing the output to carry a more personal identity than a generic computer voice.
The company’s video does not explain how much previous voice data the system requires, whether Terry’s voice was recorded before his illness, or how the voice model handles unfamiliar words, emotion and changes in speaking style. Those details matter because a short scripted demonstration tests a narrower task than open-ended conversation.
VOICE Is Separate From Neuralink’s Cursor Trials
Neuralink’s better-known demonstrations have centered on the N1 implant decoding movement intentions to control a computer cursor or other digital functions. VOICE targets a different signal: the neural activity associated with attempting to move the mouth, tongue and vocal tract during speech.
The distinction matters for people with conditions such as amyotrophic lateral sclerosis, who may retain the intention to speak while losing the muscle control needed to produce intelligible sound. Neuralink’s public materials describe the VOICE system as investigational, and the company says its devices are not approved for general consumer use.
Research groups outside Neuralink have also demonstrated brain-to-voice systems for people with ALS. A peer-reviewed study in Nature described a system that decoded activity from implanted microelectrodes into synthesized speech, including intonation and other vocal features. That work shows that voice restoration is an active research area, while Neuralink’s new post provides a company demonstration of its own implant and software stack rather than an independent clinical comparison.
What the Demonstration Establishes—and What It Does Not
The September 18 video establishes that Neuralink has reached a stage where at least one VOICE participant can train a model through attempted speech and then generate audible output from silent speech intentions. It also shows the company pairing neural decoding with a voice-generation system supplied by Grok Voice.
Public evidence remains limited. Neuralink has not released trial data for Terry, including the number of sessions, the number of words decoded, the percentage of correctly understood speech or the time required to produce each utterance. Without those measurements, viewers cannot determine how close the system is to dependable everyday communication.
For now, the concrete result is a filmed demonstration of one participant moving from mimed speech to thought-driven voice output. The VOICE trial will need broader participant data and formal performance results before the system can be judged as a reliable assistive device.