A Connexus implant let a woman speak her own words — and told attempted speech from imagined

Paradromics announced on 14 September 2026 that the first participant implanted with its Connexus brain-computer interface has used it for real-time speech and text communication. The participant — a Michigan woman in her sixties living with a progressive motor neuron disease — moved past researcher-selected copy tasks: she chose her own words, answered open-ended questions, and spoke with her grandchildren during a live phone call. The work is part of the Connect-One early feasibility study under an FDA-approved Investigational Device Exemption. Paradromics additionally reports identifying distinct neural patterns for attempted speech, where the person tries to produce sound, and imagined speech, where language is generated internally with no attempt to speak. The tracker records the result.

Why it matters

A speech neuroprosthesis is not new — the first high-accuracy real-time system was demonstrated in 2024. What is new is the word unscripted. Copying a phrase the researchers chose tests decoding against a known target; choosing your own words and answering a question nobody planned removes that scaffold entirely, and a phone call with family removes the lab. That is the difference between a system that decodes speech and a system a person uses to speak.

The attempted-versus-imagined finding is the part with the longer tail. Every implanted speech system so far has decoded the motor attempt — the signals sent toward muscles that no longer respond. That works only while the attempt exists, which for a degenerative condition is a closing window. If imagined speech carries a separable signal, decoding could outlive the ability to try, and the same implant gains a second control channel that does not compete with the first. It also opens an obvious question the field will have to answer deliberately rather than by accident: a device that can read language you did not attempt to say needs a clear boundary around when it is listening.

The limits are the usual ones and they are real. This is one participant, in an early feasibility study, and Paradromics has published no accuracy rate, vocabulary size, or durability figure. Its August clearance widened what the implant may control; this shows what it can do. Neither is a marketing approval.

What to watch

Whether accuracy and vocabulary figures are published, whether additional Connect-One participants reproduce the result, and whether imagined-speech decoding holds up outside a single case.