UCSF speech neuroprosthesis (Metzger, 2023)

One-line verdict: A surface array over speech cortex, not a penetrating one, decoded attempted silent speech at 78 words per minute. It is the benchmark most speech BCIs get compared against.

Quick tags: Recording · Cortical surface · Human · Published 23 August 2023 (Nature)

Overview

What it is: A speech neuroprosthesis built on high-density surface recordings of the speech cortex, in one clinical-trial participant with severe limb and vocal paralysis.

What was shown: Real-time decoding across three outputs: text, speech audio and facial-avatar animation. Models were trained on neural data collected as the participant attempted to silently speak sentences.

• Text: median 78 words per minute, median word error rate 25%, with a large vocabulary.
• Speech audio: intelligible, rapid synthesis, personalized to the participant’s pre-injury voice.
• Avatar: virtual orofacial movements for speech and non-speech communicative gestures.
• The decoders reached high performance with less than two weeks of training.

Limits: One participant. The abstract does not give the electrode count or array geometry, so none is stated here. An Author Correction was published on 4 July 2024.

Spec Card Grid

Identity

• Paper: “A high-performance neuroprosthesis for speech decoding and avatar control,” Nature, 23 August 2023
• Participant: one person, severe limb and vocal paralysis

Results (as reported)

• Text rate: 78 words per minute (median)
• Word error rate: 25% (median)
• Training time: under two weeks
• Outputs: text, speech audio, facial avatar

Evidence and limits

• Not covered here: array dimensions and channel count; no 3D model yet
• Compare with: intracortical speech work such as the Paradromics trial

References

• A high-performance neuroprosthesis for speech decoding and avatar control. Nature, 23 Aug 2023; Author Correction 4 Jul 2024. https://www.nature.com/articles/s41586-023-06443-4