Speech
A high-performance neuroprosthesis for speech decoding and avatar control
- Authors
- Metzger SL … Chang EF15 authors
- Institution
- University of California, San Francisco; University of California, Berkeley
- Publication
- NatureNature 2023;620(7976):1037–1046 · 23 Aug 2023
- Status
- Primary source
The 60-second view
- What problem?
- Naturalistic speed and expressivity had been out of reach for speech neuroprostheses.
- What did they do?
- Used high-density ECoG over speech cortex in a participant with severe limb and vocal paralysis to decode three outputs at once: text, speech audio and facial-avatar animation.
- What changed?
- Matched intracortical speed with a surface array, and added voice and face as output channels.
- Why does it matter?
- Published alongside the intracortical result, it showed two different interfaces converging on usable speech in the same month.
Method
- Participants
- One person with severe limb and vocal paralysis
- Interface
- High-density electrocorticography array over speech sensorimotor cortex
- Signal
- Cortical surface field potentials
- Task
- Attempted silent speech of sentences
- Decoder
- Deep-learning models for text, audio and avatar
- Training
- Less than two weeks
Results
- 78 words/min
- Median decoding rate
- 25%
- Median word error rate
Strength of evidencePeer reviewed, one participant. Decoders reached high performance with less than two weeks of training.
Figures are quoted from the published abstract and were checked on 3 Oct 2026.
Limitations
- One participant.
- Wired research system.
What this could enable
Embodied communication, and a credible surface-array route to speech that does not penetrate cortex.
Primary source
- 01
Metzger SL et al. Nature 2023;620(7976):1037–1046 (opens in a new tab)
Peer reviewedNature23 Aug 2023Checked against source 3 Oct 2026
DOI 10.1038/s41586-023-06443-4