A high-performance speech neuroprosthesis

The 60-second view

What problem?

Speech BCIs had not reached accuracies usable for unconstrained sentences over a large vocabulary.

What did they do?

Recorded spiking activity from intracortical arrays in a participant with ALS and decoded attempted speech to text.

What changed?

Reported the first successful large-vocabulary decoding and more than tripled the previous speed record.

Why does it matter?

It established that a small patch of cortex carries enough articulatory detail for open-vocabulary speech, years after speech was lost.

Method

Participants

One person with ALS who can no longer speak intelligibly

Interface

Intracortical microelectrode arrays in speech motor cortex

Signal

Intracortical spiking activity

Task

Attempted speech of prompted sentences

Decoder

Recurrent neural network to phonemes, with a language model

Results

9.1%

Word error rate, 50-word vocabulary

23.8%

Word error rate, 125,000-word vocabulary

62 words/min

Decoding rate

Natural conversation is about 160

Strength of evidencePeer reviewed, one participant. The underlying data were released and now anchor a public benchmark.

Figures are quoted from the published abstract and were checked on 3 Oct 2026.

Limitations

• One participant.
• Roughly one word in four was wrong at the large vocabulary size.

What this could enable

A shared dataset that other groups have used to cut word error rates further without new surgery.