How Fast Can a Brain-Computer Interface Decode Speech?

By neuraspeak editorial · Updated 2026-10-07 · 4 min read

The fastest peer-reviewed speech brain-computer interfaces decode attempted speech at roughly 60 to 80 words per minute, up from about 15 words per minute in 2021. Natural conversation runs at about 160 words per minute. The best systems now also reach about 97% word accuracy or better on large vocabularies, and a 2026 study documented one user averaging 56 words per minute at home over nearly two years.

How fast are the best speech BCIs today?

The top peer-reviewed speeds are 62 and 78 words per minute (wpm), both reported in Nature in August 2023. Those numbers come from structured tasks, mostly copying sentences shown on a screen. They are not open-ended conversation, and each study involved a single participant.

Speed figures need context. Words per minute depends on the task (copying sentences or talking freely), on the vocabulary size, and on whether the decoder may finish a sentence before showing it. A system can also be fast but inaccurate, so speed always has to be read alongside the word error rate, the share of words decoded wrongly. Natural English conversation runs at about 160 wpm, the reference point used in the 2023 Stanford paper.

How fast were the first speech BCIs?

The first direct speech decoder for a person who could not speak ran at about 15 words per minute. In July 2021, Edward Chang's group at UC San Francisco reported in the New England Journal of Medicine (NEJM) on a man who had been paralyzed for 16 years by a brainstem stroke. He is known publicly as Pancho. A high-density electrode grid on his speech motor cortex let him produce sentences from a 50-word vocabulary at a median 15.2 wpm, with a median word error rate of 25.6%.

For comparison, the fastest BCI text entry at the time came from a different approach. In a 2021 Nature study, a Stanford participant imagined handwriting letters and reached about 90 characters per minute, roughly 18 wpm.

What changed in 2023?

In 2023, two papers published in Nature on the same day, August 23, moved speech BCIs to large vocabularies at several times the earlier speed. At Stanford, Willett and colleagues implanted microelectrode arrays in Pat Bennett, who has ALS. She reached 62 wpm, with a 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary. The authors described this as 3.4 times the previous record.

At UCSF, Metzger and colleagues used a 253-electrode surface grid in Ann Johnson, a woman paralyzed by a brainstem stroke. Her decoder produced text at a median 78 wpm with a median word error rate of 25%. It also generated synthesized speech in a voice personalized to her pre-injury voice and animated a digital avatar's face. The decoders reached this level with less than two weeks of training.

Both systems got roughly one word in four wrong on large vocabularies, so accuracy became the next problem to solve.

How accurate are speech BCIs now?

The best published systems now get roughly 97% or more of words right on large vocabularies. In August 2024, Card and colleagues at UC Davis reported in NEJM on Casey Harrell, a man with ALS, who had four microelectrode arrays with 256 electrodes in total. On the first day of use, 25 days after surgery, the system reached 99.6% accuracy on a 50-word vocabulary after 30 minutes of calibration. On day two, it reached 90.2% on a 125,000-word vocabulary. With continued training, it kept 97.5% accuracy over 8.4 months.

In self-paced conversation, his rate was about 32 wpm, slower than the 2023 copy-task records. This is a reminder that real conversation, with pauses to think, runs slower than benchmark tasks.

How close are speech BCIs to real-time conversation?

In 2025, researchers cut the delay between trying to speak and hearing a voice from several seconds to well under a second. Earlier systems decoded a whole sentence and then spoke it. In March 2025, Littlejohn and colleagues reported in Nature Neuroscience a UCSF and UC Berkeley system that streamed speech in 80-millisecond increments. It produced the first sound within about a second of the attempt to speak, at roughly 47 wpm on average. UC Berkeley reported that the earlier version of the system had a delay of about eight seconds.

In June 2025, a UC Davis study in Nature reported a system that synthesized voice directly from intracortical signals in about 25 milliseconds. The participant could change his intonation and sing short melodies. Listeners understood almost 60% of the synthesized words, compared with 4% of his unassisted speech. Direct voice synthesis is less accurate than text decoding, but it brings a BCI closer to the natural back-and-forth of conversation.

What did 2026 add?

The main 2026 result was evidence that speech BCIs can stay fast over years of daily home use. A study in Nature Medicine, published June 15, 2026, followed the UC Davis participant for nearly two years of near-daily, independent use. He used the system for more than 3,800 hours and produced about 2 million words at an average of 56 wpm. The publisher's summary reports word accuracy above 99% on a vocabulary of more than 125,000 words, with 92% of sentences rated at least mostly correct. He also used the BCI for computer control and to keep working full time.

In March 2026, a BrainGate study in Cell Reports showed that intended speech could be decoded well above chance in a person with advanced ALS, long-standing anarthria and locked-in syndrome who depends on a ventilator. That matters because earlier participants could still make some speech sounds or movements. Neuralink began a speech-focused early feasibility study, VOICE, in October 2025, but as of this writing it had not published peer-reviewed speed or accuracy data.

The bottom line

Speech BCIs went from 15 words per minute on a 50-word vocabulary in 2021 to 62–78 wpm on large vocabularies in 2023, and then to about 97% or better accuracy and sub-second voice output. All of the leading results come from single participants in research studies, and no system yet matches the roughly 160 wpm of natural conversation.

Questions

Glossary: Speech neuroprosthesisPhoneme decodingNeural decoderElectrocorticography (ECoG)Utah array

Sources

1. Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria (Moses et al., NEJM) · 2021-07-15
2. A high-performance speech neuroprosthesis (Willett et al., Nature) · 2023-08-23
3. A high-performance neuroprosthesis for speech decoding and avatar control (Metzger et al., Nature) · 2023-08-23
4. An Accurate and Rapidly Calibrating Speech Neuroprosthesis (Card et al., NEJM) · 2024-08-15
5. A streaming brain-to-voice neuroprosthesis to restore naturalistic communication (Littlejohn et al., Nature Neuroscience) · 2025
6. An instantaneous voice-synthesis neuroprosthesis (Wairagkar et al., Nature) · 2025
7. Long-term independent use of an intracortical brain-computer interface for speech and cursor control (Nature Medicine) · 2026-06-15
8. Brain implant restores communication independence at home (Nature Medicine press release, Nature Asia) · 2026-06-16

Facts on this page are drawn from the sources listed and were accurate to the best of our knowledge on the update date shown. Company and product names are trademarks of their owners and are used for identification only. Nothing here is medical, investment or legal advice. Companies and individuals can request a correction via the corrections page.

Read next