The Speech-BCI Scoreboard: Why the Field's Headline Accuracies Don't Compare

Speech-implant accuracy records are being set on different tests, and the scores cannot be ranked against each other. The 99.6 percent accuracy reported in the New England Journal of Medicine in August 2024, described in a statement by co-senior author David Brandman of the University of California, Davis as "the most accurate speech neuroprosthesis ever reported," was scored on a vocabulary of 50 words 1 2. The field's open-vocabulary systems decode from a 125,000-word vocabulary 3. That is a 2,500-fold difference in the size of the test, 125,000 divided by 50, and it is why the headline numbers do not compare 4.

This page is ProvenBrief's scoreboard for speech brain-computer interfaces: landmark decoding results held in one maintained table, each row carrying vocabulary size where the study reports one, metric basis, speed, recording hardware, and participant count, scored with open-vocabulary mutual information (OVMI), a measure proposed by the Neural Processing Lab at the University of Oxford in September 2026 to put every row on one scale 4.

Seven headline results, seven different yardsticks

• Moses et al., 2021: a 50-word vocabulary decoded from invasive recordings of attempted speech; conveys 2.4 percent of broad spoken-English information, 4.7 percent with language-model post-processing 4.
• Willett et al., 2023 (Nature, Stanford): a 9.1 percent word error rate on 50 words and 23.8 percent on a 125,000-word vocabulary, decoded at 62 words per minute from intracortical microelectrode arrays in a participant with amyotrophic lateral sclerosis who can no longer speak intelligibly 3.
• Card et al., 2024 (NEJM, UC Davis): 99.6 percent accuracy on 50 words on the first day of use, 90.2 percent on 125,000 words the next day, 97.5 percent sustained beyond eight months; four microelectrode arrays, 256 electrodes, in a man in his 40s with ALS 1 5. A June 2026 follow-up on the same participant and implant, published in Nature Medicine, reports the system sustaining over 99 percent word accuracy on the 125,000-word vocabulary in controlled testing after nearly two years of independent home use 6 7.
• Qian et al., 2025 (Science Advances): real-time Mandarin decoding, reported as a first for tonal languages, from a 256-channel electrocorticography array in a 43-year-old woman undergoing epilepsy monitoring: a median 71.2 percent syllable-identification accuracy in single-character tasks across 394 Mandarin tonal syllables, and 73.1 percent character accuracy with a language model at 49.7 characters per minute on sentences 8 9.
• Jude et al., 2026 (Cell Reports): intracortical decoding of intended speech in a person with long-standing anarthria and locked-in syndrome, evaluated on 125,000-word vocabulary sentences 10.
• Chang et al., 2026 (Nature Neuroscience, UCSF): the first system to decode speech and upper-body gestures at the same time, tested with electrocorticography arrays over the sensorimotor cortex in three participants with vocal tract and bodily paralysis; two of the three drove a full-body virtual avatar, and for one participant the median decoding accuracy for speech and gestures together reached 100 percent across three conversational experiments 11 12 13.
• Zhao et al., 2026 (Nature Neuroscience, Fudan University and ShanghaiTech): imagined speech decoded in nine awake brain-tumor surgery patients, with temporary 256-channel electrocorticography grids and monitoring that confirmed no sound or muscle movement during imagery; the coverage describes it as the first demonstration that imagined speech can be decoded into text, audio and movement trajectories at accuracy rivaling overt speech, and its syllable classifier reached a median 80.4 percent identification accuracy for imagined syllables against 78.3 percent for spoken ones, with chance at 16.7 percent, no word vocabulary, and no words-per-minute rate 14 15.

Every row through 2025 is a single participant; the September 2026 UCSF entry was the first on this scoreboard to carry three, and the row added in this October 2026 refresh carries nine, the most on the page, with a caveat the table keeps attached: all nine were awake surgical patients with temporary electrode grids rather than people with communication impairments, and the study's authors state that validation in chronically implanted patients will be essential 14. The vocabulary column runs from 50 words to 125,000, and the two newest rows report no word count at all: one only conversational tasks, the other syllables rather than words. The scoreboard keeps those facts attached to every number.

The imagined-speech row carries its own yardstick, syllable identification, which the study reports for imagined speech, for overt speech, and against chance.

The Oxford analysis: vocabulary, not the decoder, moved the implanted systems

OVMI turns on a plain identity. The probability that a speech BCI transmits the word a user intends equals the probability that the word is in the decoder's vocabulary at all, times the accuracy within that vocabulary. Conventional metrics capture only the second factor; the Oxford authors show that scores "computed only over the words a system supports can overstate how much of a user's intended speech the system can communicate" 4.

The Stanford 2023 study is the clean demonstration because it tested both configurations in one participant. On the OVMI scale its 50-word system conveys 6.4 percent of the lexical information in broad spoken English, and its 125,000-word system conveys 72.0 percent: about 11 times more intended speech, delivered while the word error rate got worse, not better, rising from 9.1 percent to 23.8 percent 4 3. The Oxford analysis finds the largest historical gain among implanted systems came from the jump to large vocabularies, after in-vocabulary accuracy had already flattened at small sizes; on that common scale it scores the 2024 UC Davis system at 93.7 percent, the highest of the compared systems 4.

The paper's worked example makes the trade plain: a decoder that is perfect on 50 words conveys 0.28 bits against a 1,000-word communication target, while a decoder at 50 percent accuracy covering all 1,000 words conveys 3.98 bits, roughly 14 times more 4. Accuracy ranks the first system higher. Information conveyed ranks the second.

The same system scores differently depending on the conversation

Scores flip with what the user is trying to say. Measured against broad spoken English, the 50-word Willett system conveys 6.4 percent of lexical information; measured against an augmentative and alternative communication distribution, the restricted vocabulary of clinical assistive use, it conveys 40.4 percent. The 2021 Moses system moves from 4.7 percent to 30.7 percent on the same switch 4. A number reported without its target distribution is a property of the test, not of the talk.

The UC Davis arc shows the vocabulary trade inside one study, same participant and implant: 99.6 percent on 50 words after 30 minutes of training data on day one, 90.2 percent on 125,000 words the next day after 1.4 more hours, 97.5 percent sustained past eight months, with word error rates consistently below 5 percent and more than 248 hours of self-paced conversational use 5. The June 2026 follow-up stretches that arc to nearly two years of near-daily independent use: more than 3,800 hours at home, more than 183,000 sentences and close to two million words, an average of 56 words per minute, over 99 percent word accuracy on the 125,000-word vocabulary in controlled testing, and 92 percent of sentences rated accurate or mostly correct by the participant himself 6. Earlier brain-to-text systems needed roughly 16.8 training hours to reach 23.8 percent error on 125,000 words 5 3.

Speed carries the same label problem. Stanford's 62 words per minute is 3.4 times the previous record and approaches natural conversation at 160 words per minute 3; the UC Davis participant's attempted speech reaches 6.8 correct words per minute to expert listeners, and his headmouse typing 6.3 5. The UC Davis system's own reported rate has kept climbing the same ladder of labels: roughly 32 words per minute at the August 2024 report, an average of 56 words per minute by June 2026, more than eight times his unaided 6.8 2 6. The Mandarin system's 49.7 characters per minute stays in its own row because characters are not words, and this scoreboard does not convert units 8. A December 2025 review in the Journal of Intelligent Medicine notes the 62-words-per-minute decoder remains below the pace of natural conversation 16.

What the scoreboard can and cannot tell you

Rows carry only figures the studies report. The OVMI scores are the Oxford paper's, computed with a scalar estimator because studies rarely publish confusion matrices; out-of-vocabulary words count as unsupported, so paraphrase receives no credit; the measure covers lexical information, not grammar or context 4. The authors state that a higher OVMI does not necessarily imply a more practical or clinically useful paradigm: decoding perceived speech, for instance, is on its own unlikely to constitute a usable communication interface 4. The measure also aims forward: selecting a vocabulary to maximize OVMI produced a relative improvement in accuracy of up to 16.3 percent across three speech domains 4. The newest rows, the June 2026 UC Davis follow-up, the September 2026 UCSF entry, and the imagined-speech study added in the October 2026 refresh, carry their studies' reported figures only; this page holds no OVMI scores for them, so they stand next to the scored rows unscaled. The table refreshes as new results publish.