Free tools Windows power users keep installed
One-click scans. No signup required.
The reported change “from 50% to 23.5%” cannot be verified from the available evidence. The closest matching primary study, published in 2023, reported a 23.8% word error rate (WER) with a 125,000-word vocabulary—not 23.5%—and 9.1% with a 50-word vocabulary. Those are results from one participant using a particular implanted speech-decoding system, not proof of a general decoder rebuild from a 50% baseline.
What does a brain-to-text decoder do?
A speech neuroprosthesis can record neural activity while a person attempts to speak, then translate patterns in that activity into text. In the 2023 intracortical study closest to the headline, implanted microelectrode arrays recorded activity; a neural decoder mapped it to phonemes, and a language model helped produce words. The reported word error rate therefore measures the output of a combined pipeline, not the neural decoder alone.
As an Amazon Associate I earn from qualifying purchases.
That distinction matters when describing a “rebuild.” Changing the neural decoder, language model, vocabulary, or evaluation conditions could change the final text score. The evidence here does not identify a specific system rebuilt from 50% to 23.5%, or establish what was changed to produce that exact improvement.
What does word error rate measure?
WER compares a system’s output with the reference transcript. It counts substitutions, deletions, and insertions relative to the number of words in that reference. A lower score means fewer word-level errors under the particular test conditions; it does not mean the same percentage of sentences will be correct.
#1 Best Overall
WER also depends on the task. A restricted vocabulary can make word prediction easier than a much larger one, and scores from different participants, recording methods, datasets, or evaluation procedures should not be treated as a direct ranking.
What results have been reported?
| Study and recording approach | Vocabulary or benchmark | Reported result | How to interpret it |
|---|---|---|---|
| Stanford-led intracortical speech neuroprosthesis study, 2023; one participant with ALS | 125,000-word vocabulary | 23.8% WER; 62 words per minute reported for the system | The closest match to the headline’s 23.5% figure, but it is 23.8%, and the study did not establish a 50%-to-23.5% rebuild. |
| Same 2023 intracortical study | 50-word vocabulary | 9.1% WER | A separate, smaller-vocabulary condition—not a 50% baseline. |
| Separate high-density surface electrocorticography (ECoG) study, 2023; participant with severe limb and vocal paralysis | 1,024-word vocabulary | Median WER of 25%; median 78 words per minute | A different recording method and study. Its score cannot be compared with the intracortical result on percentage alone. |
| Rapid-calibration speech neuroprosthesis paper, 2024, describing training for the earlier intracortical system | Earlier system associated with the 23.8% result | 16.8 hours of neural data collected over 15 days | Training burden is important context for the earlier result; this is not the calibration requirement of every system. |
| Context-aware decoding study, 2024 | Brain-to-Text 2024 benchmark with a fine-tuned large language model | 5.77% WER | A distinct benchmark and method, not an implanted neuroprosthesis result and not confirmation of the headline’s figures. |
| Intracortical multi-user transformer study, July 2026 preprint | Pooled model across participants | Reported relative WER improvement across participants; a single directly comparable score is not stated here | A preprint, not established here as peer-reviewed; it does not substantiate the claimed 50%-to-23.5% change. |
Why the headline numbers do not establish a rebuild
The 2023 intracortical study reported 23.8% WER in its 125,000-word condition. The same study’s 50-word condition produced 9.1%, so the number 50 in that result refers to vocabulary size, not a 50% error rate. The evidence does not tie a 50% baseline to the 23.8% result, and it does not verify 23.5%.
Rank #2
To judge a claimed improvement, a reader would need both scores’ evaluation details: the participant and speech condition, recording method, vocabulary, WER calculation, words per minute, calibration data and duration, and whether the text was decoded in real time or reanalyzed offline. Without matched conditions, a lower percentage may reflect a different task rather than a better version of the same decoder.
What can be concluded about progress?
The studies show that attempted speech can be converted into text with measurable accuracy using investigational implanted systems, and that results vary with vocabulary, recording approach, and decoding method. The 23.8% result is a specific research outcome, not a consumer-device specification or a performance guarantee for other participants. The separate 2024 benchmark and 2026 preprint point to other lines of work, but their figures should remain attached to their own methods and evaluation settings.
Quick Recap
Best Value
- Learn about your brainwaves, train your meditation, and develop your own applications with the mindwave mobile wireless headset.
- Bt/ble Dual mode module and support iOS, Android, PC, and Mac platform. Detects raw-brainwaves, eeg power spectrums (Alpha, beta, etc.), esense meters for attention, meditation, and future algorithms.
- More than 100 brain training games and educational apps available from the NeuroSky online store. Uses a single AAA battery (not included) for 8-hour battery run time
Rank #4
Rank #3
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
