Multi-User Speech BCI Model Needs Fewer Than 200 Sentences for a New User
Summary
UC Davis researchers trained a transformer-based speech decoder across six people with intracortical BCIs. The pooled model cut relative word error rates by more than 50% on average compared with subject-only models and, after fine-tuning on fewer than 200 sentences from an unseen user, achieved a word error rate below 7%. The bioRxiv preprint has not been peer reviewed.
Why it matters
Reducing the amount of per-patient calibration data addresses one of the main barriers to scaling speech neuroprostheses and suggests that useful neural structure can transfer across users.
BCIwiki (bciwiki.com) – A speech decoder trained jointly across 6 participants reached a word error rate below 7% on a user it had never seen after finetuning on fewer than 200 sentences, according to a preprint posted to bioRxiv on July 27, 2026, ahead of peer review.Intracortical brain-computer interfaces can restore communication to people with vocal tract paralysis by decoding cortical activity during attempted speech into text. The best current systems, which pair a neural-to-phoneme decoder with a phoneme-to-word language model, have reached word error rates as low as 1%. That figure carries a condition: thousands of sentences of training data must first be collected from the user. For someone who has just received an implant, that means a long stretch between surgery and fluent communication, and it is a practical obstacle to scaling the technology.The team at the University of California, Davis pooled data across people instead. Their transformer-based decoder was trained jointly on 6 intracortical speech BCI participants. For every participant, regardless of sex, disease etiology or attempted speaking strategy, the multi-user model decoded more accurately than a model trained on that individual’s data alone, with relative word error rate more than 50% lower on average. The finding is notable in itself: across different etiologies and different speaking strategies, the neural activity appears to carry structure that can be shared.The held-out result carries more practical weight. Applied to a user who had contributed nothing to training, the multi-user model needed fewer than 200 sentences from that person to finetune, bringing word error rate below 7%. That figure is not in the same class as the 1% cited earlier: 1% is the best result after thousands of training sentences, while a rate below 7% is a usable level reached from fewer than 200, soon after implant. The authors read this as a route to pooling intracortical data across people for decoding models that are more accurate, more generalisable and faster to deploy. As a preprint, the work awaits peer review and larger-scale confirmation.
Compiled by BCIwiki from public sources
