BCIwiki (bciwiki.com) — Swapping a multi-billion-parameter language model for Jev, a hosted model, to rescore a speech neuroprosthesis's candidate sentences lowered word error rate to 7.5% on 978 held-out sentences from a participant with ALS, against 7.8% for both OPT-6.7b and Qwen2.5-7B, according to a preprint by Gabriele Cinà posted to arXiv on September 27, 2026. The study has not been peer reviewed.
A speech neuroprosthesis decodes attempted speech from brain activity and ends by rescoring the decoder's candidate sentences with a large language model, the only component that needs a GPU. The author notes that general language models asked to pick one sentence from a list tend to answer from a label's position in the list rather than from the sentence itself. The study instead frames rescoring as a single typed decision: one call returns a probability for every candidate, which is then combined with the decoder's own score.
The published decoder alone reaches 8.1% word error. With the decoder's weight re-tuned, Jev reaches 6.9%, versus 7.2% for OPT-6.7b and 7.4% for Qwen2.5-7B. The author reports Jev is ahead in all 4 comparisons and at most 0.2 points behind at the 95% bound.
Jev costs 0.07 USD per thousand sentences and needs no GPU. By the author's estimate, a dedicated GPU running a 7B model is cheaper per sentence only above 43% utilisation, far beyond what one user generates. End-to-end latency over the internet is 262 ms, 62 ms of it at the provider, the same order as a 7B model on a local GPU (27 ms) but not faster.