Speech & Language Neuroprostheses

Real-time speech BCI

A speech neuroprosthesis that decodes and outputs during attempted speech with latency low enough to support communication, rather than analyzing recordings offline. This demands a causal pipeline that never uses future samples, bounded compute per time-step, streaming language-model decoding, and closed-loop feedback so the user sees or hears the output and can adapt.

Moses et al. (2021) demonstrated real-time decoding of words and sentences from ECoG in a participant with anarthria; later systems raise words-per-minute and vocabulary while keeping latency near-conversational.

Offline WER computed on the same neural data is an optimistic upper bound on real-time performance, causality, latency limits, and the user's own closed-loop behavior all erode it.

Also called
online speech decoding即時語音解碼