Learning discrete neural latent spaces for high-performance speech decoding.

Abstract

Speech brain-computer interfaces (BCIs), which directly transform neural signals into intelligible voices, offer a promising avenue for people with aphasia. To decode speech information from brain signals, neural representation learning plays an important role. Existing studies mainly explored continuous neural latent spaces for speech decoding and ignored the intrinsic discrete property in speech production. Here, we propose to learn a discrete neural latent spaces by constructing a quantized representation learning network for speech decoding. Experiments with intracranial stereotactic EEG (sEEG) signals from 11 subjects demonstrated that our approach significantly improved the precision and robustness of speech decoding. These results underscore the potential of our method to improve the functionality and usability of speech BCIs for people with aphasia.