Decoding speech intent from non-frontal cortical areas

Brain–computer interfaces (BCIs) for speech restoration have largely relied on neural signals recorded from frontal speech motor regions, including ventral precentral and postcentral gyri and inferior frontal gyrus. In contrast, temporal and parietal cortices are classically associated with speech and language perception, and evidence for speech production signals in these regions remains limited. Identifying such signals could enable speech BCIs for individuals with frontal lobe damage, including expressive aphasia.

We investigated whether temporal and parietal cortices encode a decodable signal related to speech production using electrocorticography (ECoG) recordings from four participants undergoing surgical treatment for epilepsy or brain tumors. Standard clinical ECoG arrays (10-mm spacing) or high-density mini-ECoG arrays (4-mm spacing) were placed over temporal and/or parietal cortices. Participants performed a delayed speech task in which words were read silently, held in memory, and then spoken aloud, allowing dissociation of neural activity related to speech production from reading and comprehension.

High-gamma (70–300 Hz) power was extracted in 100-ms windows and used to classify speech versus silence using a linear support vector machine with strictly causal features. Across participants, speech intent was decoded with accuracies ranging from 67.2% to 80% (p < 0.03, permutation test). Demixed principal component analysis further revealed a low-dimensional separation between speech and silence states emerging up to 500 ms prior to speech onset.

These results provide evidence for a speech production–related signal in temporal and parietal cortices, supporting emerging models that implicate posterior temporal regions in speech planning. This finding expands the set of cortical areas relevant for speech BCIs and has implications for restoring communication in individuals with language and motor speech disorders.