Semantic Representations of Speech Production in Intracranial EEG

Speech neurprostheses have the potential to provide severely paralyzed patients with a means of communication. To enable the best possible decoding of speech processes from neural data, it is important to chose a representation of speech that is both meaningful to the decoding process and represented well in the neural recordings. Previously, acoustic, articulatory, and textual representations of speech have been decoded from neural recordings. Semantic representations of speech could add additional information about the content of the produced speech. In this study, we show that semantic embeddings for individual words, as extracted by a word2vec-model, can be used to reconstruct neural activity during speech production across wide-spread cortical and subcortical areas. We elucidate the temporal dynamics of reconstruction quality and show that a slight right hemisphere preference exists. These findings could be used to add semantic information into speech neuroprostheses in the future.