Inner Speech in Motor Cortex and Implications for Speech Neuroprostheses | Cell 

Past work has decoded attempted speech from the motor cortex, and now we want to know if we can decode inner speech for both easier-to-use BCIs and evaluating privacy concerns (inner speech decoded). In their first experiment with seven words, they find that (1) inner speech, reading, and listening are scaled-down versions of attempted speech (it can leak, and it is in the motor cortex), (2) there is a separable intent axis that allows for classifying inner vs attempted speech. In their second experiment with a 125k vocab and adopting previous RNNs for signal -> phonemes -> speech, they develop a usable, preferred, but much less accurate (26-54% WER as opposed to 2.5%) realtime decoding product.

Found it very difficult to read, but the scientific findings are extremely interesting once understood.