Machines Can Produce What We Speak in the Brain: A Linguistic Framework for Decoding Inner Speech

When we rehearse a sentence silently, such as ‘Hi, what’s up?’ before greeting a friend, the brain runs much of the machinery of speaking without moving a muscle. This paper asks, from a linguistic perspective, how a machine could turn that silent speech into words. We propose a framework in two parts. The first is a measurement programme that moves from overt to covert speech in controlled steps. Phonemes are first mapped while they are spoken aloud (data-A). Articulatory and laryngeal layers are then removed one at a time to isolate the neural correlates of individual distinctive features (data-B). Finally, the same items are only imagined (data-C). The second part is an inference cascade. Noisy, phoneme-level neural evidence is resolved by successive layers of linguistic knowledge: phonotactics, prosodic timing, the lexicon, morphosyntax, semantics, and discourse. A worked example shows how the fragmentary signal H..y, w…tz ..p converges on a single utterance. We argue that the phoneme is the right unit for decoding because of duality of patterning: a small, closed inventory generates an open-ended vocabulary. This architecture converges with that of current speech neuroprostheses, which pair phoneme decoders with language models and have recently begun to decode inner speech. The linguistic analysis adds three cautions that such systems rarely state. Inner speech may be condensed rather than merely noisy. A strong language model can end up writing the speaker’s sentence for them. And only inner speech addressed to someone should be decoded at all. I am open to reviews or comments on this work, as well as collaborations.