Phoneme decoding

Taking the roughly 40 phonemes of a language (or a set of phonetic features) as the decoding target, so the neural model outputs a sequence of phoneme probabilities over time that a language model then assembles into words. Phonemes form a compact, language-covering alphabet: a decoder that generalizes across about 40 units can in principle produce any word, including out-of-vocabulary ones, unlike a fixed word classifier.

Willett et al. (2023) decoded attempted-speech phonemes from two small intracortical arrays in area 6v of a participant with ALS and, combined with a language model, reached about 62 words per minute over a 125,000-word vocabulary, with word error rates on the order of 24 percent, and far lower on a restricted 50-word set.