This was a team effort led by Pavo Orepić @pavoorepic.bsky.social, with Wilson Truccolo, Eric Halgren, Sydney Cash, Anne-Lise Giraud @annelisegiraud.bsky.social, and Timothée Proix. Our deepest thanks go to the participants, and to the reviewers who helped us improve the paper!

Our interpretation: the brain makes a top-down guess about the word, and the population's phoneme code realigns to match it, as analysis-by-synthesis models predict. This was three participants and a 4×4 mm patch of cortex, so details vary between them.

We found a top-down alignment in this anterior superior temporal gyrus area: the representation for the 1st, 2nd, 3rd... phoneme peaks at the same moment as the meaning, at a fixed delay after the word starts.

How do the two codes interact? In the bottom-up hypothesis, phonetic representations build up towards a lexical one, and each phoneme is encoded a fixed time after it is heard. In the top-down case, lexical drives, and all phonemes are encoded at the same time or after the word representation.

These results were reproduced during natural conversation, with a different speaker and different recorded neurons. The manifold stays stable even when the individual neurons change.

Word meaning is also encoded in a lexical neural manifold. The phonetic and lexical manifolds partly overlap: they share at least one main axis of activity but also keep their own dimensions. So the two codes live in the same neuronal population and can interact.

Only a few single neurons responded clearly when taken in isolation, and nearby ECoG did not show any specific response. The information actually lies in the neural population: its activity unfolds in a low-dimensional neural manifold, where phonemes trace paths grouped by phonetic categories.

We recorded neurons with Utah electrode arrays in the temporal lobe of three people undergoing epilepsy monitoring. They judged whether words describing animals or objects, like "chair", named something bigger than a foot, and also listened to natural speech.

Speech is often described as a progression from sounds to meaning. But meaning also changes what we hear, e.g. the surrounding word can change how we hear an ambiguous sound. Where and how do these two levels meet in the brain?

Excited to share our new paper in Nature Communications! 🎉 Recording neurons in human temporal cortex, we found that speech sounds and word meaning are both encoded in neural manifolds, and that speech sounds line up in time with word meaning, a signature of top-down prediction. 🧵