Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Abstract lexical information is available in speech-motor cortex during conversational speech
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Jeremy Yeaton1, Zachery Fogg1, Nicholas Card1, Carrina Iacobacci1, Tiffany Brailow1, David Brandman1, Sergey Stavisky1; 1University of California, Davis
Introduction: Speech motor cortex has traditionally been understood as a substrate for the sensorimotor planning and execution of articulation. Yet its potential role in encoding higher-level linguistic structure — including lexical-syntactic category and semantic content — remains poorly characterized. Brain-computer interface (BCI) systems that record from this region offer a rare opportunity to examine action potential-resolution neural representations of language during naturalistic behavior. Here we ask whether part of speech (POS) and semantic information are decodable from chronic microelectrode array recordings in speech motor cortex, and whether such representations are stable across distinct speech contexts. Methods: We analyzed chronic intracranial recordings from four microelectrode arrays implanted in the precentral gyrus of a man with ALS. Neural data were collected during two speech conditions: unconstrained conversational speech (34,430 utterances) and cued sentence production (7,849 utterances). Because the participant minimally vocalizes during speech production with the BCI, word alignment was based on the timing of the decoded phonemes. For each word, we extracted spike-band power in 80ms bins and trained recurrent neural networks (RNNs) to decode POS (noun, verb, adjective, adverb, proper noun) and semantic properties captured via word embeddings. The top 3 most frequent tokens from each POS class were excluded during training. All models were trained on the conversational data, and tested on held-out in-vocabulary conversational, out of vocabulary conversational, and cued production. Averaged phoneme information was regressed out prior to analyses. Decoding performance was evaluated using cross-validated accuracy and compared across the two speech contexts. Results: An RNN trained on neural data from -2000ms to 500ms about word onset was able to decode POS correctly for 54.8% of in-vocabulary tokens (chance = 20%), and 49.5% for out of vocabulary tokens. This performance dropped to 36.2% when tested on cued production. Verbs had the highest decoding accuracy, followed by nouns. Peak decoding accuracy was ~1500ms before the first decoded phoneme. In a temporal generalization analysis, POS category representations were relatively stable from -2500ms until word onset. In separate analyses, we found better-than-chance correlations between decoded representations and word embedding vectors from around -2500ms to 1000ms in the conversational condition, but not the cued production condition. Conclusions: We found decoding accuracy was substantially higher during conversational speech than during cued sentence production for both lexical-syntactic and lexical-semantic decoding, despite equivalent low-level motor demands across conditions. This asymmetry was consistent across part-of-speech classes and semantic dimensions. These findings demonstrate that human speech motor cortex carries decodable higher-order linguistic information beyond articulation, and that the structure of these representations is not fixed but rather scales with the naturalness of the speech context. We interpret the advantage for conversational speech as reflecting richer top-down linguistic context during spontaneous language use, which may stabilize or sharpen lexical and semantic representations. These results challenge a strictly motor view of the precentral gyrus and suggest that speech motor cortex participates in broader language processing in a context-dependent manner.
Topic Areas: Language Production, Speech Motor Control