Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks

Phonemic decoding of speech articulation using stereo-electroencephalography 

Poster B94 in Poster Session B, Tuesday, October 24, 3:30 - 5:15 pm CEST, Espace Vieux-Port

Aditya Singh1, Tessy Thomas1, Xavier Scherschligt1, Nitin Tandon1; 1UT Health

Recent research has shown that brain-computer interfaces (BCIs) can be used to decode speech from neural activity, with the goal of aiding patients with speech impairments, such as dysarthria or aphasia. Stereo-electroencephalography (sEEG) has emerged as a means of achieving widespread coverage across multiple regions of the language network with minimal surgical risks. In this study, we employed the articulation of tongue twisters to demonstrate the use of sEEG to map and decode speech. This also provides novels ways to evince unique neural signatures of the speech network, going beyond extracranial localization and surface-based measures derived from subdural grid arrays, contributing to a refinement of existing language models. 25 patients, with sEEG electrodes implanted for seizure onset localization, participated in a tongue twister production task. Each tongue twister comprised four words with matched phonological onsets, systematically increasing articulatory complexity (e.g. just rum rug jump). Patients silently read the words, followed by two rounds of reading aloud and two rounds of memory-based production. Phoneme decoding was performed using linear classification models trained on broadband gamma activity (BGA: 70-150 Hz) recorded from the sEEG electrodes during reading and memory trials. The training and testing samples consisted of a 600ms BGA-window centered at the onset of articulation for each phoneme, and accuracy was evaluated using 10-fold cross-validation. Covert trials were separately decoded using an encoder-decoder model with a connectionist temporal classification loss function. Decoding was performed either during articulation or a pre-articulatory period to evaluate the predictive properties encoded in the latent neural information. We achieved remarkable decoding accuracies, with a peak accuracy of 25.2% during the articulatory window and 22.6% during the pre-articulatory window in the best-performing patient (chance accuracy at 7.7%). Across all 25 patients, the average accuracy was 13.3% +/- 4.4% during articulation, and 11.6% +/- 5% during pre-articulatory periods, surpassing chance accuracy of 5.1% +/- 1%. Notably, these decoding results consistently outperformed those obtained through ECoG and MEA-based decoding studies using similar linear models, while still being considerably safer in terms of intra-operative neural trauma. Our findings revealed activation in posterior middle temporal gyrus, dorsal frontal cortex, inferior frontal gyrus, and superior parietal cortex. Moreover, we observed distinct separation in the latent articulatory space for manner and place of articulation, particularly within superior temporal gyrus and sensorimotor cortex. These regions have been associated with predictive encoding and monitoring of articulation, notably emphasized during tasks which have a larger stress on generating phonologically correct sounds when given stimuli with higher phonological load and lexical bias. Spatial decoding patterns from pre-articulatory models validated these predictive encoding sites, highlighting the presence of information-rich electrodes sparsely distributed across frontal and parietal regions, including deep sulcal sites inaccessible to subdural grids.This study significantly contributes to the development of more precise and high-fidelity assistive communication devices for individuals affected by neurodegenerative disorders that impact speech articulation and production. Furthermore, our sEEG findings refine models of the distributed speech network derived solely from extracranial neuroimaging measures, enhancing our understanding of the underlying neurobiology of language.

Topic Areas: Computational Approaches, Speech Motor Control

SNL Account Login

Forgot Password?
Create an Account