Neural dynamics of between-speaker convergence during speech conversation: a dual-EEG study
Résumé
When two people engage in a conversation, their speech tend to converge towards each other. Little is known about the neural mechanisms contributing to this phenomenon. We used the Word Domino task where two speakers take turns in chaining bi-syllabic words according to a rhyming rule: the first syllable has to rhyme with the last syllable of the previous word. We developed a robust automatic method to extract quantitative indexes of convergence at the single word pair, with minimal a-priori hypothesis, from Mel-frequency cepstral coefficients (MFCCs). A data driven, text independent, automatic speaker identification technique, based on GMM-UBM (Gaussian Mixture Model-Universal Background Model) was used to extract convergence. The Gaussian components model the underlying broad phonetic features that characterize a speaker's voice. Each speaker dependent models was tested on speech produced by their conversational partners. Model goodness of fit define the convergence index. Dual-EEG (Electroencephalography) was recorded to search for the neural underpinnings of speech convergence. This was done by splitting the neural signals of both the speaker's (-500ms -- 0ms before speech onset) and listener's (0ms -- 500ms during listing) brain into convergent and non-convergent epochs. Convergence induced significant desynchronization in the speaking brain in the 12-14 Hz band -270ms to -190ms before speaking onset, in a fronto-central anterior scalp region (F5 FC5 FT7 F7). The listening brain showed desynchronization in the high-beta band (24-29 Hz; 32-34 Hz) with a left (F3 F5 F7 FC5 FC3 C3 C5) and right (AF8 F6 F8 FT8 FC6) fronto-central scalp topography between 50ms and 100ms after listening onset. This results demonstrate that low and high beta oscillatory dynamics contribute to the emergence of speech convergence. Topography suggests the involvement of partially lateralized fronto-central regions possibly including Broca's area. Our hyperscanning approach offers insight on the inter-brain neural dynamics at play in natural conversation.