Contextual Prediction Tunes the Tempo of Speech Segmentation

Real-time speech comprehension draws on two mechanisms usually studied apart: bottom-up parsing of syllable-sized units by theta-range neural activity, and top-down prediction of next tokens based on context. Combining behavior and neurocomputational modeling, we asked how context informs syllable parsing. Human listeners (n=110) transcribed spoken sentences with altered temporal structure. Sentences were time-compressed threefold, then spread out again by interspacing silences. In a first experiment, silent gaps were either placed at natural syllable boundaries or interspaced with arbitrarily segmented speech chunks at fixed intervals. In a second experiment, gaps were placed at syllable boundaries and syllables delivered either at their natural pace (quasi-periodic) or at fixed intervals (periodic). Recognition peaked at presentation rates between 8 and 13 Hz, above the canonical theta-syllable range, and alignment to the syllabic structure outperformed time-based segmentation. The contextual predictability of each word, estimated by a language model, predicted recognition only when natural syllable boundaries and timing were preserved. This sensitivity to predictability was reproduced in a computational model where word-to-syllable prediction is gated by a (top-down) beta rhythm. Speech can thus be understood under heavy temporal degradation, and beyond the canonical theta-syllable range, if syllable timing naturalness allows context to be used.