Next-Generation Neural Mass Models Reproduce Features of Speech Processing

Abstract Segregation of speech into syllables is a key step in neural speech processing. It relies on the alignment of neural activity with the rhythmic structure of speech. Two neurological mechanisms have been suggested: phase-resetting of existing neural oscillations, or evoked responses to acoustic features of the speech signal. From EEG experiment, it is known that a distinct feature of this neural entrainment is that its strength is correlated with phonemes sharpness. Here, we reproduce this EEG experiment in-silico. In particular, we use a bio-physical neural mass model to simulate the neural response to the same near-isochronus, consonant–vowel stimuli as the EEG study. We compare this to the response of two other, phenomenological, models that are designed to represent the proposed mechanisms: one a phase-resetting oscillator, the other an evoked-response filter. The biophysical neural mass model captures population-level oscillations without being tailored to speech. All three models can reproduce the correlation between phoneme sharpness and entrainment strength. While both the phase-resetting and evoked models succeed in reproducing the experimental result, the phase-resetting model is arguably more parsimonious since it does not require the preprocessing the evoked model needs to convert a sound envelope into features. Interestingly, the neural mass model also succeeded, without being tailored to the stimulus, unlike the other two models. For example, it does not respond at a specific stimulus-targeted frequency and it is robust across different parameter settings. Our results suggest that generic oscillatory dynamics in cortical populations may be sufficient to generate sharpness-dependent entrainment to speech. Significance Statement Syllable segregation is a crucial but challenging task as syllables in natural speech lack clear boundaries. Yet human brains perform this task effortlessly. It is known that speech causes alignment of neural activity to the syllabic rhythm, supporting segregation by predicting syllable timing, but the mechanisms responsible are unknown. We explore whether a biophysical neural mass model can recreate an observed sharpness-dependent tuning of this entrainment, and compare to two phenomenological models. Crucially, the neural mass model recreates the phenomenon with little tuning, demonstrating that generic cortical circuit dynamics are sufficient to capture sharpness-dependent entrainment. This biophysical model offers a bridge between the neurobiology of the auditory cortex and the emergent dynamics of speech processing that phenomenological models cannot provide. Competing Interest Statement The authors have declared no competing interest. Footnotes Conflict of interest statement: The authors declare no competing interests. https://github.com/Modelling-Speech-Processing/Shannon_et_al