How does the human brain recognize a word using time-yoked auditory computations? We’re working on that now.

We replicate this result using naturally fast and slow speech.

We applied the TCI method to precise human intracranial recordings. We find that integration windows are yoked to absolute time everywhere we look in the auditory cortex, including non-primary regions of the superior temporal gyrus.

We show this approach can distinguish time- vs. structure-yoked integration from computational models. For example, we uncover a transition from time- to structure-yoked integration across the layers of a deep neural network model trained to transcribe natural speech.

Integration windows were measured using the temporal context invariance (TCI) method, which is applicable to virtually any sensory response, including highly nonlinear systems like the brain. TCI method estimates the smallest segment yielding a context-invariant response.

We measured integration windows after time-compressing and stretching speech.

Structure-yoked prediction: integration window scales with speech rate
Time-yoked prediction: integration window is constant across different speech rates

Auditory and cognitive models often cast neural integration in terms of time vs. structure, respectively. Time- and structure-yoked integration are distinct because speech structures (e.g., phonemes) have highly variable durations.

Human auditory cortex integrates information in speech across absolute time (e.g., 200 ms), not phonemes, syllables, words, or any other time-varying speech structure: www.nature.com/articles/s41...