Abstract

Several labs have recently developed relative weights procedures for estimating crucial time-frequency regions of the speech spectrogram. Due to the large degree of acoustic variability in natural speech, these spectrogram-based methods work best with rather limited sets of brief (e.g., syllable-length) utterances. We have developed a procedure that measures relative weights on the modulation power spectrum (MPS), which is obtained from the 2D Fourier transform of the speech spectrogram. Modulation patterns are relatively stable across utterances, which allows measurement of relative weights in the MPS domain using more ecologically-valid sets of continuous utterances (e.g., sentences). Our procedure works by applying randomly-shaped filters to the MPS over many trials and relating the filter patterns to an outcome measure using reverse correlation. Here, we estimate relative weights in the MPS domain for two outcome measures: keyword recognition/intelligibility (perceptual weights) and physiological activity measured with blood-oxygen-level-dependent fMRI (neural weights). Perceptual weights indicate that a circumscribed region of the MPS supports intelligible speech perception. Patterns of neural weights differ across regions of the auditory cortex—namely, primary auditory areas encode a broad range of the MPS while downstream regions converge toward a circumscribed representation that strongly resembles the pattern observed in perceptual weights.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call