Noise differentially impacts phoneme representations in the auditory and speech motor systems

Abstract

Although it is well accepted that the speech motor system (SMS) is activated during speech perception, the functional role of this activation remains unclear. Here we test the hypothesis that the redundant motor activation contributes to categorical speech perception under adverse listening conditions. In this functional magnetic resonance imaging study, participants identified one of four phoneme tokens (/ba/, /ma/, /da/, or /ta/) under one of six signal-to-noise ratio (SNR) levels (-12, -9, -6, -2, 8 dB, and no noise). Univariate and multivariate pattern analyses were used to determine the role of the SMS during perception of noise-impoverished phonemes. Results revealed a negative correlation between neural activity and perceptual accuracy in the left ventral premotor cortex and Broca's area. More importantly, multivoxel patterns of activity in the left ventral premotor cortex and Broca's area exhibited effective phoneme categorization when SNR ≥ -6 dB. This is in sharp contrast with phoneme discriminability in bilateral auditory cortices and sensorimotor interface areas (e.g., left posterior superior temporal gyrus), which was reliable only when the noise was extremely weak (SNR > 8 dB). Our findings provide strong neuroimaging evidence for a greater robustness of the SMS than auditory regions for categorical speech perception in noise. Under adverse listening conditions, better discriminative activity in the SMS may compensate for loss of specificity in the auditory system via sensorimotor integration.

Keywords: fMRI; forward sensorimotor mapping; multivariate pattern analysis; speech categorization.

Fig. 1.

Behavioral performance and activation elicited by phoneme perception and production. (A) Group mean accuracy and RT across phonemes as a function of SNR. (B) Group mean accuracy and RT for identifying individual phoneme as a function of SNR. NN, no noise condition. The error bars indicate SEM. (C) Activation elicited by phoneme identification without masking noise relative to silent baseline. (D) Activation elicited by subvocal phoneme production relative to rest. Maps are thresholded at FWE-corrected P < 0.01 with a cluster size ≥ 732 mm³ for both perception and production. (E) Conjunction analysis of phoneme perception and subvocal phoneme production.

Fig. 2.

Regions showing significant within-subject correlation between BOLD signal and behavioral accuracy. Maps are thresholded at FWE-corrected P < 0.01 with a cluster size ≥ 342 mm³. Notably, BOLD activity in dorsal and ventral regions exhibited opposite correlations with accuracy. aSTG/aMTG, anterior superior temporal gyrus and anterior middle temporal gyrus; Ins/Broca, insula and Broca’s area; IPL, inferior parietal lobule; MFG, middle frontal gyrus; pMTG, posterior middle temporal gyrus; PMv, ventral premotor cortex; pSTG, posterior superior temporal gyrus.

Fig. 3.

PSI maps and within-subject correlation between the PSI and behavioral accuracy. PSI maps when the phonemes were presented (A) without noise and (B) with noise (average across five SNRs). Maps are thresholded at FWE-corrected P < 0.05 with a cluster size ≥ 293 mm³ (without noise) or 439 mm³ (with noise). (C) Regions that showed a significant within-subject correlation between the PSI and perceptual accuracy (FWE-corrected P < 0.05, clusters ≥ 293 mm³). (D) PSI in two selected ROIs (left IFG and left PMv) that showed a significant correlation between the PSI and accuracy. (E) PSI in bilateral STG ROIs that exhibited significant (FWE-corrected P < 0.05) phoneme-specific encoding under the no noise condition. *P < 0.05, **P < 0.01, by one-sample t tests. NN, no noise condition. The error bars indicate SEM. IFG, inferior frontal gyrus; Ins, insula; IPL, inferior parietal lobule; PMv, ventral premotor cortex; poG, postcentral gyrus; Precu, precuneus; STG, superior temporal gyrus.

Fig. 4.

PSI as a function of SNR in ROIs that showed a significant correlation between BOLD signal and accuracy. (A) Locations of ROIs are displayed on a template brain. ROIs were spheres with a 10-mm radius around the peak coordinates from univariate regression analysis. Peak Talairach coordinates are (–44, 0, 28) for the left PMv, (–40, 20, 5) for the left Ins/Broca, (–45, –32, 9) for the left pSTG, and (–36, –48, 38) for the left IPL. Data from the left insula and Broca’s area (Ins/Broca) were pooled because a sphere with a 10-mm radius centering at the above coordinate occupied a part of both areas. (B) PSI in four ROIs as a function of SNR. (C) Difference in the PSI between the left Ins/Broca and pSTG ROIs as a function of SNR. *P < 0.05, **P < 0.01, by one-sample t tests. See Fig. 2 legend for abbreviations.