Hierarchical organization of human auditory cortex: evidence from acoustic invariance in the response to intelligible speech

Abstract

Hierarchical organization of human auditory cortex has been inferred from functional imaging observations that core regions respond to simple stimuli (tones) whereas downstream regions are selectively responsive to more complex stimuli (band-pass noise, speech). It is assumed that core regions code low-level features, which are combined at higher levels in the auditory system to yield more abstract neural codes. However, this hypothesis has not been critically evaluated in the auditory domain. We assessed sensitivity to acoustic variation within intelligible versus unintelligible speech using functional magnetic resonance imaging and a multivariate pattern analysis. Core auditory regions on the dorsal plane of the superior temporal gyrus exhibited high levels of sensitivity to acoustic features, whereas downstream auditory regions in both anterior superior temporal sulcus and posterior superior temporal sulcus (pSTS) bilaterally showed greater sensitivity to whether speech was intelligible or not and less sensitivity to acoustic variation (acoustic invariance). Acoustic invariance was most pronounced in more pSTS regions of both hemispheres, which we argue support phonological level representations. This finding provides direct evidence for a hierarchical organization of human auditory cortex and clarifies the cortical pathways supporting the processing of intelligible speech.

Figure 1.

Spectrograms for a sample set of stimuli based on the sentence, the shop closes for lunch.

Figure 2.

Group results from a standard BOLD amplitude subtraction analysis (P = 0.05, false discovery rate corrected) projected onto a surface-rendered template brain showing regions that respond more to intelligible speech (clear + NV) than unintelligible speech (rot + rotNV).

Figure 3.

Activation of auditory core regions (HG) to 8-Hz amplitude-modulated wideband noise versus rest (scanner noise), P < 0.001, uncorrected. (A) Axial and coronal slices showing activation in HG and immediately surrounding tissue. Images are in radiological convention (left = right). (B) Average signal time course for each speech condition for the HG ROI in each hemisphere. Note the virtually identical average response to clear and rot speech.

Figure 4.

Location of the peak voxel in STS ROIs for each subject projected onto a surface-rendered template brain. ROIs were identified using the contrast, clear minus rot with a threshold of P < 0.0001, uncorrected. Red, aSTS; green, mSTS; yellow, pSTS. Note the nonoverlapping distributions of these ROIs across subjects.

Figure 5.

Classification accuracy (proportion correct) for the 4 classification contrasts in the 3 left hemisphere ROIs and 4 right hemisphere ROIs. The left 2 bars in each graph are “intelligibility” contrasts (intelligible vs. unintelligible) and the right 2 bars in each graph are acoustic contrasts (acoustically different intelligible vs. intelligible and acoustically different unintelligible vs. unintelligible). Thick black horizontal line indicates chance (0.5) and thin black line marks the upper bound of the 95% confidence interval determined via a bootstrapping procedure. As the bootstrapping procedure was calculated separately for each classification contrast, the 95% confidence interval boundary can vary from one condition to the next. LH, left hemisphere; RH, right hemisphere; clear, clear speech; rot, rotated speech; NV, noise-vocoded speech; rotNV, rotated noise-vocoded speech. *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 6.

Acoustic invariance across the 7 ROIs as measured using an acoustic invariance index (see text). This index effectively ranges from −1 to 1 where positive values indicate higher degrees of acoustic invariance and negative values indicate lower degrees of acoustic invariance (i.e., more acoustic sensitivity). *P < 0.05, **P < 0.01, 2 tailed.