Figure 1. Schematic of stimuli and imaging protocol.

(A) Cochleagrams for two example natural sounds (left column) and corresponding synthetic sounds (right four columns) that were matched to the natural sounds along a set of acoustic statistics of increasing complexity. Statistics were measured by filtering a cochleagram with filters tuned to temporal, spectral, or joint spectrotemporal modulations. (B) Schematic of the imaging procedure. A three-dimensional volume, covering all of ferret auditory cortex, was acquired through successive coronal slices. Auditory cortical regions (colored regions) were mapped with anatomical and functional markers (Radtke-Schuller, 2018). The rightmost image shows a single ultrasound image with overlaid region boundaries. Auditory regions: dPEG: dorsal posterior ectosylvian gyrus; AEG: anterior ectosylvian gyrus; VP: ventral posterior auditory field; ADF: anterior dorsal field; AAF: anterior auditory field. Non-auditory regions: hpc: hippocampus; SSG: suprasylvian gyrus; LG: lateral gyrus. Anatomical markers: pss: posterior sylvian sulcus; sss: superior sylvian sulcus. (C) Response timecourse of a single voxel to all natural sounds, before (left) and after (right) denoising. Each line reflects a different sound, and its color indicates its membership in one of 10 different categories. English and non-English speech are separated out because all of the human subjects tested in our prior study were native English speakers, and so the distinction is meaningful in humans. The gray region shows the time window when sound was present. We summarized the response of each voxel by measuring its average response to each sound between 3 and 11 s post-stimulus onset. The location of this voxel corresponds to the highlighted voxel in panel B. (D) We measured the correlation across sounds between pairs of voxels as a function or their distance using two independent measurements of the response (odd vs. even repetitions). Results are plotted separately for ferret fUS data (left) and human fMRI data (right). The 0 mm datapoint provides a measure of test–retest reliability and the fall-off with distance provides a measure of spatial precision. Results are shown before and after component denoising. Note that in our prior fMRI study we did not use component denoising because the voxels were sufficiently reliable; we used component-denoised human data here to make the human and ferret analyses more similar (findings did not depend on this choice: see Figure 1—figure supplement 2). The distance needed for the correlation to decay by 75% is shown above each plot (${\tau }_{75}$). The human data were smoothed using a 5 MM FWHM kernel, the same amount used in our prior study, but fMRI responses were still coarser when using unsmoothed data (${\tau }_{75}$ = 6.5 mm; findings did not depend on the presence/absence of smoothing). Thin lines show data from individual human (N = 8) and ferret (N = 2) subjects, and thick lines show the average across subjects.

Figure 1—figure supplement 1. The effect of enhancing reliable signal using a procedure similar to ‘denoising source separation (DSS)’ (see ‘Denoising part II’ in Materials and methods) (de Cheveigné and Parra, 2014).

(A) Voxel responses were denoised by projecting their timecourse onto components that were reliably present across repetitions and slices. This figure plots the test–retest correlation across independent splits of data before (x-axis) and after (y-axis) denoising (data from experiment I). Each dot corresponds to a single voxel. We denoised either one split of data (blue dots) or both splits of data (green dots). Denoising one split provides a fairer test of whether the denoising procedure enhances SNR. Denoising both splits shows the overall effect on response reliability. The theoretical upper bound for denoising one split of data is shown by the black line. The denoising procedure substantially increased data reliability, with the one-split correlations hugging the upper bound. This plot shows results from an eight-component model. (B) This figure plots split-half correlations for denoised data (one split) as a map (upper panel), along with a map showing the upper bound (lower panel). Denoised correlations were close to their upper bound throughout auditory cortex. (C) This figure plots the median denoised correlation across voxels (one split) as a function of the number of components used in the denoising procedure. Gray line plots the upper bound. Shaded areas indicate the 95% confidence interval, computed via bootstrapping across the sound set. Results are shown for both experiments I (left) and II (right). Predictions were near their maximum using approximately eight components in both experiments (the eight-component mark is shown by the vertical dashed line).

Figure 1—figure supplement 2. Effect of component denoising on human fMRI results.

This figure plots normalized squared error (NSE) maps comparing natural and synthetic sounds in humans both before (top) and after denoising (bottom) by projecting onto the six reliable components identified in our prior work (Norman-Haignere et al., 2015). We used component-denoised data for all species comparisons to make the analyses more similar, but results were similar without denoising. The bottom panel is the same as that shown in Figure 2E and is reproduced here for ease of comparison. Results are based on 12 human subjects.