Overview of candidate encoding models.

<p>(A) Joint modulation selective filters. (A1) Joint frequency specific: the spectrogram is filtered with a bank of modulation selective filters at different spectral modulations (Ω), temporal modulations (ω), and direction (upwards/downwards). The output of the filter bank is averaged across time and direction to yield a reduced representation of modulation energy as a function of Ω, ω, and frequency. The joint frequency-specific MTF-based model predicts that fMRI responses vary linearly with this representation, i.e. sounds that differ with respect to any of the three dimensions will elicit different responses. (A2) Joint frequency non-specific: the 3D modulation representation is averaged across frequency to yield a global measure of modulation energy. By concatenating modulation and frequency content (not shown here, see tonotopy model), the joint frequency non-specific model predicts separate processing for global, joint modulations and frequency. (B) Independent modulation selective filters. (B1) Independent frequency-specific: the spectrogram is filtered with purely spectral and purely temporal modulation selective filters and the output is averaged over time. This yields separate representations of spectral and temporal modulation energy as a function of frequency. The independent frequency-specific model predicts that the response of a voxel dedicated to spectral (temporal) processing will not be affected by a change in temporal (spectral) modulation content. (B2) Independent frequency non-specific: the two separate representations of spectral and temporal modulation energy are averaged across frequency to yield the global spectral and temporal modulation content. This representation is concatenated with the frequency content (not shown here, see tonotopy model) to simulate separate processing for frequency, spectral and temporal modulations. (C) Tonotopy model: the spectrogram is averaged over time and voxels are modeled as frequency selective units, whose response varies linearly with the frequency content of the input stimuli.