Figure 6.
Cyclo-spectral template matching model. (A) An overview of the model. The incoming sound waveform is transformed into a spectro-temporal representation at the early auditory stages. In stage 1, the signal within each spectral channel is passed through a filterback of deconvolution filters of varying Q, constructed as two-sides derivatives of the gammatone function. In the last stage, through lateral inhibition, the most active Q based channel is enhanced and averaged over time. (B) The output of the stage 2 of the model of an input signal (black trace, inverted for clarity of figure) with f = 1000 Hz, Q = 2, rate 1 with 10% Gaussian noise (gray trace, inverted for clarity of figure) for with Q = 0.5 (green trace), 2 (red trace), and 8 (blue trace). The convolution of the input and was normalized by the maximum value and half-wave rectified. The output for the filter with Q = 2 captures all the peaks present in the original signal, while the output for the filter with Q = 8 smoothes over neighboring peaks, and output for filter with Q = 0.5 misses the peaks because of the noise. (C) The deconvolving filter with matching Q relays the most information about the original signal. The information rate of the output of stage 2 of the model as in (B). Note that the red trace (Q = 2) depicts a greater information content transmitted by the matching deconvolving filter. The green trace (Q = 0.5) exhibits most power in the noise band. The blue trace (Q = 8) exhibits lower information than the red trace over all frequency bands. (D) Examples of cyclo-spectral templates: Predicted mean activation of cyclo-spectral (Q–f) channels. Sounds with constant Q across spectral bands, such as the naturally perceived sounds with Q = 2, will activate predominantly the channels of the same Q across spectral channels. Sounds from control 1, composed of chirps whose temporal structure is constant, activate the channels with varying Q across spectral bands: Q will be higher for higher frequencies.