Correlation Network Model applied to F0-adaptive spectral estimation

Correlation Network model. This model (hereafter CN model) is an abstract model of auditory processing that allows a range of auditory signal-processing functions to be implemented in a uniform way: pitch, timbre, localization, segregation. It consists of three modules (Fig. 1). The first calculates arrays of running autocorrelation (monaural) and crosscorrelation (binaural) coefficients. The second forms a linear combination of coefficients produced by the first module. The third controls the parameters of the second module while monitoring its output. It is responsible for producing the behavior needed for each function (pitch, etc.). Fast signal processing is limited to the first module, while the second and third handle relatively slowly varying quantities. This might ease mapping of the model to the auditory system (first module to brainstem, second and third to more central levels). The CN model can implement any model that operates on a quadratic statistic (power or correlation) of linear combinations of delayed versions of its inputs. In particular it can implement cancellation models (Durlach, 1963; de Cheveigné, 1993). It has also proved useful as a basis for F0 estimation (de Cheveigné and Kawahara, 2002). While the CN model addresses auditory processing, here we treat it as a digital signal processing model to demonstrate how the desired functionality can be obtained. The basic ingredients of the CN model are arrays of running autocorrelation (AC) and crosscorrelation coefficients (here only the former are considered). Using a sampled-signal notation, the AC function of the signal at the left ear is: