Abstract

The Correlation Network model serves as a framework for models of auditory processing for pitch, timbre, localization and sound segregation. It comprises three modules. The first module calculates arrays of running correlation coefficients (two autocorrelation arrays, one crosscorrelation array). Each array is two dimensional, indexed in time measured relative to a sliding origin (the ”present”), and lag. If peripheral frequency analysis is taken into account, the arrays have a third dimension: tonotopy. Integration over a sliding window in the correlation calculation removes most of the fine time structure, so the output of the first module consists of slowly varying values. The second module calculates a weighted sum of its inputs. The third module controls the weights of the second module while monitoring its output, and is responsible for producing the behavior expected from the model. Based upon the Correlation Network model, a wide range of models of pitch, timbre and binaural processing can be implemented, in particular those involving correlation and cancellation. It offers a uniform basis for these operations with a simple mapping to known anatomy (module I to the brainstem, modules II and III to midbrain and beyond). It allows complex models (such as multi-stage cancellation) to be cast in relatively simple and plausible terms. It provides useful inspiration for signal processing tasks such as F0 estimation, spectral estimation and source separation. 1.

Citations