About Correlograms


The purpose of the correlogram is to model the temporal representation of sound in the brain. The auditory system, especially at lower levels, does a spectacular job of preserving the timing of auditory events. This is true even with large changes in loudness. Thus the correlogram is an ideal way to summarize the temporal periodicity of an auditory signal. These temporal events are a good model of pitch perception, and this was the first reason for the correlogram's success.

Licklider's Correlograms

The neural architecture of Licklider's correlogram looks like this.

Licklider's original caption: Schematic diagram of overall analyzer. At the bottom is the uncoiled cochlea. Its lengthwise dimension and the corresponding dimension in the neural tissue above it is designated the x-dimension. The cochlea performs a crude frequency analysis of the stimulus time function, distributing different frequency bands to different x-positions. In the process of exiting the neurons of the auditory nerve, the outputs of the cochlea filters are rectified and smoothed. The resulting signals are carried by groups of neurons A to the autocorrelators B, whose delay- or tau-dimension is orthogonal to x. The outputs of the autocorrelators are fed to higher centers over the matrix of channels C, a cross-section through which is called the (x, tau)-plane. (Output arrows arise from all the dots; some are omitted in the diagram to avoid confusion.) The time-varying distribution of activity in the (x, tau)-plane provides a progressive analysis of the acoustic stimulus, first in frequency and then in periodicity.

This is an image of the first correlogram, as proposed by J. C. Licklider. The horizontal axis is labeled x and represents position along the cochlea. We now rotate the correlogram so that frequency is along the vertical axis. The base of the cochlea is sensitive to high frequencies, and is at the top of the plot.

Modern Correlograms

With the increase of computing power, the correlogram has been employed as a representation for sound separation (Weintraub, 1985) and pitch perception. These papers and others are listed in the bibliography.

It is important to separate the idea---measuring periodicity is a good way for the auditory system to understand sound---from the implementations (autocorrelation, stabilized image, or various forms of mechanical or neural delay lines.) There are many ways to calculate a correlogram-like representation, but all aim to summarize the temporal information in auditory firings. This is in contrast to models that only consider the average firing rate of a neuron.

Most of the correlograms in this web site were computed using Lyon's Cochlear Model in concert with an implementation of the correlogram based on autocorrelation, and computed for efficiency reasons using an FFT. An alternate model, based on trying to find a stable representation of the neural firings that are flowing through the system, is called the stabilized auditory image. While a correlogram based on autocorrelation is simpler and easier to analyze analytically, is is perhaps not as realistic as the stabilized image.

For more examples

To see collection of correlograms, showing what the correlogram looks like for many different sounds, go to the Examples.