Sound Localization With Stochastic Cross-Correlation

Members: Tara Julia Hamilton, Jonathan Tapson, Suma George, Qian Liu

In this project we used the stochastic cross-correlator chip to try and do localisation using the silicon cochlea data.


The XCORR is a stochastic cross-correlator chip that was designed by Tara and Jon. This chip is based on the cross-correlation algorithms developed by Jon. The concept is illustrated below.

The XCorr chip has the following top-level operation.

The basic elements of the XCorr are tuned via a serial interface. There are 144 parameters to tune as well as an AER and scanner output interface. Unfortunately the AER interface was not up-and-running at Telluride this year. The scanner interface worked and the parameters of the chip were eventually tuned so that the XCorr can be used in experiments.

Preliminary testing of the XCorr

In order to prove that the XCorr worked, a square wave input is input into the chip with the inputs offset. Counting the positive and negative edges the offset can be determined. The images below show the histograms for cross-correlation of two square waves 90 degrees out of phase and 180 degrees out of phase. The delay is shown in the peaks in the histogram.

Cochlea Data

The cochlea data was collected by Dmitria and Anna. This data is for both males and females at different angles from the binaural cochlea. Example data from the silicon cochlea is given below. Each "ear" is colour coded.

The cochlea data was very noisy and difficult to analyse by eye, although it should be fine for the XCorr chip. In order to run it through Xcorr and the FPAA, Suma and Tara had to change the data into real pulses for each "point". This is a big job and is ongoing post-workshop. Also Shih-Chii will provide further data using sinusoids to help with analysis. This project is ongoing but the results will be exciting!