Correlation computations are widely used in template-based temporal pattern recognition, but are expensive to compute in real time on DSP systems. We have designed and fabricated a low-power chip which uses current-mode analog circuits to compute the correlation between an auditory input signal in the time-frequency domain and a stored binary template. We use a bucket-brigade device (BBD) to accumulate partial column sums over time at rates consistent with the auditory-band input. Experimental results of the correlator demonstrate its correct operation over a wide range of operating frequencies.
I . INTRODUCTIONMany time-based classification systems compute the correlation of an incoming discrete-time signal or signals with a predetermined template [ 11. While for speech and other complex long-term signals it is necessary to perform dynamic time warping (DTW) or similar weighting of the incoming signal [ 2 ] , for transients (short-term acoustic events of less than approximately 1/10 of a second), a simple correlation in the time-frequency domain yields accurate classification results [3].A general form of the simple correlation iswhere M is the number of frequency channels of the input, N is the maximum number of time bins in the window, IC is the array of input signals split into frequency bands, p , is the matrix of template pattern values for pattern z , and t is the current time.This formula produces a running correlation c,[t] of the input array with the template z. For large M and N , this algorithm can be expensive to execute on a DSP in terms of speed and power requirements. However, our approach lends itself elegantly to low-power parallel analog computation in the form of MOS transistor circuits operating primarily in the subthreshold mode.In the following discussion, we present our algorithm as a set of incremental modifications to the baseline algorithm of Equation (1). First, we normalize the input, which is essential for the steps which follow. Next, we take time derivatives of the input and templates, In this form, we are able to make the template values binary without significantly increasing classification error rates. Finally, we move the time-differencing to the output, yielding the simplest possible form of the equation.
SIMPLIFYING THE CORRELATION EQUATIONWe assume the input x to the system is an acoustic signal, split into M frequency bands. We compute the (rectified) energy envelope for each band, denoted y, and then normalize these system inputs by the functionwhere 8 is a threshold value included to suppress noise during quiet invervals in the input. Pineda et a1. [4] have shown that the normalized input representation is essential to significantly simplify the pattern classifier algorithm, greatly reducing the size and complexity of the hardware implementation but not degrading the classification result.To make the calmputation less expensive, first we use the time difference between samples of the input and compare it to the difference between samples of the template. The effect of thi...