Calculation of Classification Functions Using a Large Number of Spoken Word Samples

Abstract

A two-phase procedure for calculation of multiple linear classification functions has been developed. Phase I approximates the distribution of samples in the measurement space by a series of linear functions and preselects only those samples lying in the critical region. Phase II utilizes the reduced number of samples for calculation by a linear programming algorithm, the multiple linear classification functions that can separate given samples strictly. The procedure was applied to the problem of spoken-word recognition. About 3300 samples of Japanese spoken digits, uttered by 55 male speakers and represented as 56-dimensional measurement vectors by a vocoder-type frequency analyzer are collected. During Phase I, the number was reduced to 16 linear functions on the average. Precise classification functions were then calculated for each digit in Phase II. Most of the digits were separated by a single linear function while others were separated by two functions. An error rate of less than 0.2% was achieved for 550 samples newly uttered by the same 55 speakers.