Implementation of an isolated word recognition system

Abstract

A speaker-independent Arabic digits recognition system is implemented which uses template matching of input utterances with a stored set of multiple templates for each digit. The system is based on the LPC parameters for features, the log likelihood ratio for a distance function between frames, the procedure of dynamic time warping for time-normalization between test and reference utterances and the K-NN rule for decision criterion. Four utterances for each digit from each speaker were collected to form a data-base of 80 replications for every word. The reference templates were obtained from a statistical clustering analysis of this data-base. Two implementations were considered; in one approach the LPC features are extracted at a fixed rate while in the other approach they are extracted at a variable rate by merging similar neighbouring fixed size frames. Both implementations were tested against a three category test data-base, (a) utterances used to train the system, (b) utterances not used to train the system but by speakers used in the training phase, and (c) utterances from speakers not used in the training phase.