Evaluation of some distance measures for speaker independent isolated word recognition

In this paper a number of distance measures for speaker independent recognition of isolated words are proposed. These distance measures use auto-correlation coefficients alone or autocorrelation and linear predictor coefficients as feature paramaters of the speech samples. One measure used is the measure in discrete l2space of linear functionals. Some of the other distance measures used are the "nearest in angle" or the normalized correlation measure and the Mahalanobis distance. Actual evaluation of these distance measures is then performed using a standard 40 word reading test vocabulary spoken by 25 different speakers. All the above measures give good recognition results. The best distance measure has given a recognition rate of 87.3%.