A reference speech recognition algorithm for benchmarking and speech data base analysis

A complete connected word reference algorithm has been developed at NBS. It provides score statistics and confusibility measures as well as word decisions on the speech data base on which it is run. The basic algorithm is like the Bridle, Brown, and Chamberlain [1] algorithm with Euclidean distances on mel scale cepstral coefficients. There are options as to the particular spectral features, the type of spectral comparison measures and the training scheme. Score statistics are collected and a measure of confusibility is computed for each word in the vocabulary. The algorithm has been tested on the Texas Instruments Isolated Word Data Base [2]. The most confusible words in the data base are no and go. In addition the performance of the reference system has been compared with several generic types of recognizers.