Experiments in connected word recognition

Abstract

This paper deals with different versions of a one-stage dynamic programming algorithm for connected word recognition using whole word templates. To test the different versions of the recognition algorithm, a set of one hundred seven-digit strings spoken by a male and a female speaker was used, which results in 700 single-digit recognition tests for each speaker. The first aspect to be studied is concerned with a reduced and variable frame rate of the acoustic parameters, in order to cut down the computational expenditure of the algorithm. An adapted version of the so-called trace segmentation, which had been found very useful in isolated word recognition, turns out to be not suited to connected word recognition without substantial modifications. Then, five different sets of local path constraints in the dynamic programming recursion are studied and evaluated. The experimental results indicate no critical dependence on the details of the local path constraints as long as an approximate compensation for differences in path lengths is performed. Finally, the effect of introducing a pruning threshold into the recognition algorithm is considered. This pruning technique was able to reduce the computational expenditure to 21% without loss in recognition accuracy.