Automatic segmentation of TIMIT by dynamic programming

We propose an algorithm based on the principle of dynamic programming for the automatic segmentation of continuous speech into phoneme-like units. A measure of local dissimilarity among consecutive feature vectors is combined with a knowledge of the expected statistical distribution of the segment lengths within a dynamic programming framework to obtain an optimal placement of segment boundaries. We compare the performance of our algorithm with the performance of two recently-proposed alternatives by measuring how closely the hypothesised boundaries match the TIMIT phone boundaries. The results showed that we are able to improve on the performance of the two contrasting approaches. Furthermore, we show that a hybrid approach which combines aspects of all three algorithms leads to even better results.