A new algorithm for isolated word recognition with large within-class variability

Large variations in the acoustic patterns of successive utterances of the same word may occur with speaker-independent speech recognition system or with single speakers afflicted with certain pathologies, such as cerebral palsy. If the variability is sufficiently large , the performance of time-warping algorithm based on dynamic programming may be unsatisfactory. This paper describes the theoretical development , implementation and a test of an algorithm designed to overcome this difficulty in situations where most of the variability occurs in successive utterances of the consonants. The results obtained using the test data showes (90%) correct recognition rate while Dynamic Programming algorithm gave only (55.5%) using the same data.