Isolated word recognition using continuous state transition-probability and DP-matching

A report is presented on a novel application of the Markov model to an automatic speech-recognition system, in which the feature vectors of speech represent the states of the Markov model, the transition probability of the states is represented by a multidimensional normal density function of the feature vector, and the DP (dynamic programming) matching algorithm is used for calculating the optimum time sequence of the states. Based on experimentation with the system in a speaker-independent mode, using a vocabulary of ten Japanese single-digit numerals, the current system is shown to be more effective than recognizers using Maharanobis´ distance, Euclidean distance, or the absolute distance