A speech recognition method based on feature distributions

摘要

A new speech recognition method based on feature distributions is presented. The speech pattern of an utterance is represented by a sequence of feature vectors. The vector space occupied by the feature vectors is partitioned into a number of subspaces. The feature vectors in a subspace are considered as the random samples generated by one of the feature distribution functions which are obtained by an iterative training algorithm. Each feature distribution function is associated with a certain state. Then an utterance is represented by a sequence of states, and a class of speech patterns is interpreted by a Markov model. The number of states and the state-transition probabilities for a Markov model are determined by the training utterances without the presumption of the model structure. The Viterbi algorithm is employed in the speech recognition procedure. The experimental results show that the proposed method is superior to the conventional hidden Markov model method.