An isolated word recognizer system based on corrective training

Abstract

A corrective training method of the gradient type which is based on the modification of the state transition probabilities is developed. To increase the discrimination between two HMMs (hidden Markov models) lambda /sub 1/ and lambda /sub 2/, Viterbi's algorithm is used to segment the sequence of observations, obtaining for the state i and the sequences O/sup (1)/ and O/sup (2)/ the permanencies in the state i: n/sub i//sup (1)/ n/sub i//sup (2)/, respectively. With this value, the statistics 'of the model are estimated. After a few iterations an acceptable convergence is obtained.<>