IMPROVING THE RECOGNITION PERFORMANCE BY USING A PARALLEL-BRANCH SUBUNIT MODEL-BASED ON MISRECOGNISED DATA

Abstract

An initialisation and training method using a parallel-branch subunit model is obtained which uses a continuous hidden Markov model for improved recognition performance. The model is obtained by adding a new subunit branch based on misrecognised data in the training data to the previous parallel branches for that subunit. This procedure is shown to be efficient and gives good word recognition performance.