Sub-Word HMM Recognition: An Investigation of Phone Context Modelling and Improved Discrimination

In this paper two techniques are presented for improving the performance ₒf sub-word recognition with open vocabularies. The rst technique uses a new sub-triphone unit, called a honicleI to allow triphone models which have not been encountered in the training ata to be built from contexts which have been suf ciently trained. The second uses linear discriminant analysis (LDA) to improve the discrimination between the sound classes. The two techniqiies have been evaluated for s eaker de endent trainin on an open vocabulary task. e recogniser is based on hi den Mar 0v modelling HMM) techniques with continuous distributiom.