Recognition of Chinese speech using hybrid HMM/HNN models

To discriminate the complete sets of HMM for Chinese initials and finals, we construct hierarchical neural networks (HNN) integrating the knowledge of perceptual confusions among Chinese initials and finals developed by Zhang (1982). Instead of using a large monolithic neural network, the HNN system employs a large set of hierarchically organized but relatively small neural networks to perform the probability density estimation. The parameters of all neural nets in the HNN are automatically trainable using the GEM algorithm. We report results on the 1267 Chinese syllable corpus using this kind of hybrid HMM/HNN model