5th International Conference on Spoken Language Processing
To integrate the hierarchy structure of discrimination between all HMM states for Chinese Initials and Finals, we constructed in this paper Hierarchical Neural Networks (HNN), which differ from Jordan's HME in such extensions as more complex parameterization for gate and/or expert and dimension-reduced expert network. With these extensions, we can reuse those pre-trained simple node networks in a hierarchy structure (HNN), and fine-tune them jointly by Generalized Expectation Maximization (GEM) algorithm. The proposed HNNs were used within hybrid HMM-ANN models to perform the estimation of posterior probabilities for HMM states. Instead of using a large monolithic neural network, the HNN system can be trained in a short time compared with MLP estimator and result in a speed-up in decoding time over the conventional systems. We have applied the proposed hybrid HMM-HNN method to the recognition task of Chinese Continuous Speech., achieve a promising word error rate of 26.4%.
Bibliographic reference. Jia, Ying / Du, Limin / Hou, Ziqiang (1998): "Hierarchical neural networks (HNN) for Chinese continuous speech recognition", In ICSLP-1998, paper 0415.