Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

A Multi-State NN/HMM Hybrid Method for High Performance Speech Recognition

Dong Yu, Taiyi Huang, Dao Wen Chen

National Lab of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, People's Republic of China

This paper proposed a novel Multi-State NN/HMM Hybrid Method named Multi-State Gaussion Competitive Neural Network (MSGCNN) for High Performance Speech Recognition. The basic idea of this new approach is integrating the Viterbi algorithm into a Gaussion Competitive Neural Network (GCNN). Comparing with Self-Aligning Networkp] it is more successful for it has both the Time alignment ability of HMM and the strong discrimination capability of Neural Network (NN). Moreover, because GCNN has better performance than basic Multilayer Perceptron (MLP), the novel system has many strong points such as faster training, more robust and noise immunity. An all Chinese syllable recognition system have been established based on MSGCNN and the comparative experiments confirmed the good characters stated above.

Full Paper

Bibliographic reference.  Yu, Dong / Huang, Taiyi / Chen, Dao Wen (1994): "A multi-state NN/HMM hybrid method for high performance speech recognition", In ICSLP-1994, 1503-1506.