Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

State-CodeBook based Quasi Continuous Density Hidden Markov Model with Applications to Recognition of Chinese Syllables

Ren-Hua Wang, Hui Jiang

Speech Communication Lab, University of Science and Technology of China, Anhui, China

In this paper a new improved type of HMM, called State-CodeBook based quasi continuous density HMM (SCBHMM) is proposed and is tested in the recognition of Chinese syllables . The SCBHMM is composed of a set of model parameters, which can explicitly incorporate more acoustic characteristics of speech under limited training data. Here, the observation probability is associated with the static feature of speech within certain state and state transition probability is related to the temporal variations in speech spectra. SCBHMM suggests an effective method to integrate static and dynamic features in speech recognition. Preliminary experiments on the standard Chinese Speech Database CRDB x, showed that the proposed SCBHMM not only achieved a great improvement over the original HMM, but also greatly reduced the computation consumption in the training process. An accuracy of more than 92% for the top one candidate and 99% for the top five candidates was achieved in the Chinese Syllables recognition.

Full Paper

Bibliographic reference.  Wang, Ren-Hua / Jiang, Hui (1994): "State-codebook based quasi continuous density hidden Markov model with applications to recognition of Chinese syllables", In ICSLP-1994, 211-214.