ISCA Archive ICSLP 2002
ISCA Archive ICSLP 2002

The structure and its implementation of hidden dynamic HMM for Mandarin speech recognition

Feili Chen, Jie Zhu, Wentao Song

Speech is time variable stochastic process. But the HMM model is stationary and time-independent. We present a new model as Hidden Dynamic HMM model to describe the dynamic property of speech. A hidden layer of dynamic property is interpolated between the observation and state. Estimated Dynamic Property and Predicted Dynamic Property is introduced to describe the hidden dynamic property. The result shows that the HDHMM model can achieve good improvement in different tasks of speech recognition.