A Study Of Modeling Of Inter-Frame Correlationin Segment-Based Models

Acoustic modeling is very important for improving the performance of speech recognition system. HMM is widely used in speech recognition system presently. But one of the assumptions of HMM is state-conditioned stationarity of the observation vectors, implying that each state is a stationary source generating independent identically distribute (i.i.d.) observation vectors. This assumption of stationary is not valid as time sequence of the observation vectors is highly correlated. Thus, modeling of time correlation of a sequence of observation has become the most challenging topics of current research for acoustic modeling. In this thesis, we discuss the modeling of inter-frame correlation for speech recognition in the framework of stochastic segment model. Including: (1) We describe the motivation of stochastic segment model and develop linear dynamical system model for speech recognition. Stochastic segment model can better capture the dynamics structure over the segment of speech. Thus, we look at speech on a segmental level rather than on a frame-by-frame basis. Moreover, many experiments show linear models are adequate in modeling of inter-frame correlation. So we think that linear segment-based models can better characterize speech signal. (2) The EM algorithm for linear system identification is implemented and an initialization of the EM algorithm is proposed by using statistical analysis. EM algorithm has become one of the methods of choice for ML estimation. But it often converges slowly. If model parameter is suitably initialized, the speed of convergence is quicker. The initialization method of the algorithm proposed can quicken speed of convergence, and the algorithm is numerically very stable.(3) We proposed two ideas: one is the smoothing of trajectory and the other is context modeling. Smoothing of trajectory imply that linear dynamical system is stochastic trajectory modeling. By context modeling, we can provide an exact mathematical model for each syllable. The experiment results show that both smoothing of trajectory and context modeling can improve the performance of the system...