IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO..pdfVIP

IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 1, JANUARY 2004 47 Target-Directed Mixture Dynamic Models for Spontaneous Speech Recognition Jeff Z. Ma, Member, IEEE, and Li Deng, Senior Member, IEEE Abstract— In this paper, a novel mixture linear dynamic model the speech models some key dynamic properties of the human (MLDM) for speech recognition is developed and evaluated, where speech process. several linear dynamic models are combined (mixed) to represent In recent research work along this direction [4], [8]–[10], different vocal-tract-resonance (VTR) dynamic behaviors and the mapping relationships between the VTRs and the acoustic [21], we have been developing a statistical coarticulatory model observations. Each linear dynamic model is formulated as the for spontaneous speech recognition. Our new approach is based state-space equations, where the VTRs target-directed property on statistical Bayesian theory, yet is a radical departure from is incorporated in the state equation and a linear regression the current HMM-based approach. Rather than using a large function is used for the observation equation that approximates number of unstructured Gaussian mixture components to ac- the nonlinear mapping relationship. A version of the generalized count for the tremendous variation in the observable acoustic EM algorithm is developed for learning the model parameters, where the constraint that the VTR targets change at the segmental data of highly coarticulated spontaneous speech, the new model level (rather than at the frame level) is imposed in t