EUROSPEECH '95 Fourth European Conference on Speech Communication and Technology Madrid, Spain September 18-21, 1995

Stochastic Trajectory Models for Speech Recognition: An Extension to Modelling Time Correlation

Mohamed Afify (1,2), Yifan Gong (1), Jean-Paul Haton (1)

(1) CRIN/CNRS - INRIA-Lorraine, BP 239, Vandoeuvre, France
(2) National Telecommunication Institute, Nasr City, Cairo, Egypt

Based on experimental evidence from other research work, we assume that speech signals in consecutive analysis frames are correlated and we explore the correlation for speech recognition in the framework of stochastic trajectory models (STM). In this paper, we extend our previously proposed stochastic mixture trajectory models to model such a time correlation. Specifically, we consider the observation sequence of a speech signal as being generated by a mixture of random trajectory generators, and explicitly model the time evolution of the mean vector of each trajectory as the sum of a first order non-observable AR process and a state dependent random process. We describe the main results of the formulation, parameter estimation and sentence recognition. Evaluated on a speaker-dependent continuous speech recognition task, the proposed approach reduced average word error rate by about 25%, compared to a baseline STM system in which frames are assumed to be independent.

Full Paper

Bibliographic reference.  Afify, Mohamed / Gong, Yifan / Haton, Jean-Paul (1995): "Stochastic trajectory models for speech recognition: an extension to modelling time correlation", In EUROSPEECH-1995, 515-518.