A Fast Framework for the Constrained Mean Trajectory Segment Model by Avoidance of Redundant Computation on Segment

Computational Linguistics and Chinese Language Processing Vol. 11, No. 1, March 2006, pp. 73-86 73 © The Association for Computational Linguistics and Chinese Language Processing [Received Feb. 24, 2005; Revised Oct. 8, 2005; Accepted Oct. 12, 2005] A Fast Framework for the Constrained Mean Trajectory Segment Model by Avoidance of Redundant Computation on Segment1 Yun Tang∗, Wenju Liu∗, Yiyan Zhang+ and Bo Xu∗ Abstract The segment model (SM) is a family of methods that use the segmental distribution rather than frame-based density (e.g. HMM) to represent the underlying characteristics of the observation sequence. It has been proved to be more precise than HMM. However, their high level of complexity prevents these models from being used in practical systems. In this paper, we propose a framework that can reduce the computational complexity of the Constrained Mean Trajectory Segment Model (CMTSM), one type of SM, by fixing the number of regions in a segment so as to share the intermediate computation results. Our work is twofold. First, we compare the complexity of SM with that of HMM and point out the source of the complexity in SM. Secondly, a fast CMTSM framework is proposed, and two examples are used to illustrate this framework. The fast CMTSM achieves a 95.0% string accurate rate in the speaker-independent test on our mandarin digit string data corpus, which is much higher than the performance obtained with HMM-based system. At the mean time, we successfully keep the computation complexity of SM at the same level as that of HMM. Keywords: Speech Recognition, Segment Model, Mandarin Digit String Recognition 1. Introduction The Hidden Markov Model (HMM) [Rabiner et al. 1993] has been used successfully for 1 This work was supported in part by the China National Nature Science Foundation (No. 60172055) and the Beijing Nature Science Foundation (No.4042025). ∗ Institute of Automation, Chinese Academy of Sciences, P.O.Box 2728, nlpr, Beijing 100080 Tel: +086 01062659279 Fax: +086 01062551993 E-mail: tangyun@nlpr.ia.ac.cn + Institute of Scientific and Technical Information of China 74 Yun Tang et al. acoustic modeling in many speech recognition systems. Given the state sequence, feature vectors are assumed to be conditionally independent, and the task of extracting the trajectory can be elegantly achieved by applying the Viterbi algorithm frame by frame. However, the above assumption is far from realistic, which limits the HMM’s ability to capture the relations within a segment. Another weakness of HMM is that it is not accurate enough to represent a non-stationary observation sequence by means of a piecewise constant state [Deng et al. 1994; Hon et al. 1999]. In order to handle these problems, a lot of methods have been proposed, including SM [Ostendorf et al. 1996], which is a family of methods among them. SM is totally different from HMM in terms of its segmental decoding method and potential for accomplishing some tasks effectively that are naturally difficult for an HMM based system, since it integrates more segmental information into the decoding process, produces the n-best list during the decoding process etc. …

Content truncated to 500 words