Fast Search Algorithms for Connected Phone Recognition Using the Stochastic Segment Model

Fast Search Algorithms for Connected Phone Recognition Using the Stochastic Segment Model V.Digalakist M.Ostendorf t J.R. Rohlicek t Boston University 44 Cummington St. Boston, MA 02215 :~ BBN Inc. 10 Moulton St. Cambridge, MA 02138 ABSTRACT In this paper we present methods for reducing the compu- tation time of joint segmentation and recognition of phones using the Stochastic Segment Model (SSM). Our approach to the problem is twofold: first, we present a fast segment classification method that reduces computation by a factor of 2 to 4, depending on the confidence of choosing the most probable model. Second, we propose a Split and Merge seg- mentation algorithm as an alternative to the typical Dynamic Programming solution of the segmentation and recognition problem, with computation savings increasing proportionally with model complexity. Even though our current recognizer uses context-independent phone models, the results that we report on the TIMIT database for speaker independent joint segmentation and recognition are comparable to that of sys- tems that use context information. Introduction The Stochastic Segment Model [6, 10, 9] was proposed as an alternative to hidden Markov models (HMMs), in order to overcome the limiting assumptions of the latter that ob- servation features are conditionally independent given the underlying state sequence. The main disadvantage of the SSM over other methods is its computational complexity, which can be attributed to the fact that dropping the con- ditional independence assumption on the state sequence in- creases the size of the effective state space. Having ad- dressed several issues on the classification accuracy of the SSM in our previous work [2], we now move to speaker in- dependent connected phone recognition and choose to con- centrate on the important issue of computational complexity of the SSM recognition algorithm. We present two new al- gorithms which together achieve a significant computation reduction. This paper is organized as follows. A brief overview of the SSM is given in this section. For a more detailed de- scription of the model the reader is referred to [6, 10]. In the next section, we present a new algorithm for fast phone classification with the SSM. Next, we address the issue of joint segmentation and recognition. The originally proposed solution for this problem used a Dynamic Programming al- gorithm to find the globally maximum likelihood phone se- quence. After we examine the computational complexity of this method, we look at an alternative approach to this problem based on local search algorithms. We propose a different search strategy that reduces the computation sig- nificantly, and can be used in conjunction with the classifi- cation algorithm. Although theoretically the new algorithm may not find the global maximum, experimental results on the TIMIT database, show that the recognition performance of the fast algorithm is very close to the optimum one. An observed segment of speech (e.g., a phoneme) is rep- resented by a sequence of q-dimensional feature vectors Y = [Yx Y2 ... YL], where the length L is variable. The stochastic segment model for Y has two …

Content truncated to 500 words