Abstract—The paper presents a complete discrete statistical framework, based on a novel vector quantization (VQ) front-end process. This new VQ approach performs an optimal distribution of VQ codebook components on HMM states. This technique that we named the distributed vector quantization (DVQ) of hidden Markov models, succeeds in unifying acoustic micro-structure and phonetic macro-structure, when the estimation of HMM parameters is performed. The DVQ technique is implemented through two variants. The first variant uses the K-means algorithm (K-means- DVQ) to optimize the VQ, while the second variant exploits the benefits of the classification behavior of neural networks (NN-DVQ) for the same purpose. The proposed variants are compared with the HMM-based baseline system by experiments of specific Arabic consonants recognition. The results show that the distributed vector quantization technique increase the performance of the discrete HMM system. Keywords—Hidden Markov Model, Vector Quantization, Neural Network, Speech Recognition, Arabic Language I. INTRODUCTION UTOMATIC Speech Recognition (ASR) can be viewed as a successive transformations of the acoustic micro- structure of the speech signal into its implicit phonetic macro-structure. The main objective of any ASR system is to realize the mapping between the two structures. The hidden Markov model (HMM) is actually the most used approach to the ASR. Several types of HMMs as discrete, continuous and semi continuous HMMs [1], [2] have been developed and applied to the ASR. The discrete HMM (DHMM) is attractive in terms of algorithmic complexity; that is why, it has been investigated in several studies [3], [4], [5], [6]. Recently, in the context of the prodigious growth of network applications, discrete HMM-based speech recognition systems that use a Vector Quantization (QV) front-end process constitute a very useful and inexpensive solutions [7], [8]. Manuscript received October 5, 2005. M. Debyeche, Laboratory of Signal Processing and Speech communication (LCPTS), faculty of Electronics and Computer Sciences, (USTHB), P.O. Box 32 El-Alia, Bab-Ezzouar, Algeirs, Algeria.(phone:+213 72 63 86 44; fax:++213 21 24 71 87; e-mail: mdebyeche@caramail.com and mdebyeche@usthb.dz ). J.P Haton, LORIA/INRIA-Lorraine, 615 rue du jardin botanique, P.O. Box 101, F-54600, Villers-lès-Nancy, France (e-mail: jph@loria.fr). A. Houacine, Laboratory of Signal Processing and Speech communication (LCPTS), faculty of Electronics and Computer Sciences, (USTHB), P.O. Box 32 El-Alia, Bab-Ezzouar, Algeirs, Algeria.(e-mail: ahouacine@usthb.dz). In this scenario, it is highly desirable to perform compression of acoustic features, but it is crucial that the VQ involved in the front-end stage does not introduce noise that degrades the recognition accuracy. This is the dilemma. In fact, discrete HMM inherently suffers from some problems due to the Vector Quantization (VQ) process. The lack of sufficient training data involved by the VQ causes poor HMM parameter estimation, and this inevitably leads to a degradation of recognition performance. This paper is dedicated for improving accuracy issues of discrete HMM- based ASR systems. It proposes a complete discrete statistical framework, based on the use of a novel VQ-based front-end process. This new approach performs an optimal distribution of VQ codebooks on HMM states. This technique, which has been named the distributed vector quantization (DVQ) of hidden Markov models, succeeds in unifying acoustic micro- structure and phonetic macro-structure, when the parameter estimation of HMM is performed. The DVQ technique is implemented through two variants. The first variant uses the K-means algorithm (K-means-DVQ) to optimize the VQ, while the second variant exploits the benefits of the classification behavior of neural networks (NN-DVQ) for the same purpose. The evaluation is done by focusing on specific Arabic consonants: emphatic and back consonants. The characterization of these consonants has captured the interest of many researchers, since they are specific to the Arabic language [9]. The paper is structured as follows: after the first, introductory section, we present in the second section the wellknown statistical paradigm used for speech recognition represented by the HMM. In section 3 we depicts the framework of distributed vector quantization. Section 4 reports the comparative results of trials that aim to evaluate the proposed techniques by focusing on some specific Arabic phonemes. Finally, we summarize our major findings in section 5. II. VQ/HMM SYSTEM FOR WORD RECOGNITION To illustrate an application of HMMs for speech recognition, we present in Fig.1 our implementation of an isolated word recognition system based on discrete hidden Markov models. We have a vocabulary of L words to be recognized, and each word is to be modeled by a disctinct HMM. The training sets consist of K utterances of each word, pronounced by one or more speakers. In order to obtain a word recognizer, we performed the following steps: A New Vector Quantization front-end Process for Discrete HMM Speech Recognition System M. Debyeche, J.P Haton, and A. Houacine A World Academy of Science, Engineering and Technology International Journal of Electronics and Communication Engineering Vol:1, No:6, 2007 1627 International Scholarly and Scientific Research & Innovation 1(6) 2007 scholar.waset.org/1307-6892/2743 International Science Index, Electronics and Communication Engineering Vol:1, No:6, 2007 waset.org/Publication/2743