148 The International Arab Journal of Information Technology, Vol. 2, No. 2, April 2005 A Connectionist Expert Approach for Speech Recognition Halima Bahi and Mokhtar Sellami Department of Computer Science, University of Annaba, Algeria Abstract: Artificial Neural Networks (ANNs) are widely and successfully used in speech recognition, but still many limitations are inherited to their topologies and learning style. In an attempt to overcome these limitations, we combine in a speech recognition hybrid system the pattern processing of ANNs and the logical inferencing of symbolic approaches. In particular, we are interested in the Connectionist Expert System (CES) introduced by Gallant [10], it consists of an expert system implemented throughout a Multi Layer Perceptron (MLP). In such network, each neuron has a symbolic significance. This will overcome one of the difficulties encountered when we built an MLP, which is how to find the appropriate network configuration and will provide it with explanation capabilities. In this paper, we present a CES dedicated to Arabic speech recognition. So, we implemented a neural network where the input neurons represent the acoustical level, they are defined using the vector quantization techniques. The hidden layer represents the phonetic level and according to the Arabic particularities, the used phonetic unit is the syllable. Finally, the output neurons stand for the lexical level, since they are the vocabulary words. Keywords: Artificial intelligence, speech recognition, hybrid system, neuro-symbolic integration, expert system, neural networks. Received February 23, 2004; accepted July 8, 2004 1. Introduction The Artificial Intelligence (AI) approach tries to reproduce the natural human reasoning which incorporate several approaches of reasoning in particularly in perception problems. This allows us to recognize and to react instantly to sensory cues. This kind of hybrid intelligence has inspired AI researchers to combine multiple artificial methods and several information sources to deal with knowledge in an attempt to simulate human thought. Some researches in this area deal with the integration of expert systems and neural networks [7, 10, 16, 17, 20, 23]. In particular, we are interested in the connectionist expert system introduced by Gallant [10], which is a multi layer perceptron with symbolic aspect related to domain knowledge. Our system is dedicated to Arabic speech recognition; it is an MLP which recognizes isolated spoken words in Arabic. We attached to its architecture a symbolic meaning. So, the input layer represents the acoustical level, the hidden layer the phonetic level, and the output layer, stands for the lexical one. In this paper, we describe our investigations throughout the expert system-neural network integration, and we propose an integration approach which we applied to Arabic speech recognition. The remainder of the paper is structured as follows. In the second section 2, we give a brief introduction of expert neural networks. In section 3, we present the connectionist expert system approach. In section 4, we describe the conceptual elements of our recognizer. In section 5, we give implementation issues. The obtained results are presented in section 6. Finally, conclusion is drawn and perspectives are presented. 2. Expert Neural Networks 2.1. Expert Systems An expert system consists of programs that contain knowledge bases and a set of rules that infer new facts from knowledge and from incoming data. The rules are used in the inference process to derive new facts from given ones. The strength of expert systems is the high abstraction level. Knowledge can be declared in a very comprehensive manner, making possible to easily verify the knowledge base with the domain experts. The system also gives explanations for the given answers in the form of inference traces. Typical weakness is dealing with incomplete, incorrect and uncertain knowledge. Also, the system does not learn anything by itself. 2.2. Artificial Neural Networks An Artificial Neural Network (ANN) is basically a dense interconnection of simple, non-linear computation elements called “neurons”. It is assumed that a neuron has N inputs, labeled x 1 , x 2 , .., x N , which are summed with weights w 1 , w 2 , …, thresholded, and