Hybrid Fuzzy HMM System for Arabic Connectionist Speech Recognition SINOUT D. SHENOUDA Computer Science Department American University in Cairo P. O. Box 2511 Cairo EGYPT DR. FAYEZ W. ZAKI Electronics and Communications Engineering Mansoura University Faculty of Engineering EGYPT DR. AMR GONEID Computer Science Department American University in Cairo P. O. Box 2511 Cairo EGYPT Abstract:- In this paper, a new Arabic connectionist speech recognition system is presented. This recognition system is based on the combination of the fuzzy integral and measure theory [1] and Hidden Markov Model (HMM) [2] using the CSLU toolkit. The CSLU toolkit [3] is a research and development software environment that provides a powerful and flexible tool for research in the field of spoken language understanding. The objective of this paper is to design a hybrid Fuzzy HMM (FHMM) system for Arabic speech recognition. This system is based on a novel Hidden Markov Model with fuzzy logic and fuzzy integral theory. In this context, the fuzzy integral is used to relax the independence assumptions that are necessary with probability functions. Interestingly, it should be noted that one particular case in the choice of fuzzy integral (the Choquet integral), fuzzy measure (probability measure), and fuzzy intersection operator (multiplication), reduces the generalized fuzzy HMM to the classical HMM. The traditional HMM and the proposed Fuzzy HMM systems were implemented by computer simulation and a performance comparison was carried out. It is noticed that, there are some improvements in recognition accuracy in case of the Fuzzy HMM (FHMM) system over the classical HMM recognition system. The FHMM recognition system accuracy varies from 93.36% to 98.36% depending on the data set used whereas the classical HMM’ accuracy varies from 91.27% to 94.60% for the same data sets. Key-Words: - Speech recognition – Signal processing – Speech processing – Hidden Markov Model – Fuzzy logic – Fuzzy integral theory – Fuzzy measure - Man-machine communications. 1 Inroduction The main goal of Automatic Speech Recognition (ASR) is to develop techniques and systems that enable computers to accept speech as input. Speech is the human’s most efficient communication media. Using speech in man-machine communication is more comfortable and position independent than any other media that require more concentration, and restrict movement. In speech recognition techniques, pattern comparison can be performed in several ways depending on the specifics of the recognition system. There are many approaches to ASR. The main approaches are template-based, knowledge-based, and stochastic-based approaches. There are other new approaches, which are Artificial Neural Network (ANN), and fuzzy logic. The main problem in speech recognition, as with other complex tasks that require some form of intelligence, is the amount of information that must be examined before making a classification or decision. Speech recognition is an extremely complex pattern-matching problem. The complexity arises from the variability in speech rate, pitch, volume, and emotion. Together with the natural differences in individual human voice production systems, these factors produce variable and nonlinear waveforms. As if these challenges were not enough, a speech recognition system must also deal with non-speech sounds and environmental noise. There are still many research problems to resolve in speech recognition, as it is still often not completely robust or efficient for certain applications. Nevertheless, speech recognition systems today can obtain high accuracy with the utilization of neural networks, fuzzy logic and hidden Markov models (HMM). Today, the HMM is the most widely used, and its strong mathematical base allows many new studies to improve its efficiency. Recently, a novel generalization of HMM has been introduced [4] and successfully applied to hand-written recognition [5]. Proceedings of the 5th WSEAS Int. Conf. on Signal Processing, Robotics and Automation, Madrid, Spain, February 15-17, 2006 (pp64-69)