Acoustic-phonetic recognition in the ARABEX system

In this paper we consider the knowledge-based approach whose main advantages are pointed out, particularly the possibility of incrementally building up a complex system and updating and maintaining a knowledge base, and the capability for the expert system to explain its reasoning process. Our expert system uses intensively a large amount of knowledge derived from the performance emulation on spectrogram-reading database. The knowledge is coded in the Arabic decoder system under the forms: procedures for segmentation and labelling of segments into gross phonetic classes; contextual production rules; and decoding strategies. The output is a string of segments containing alternative phonemes, together with their matching score. Three male speakers are used for the recognition step. Our system is developed to serve as a part of an automatic speech recognition system for continuously spoken Arabic sentences and can be considered as as important step towards a system for oral dialogue between man and machine.