The ARABEX speech recognition system

The author describes an expert system which uses intensively a large amount of a spectrogram-reading database. A system overview is given and the principal stages of the Arabic decoder are developed, namely: (1) selection of the class for each speech segment to enable the guidance of the phonetic decoding process; and (2) decoding strategy using bottom-up analysis between segmentation and phonetic identification. After low-pass filtering and sampling, the speech signal is submitted to a set of preprocessing algorithms to be segmented and then classified into phonetic classes. With another set of processing algorithms to extract the pertinent acoustic features, combined with a set of production rules, each of these deduced phonetic segments is assigned to a list of phonemes. The output is a string of segments containing alternative phonemes, together with their matching score