Syllabic Markov models of Arabic HMMs of spoken Arabic using CV units

We survey evidence - orthographic distributional phonological and psycholinguistic - in favor of a model of Arabic speech sounds based on the CV unit and extensive use of the silent sukuun vowel. We then construct a small-vocabulary multi-speaker CV HMM similar to the phonemic HMMs based on tied triphones that are widely used in speech recognizers for English and other European languages. Using experimental measures of recognition accuracy and trainability, we demonstrate that the CV type of model outperforms a standard tied triphone recognizer for Arabic speech, using Cohen's kappa ration for statistical comparison. Finally we argue that models based on CV units may also lead to better stemmers, spell-checkers and other natural language processing tools for Arabic.