Previous research on large-vocabulary automatic speech recognition (ASR) has mainly concentrated on European and Asian languages. Other language groups have been explored to a lesser extent, for instance Semitic languages like Arabic and Hebrew. These languages possess certain characteristics which present problems for standard ASR systems. For example, the written representation of Arabic normally lacks short vowels and other information present in the spoken form, which makes it difficult to utilize textual training data. Furthermore, Arabic has a complex morphological structure, which is characterized not only by a high degree of affixation but also by the interleaving of vowel and consonant patterns. This leads to a large number of possible word forms, which complicates the robust estimation of statistical language models.

In this talk I will give a detailed overview of the problems Arabic presents for ASR and of the solutions that will be pursued during the 2002 JHU workshop. The main focus will be on a novel approach to language modeling ('factored language modeling'). Under this approach individual word forms are decomposed into multiple morphological components or streams; n-gram statistics are computed separately for each stream and are combined to predict the probability of a word form given previous word forms. It will be shown how the structure of such a model can be learned automatically from data and how it generalizes to other areas of speech processing which utilize n-gram modeling techniques.