Abstract
Most of the available resources of colloquial Arabic speech are transcribed without diacritics. Those diacritics provide short vowels and other pronunciation information and by omitting them a considerable amount of ambiguity is introduced. In this paper, we propose the use of an automatic diacritisation method as front-end for training of automatic speech recognition systems of colloquial Arabic. The system used is based on conditional random fields that are trained on speaker and contextual information. This method outperforms other reported methods in diacritisation colloquial Arabic by 13.2% relative. The empirical experiments show that applying this method on acoustic model training transcriptions improves the recognition performance in Levantine colloquial Arabic by 1.8% relative.