A Baseline Speech Recognition System for Levantine Colloquial Arabic

The Arabic language is characterized by the existence of many different colloquial varieties that significantly differ from the standard Arabic form. In this paper, we propose a state-of-the-art speech recognition system for Levantine Colloquial Arabic (LCA). A fully continuous context dependent acoustic model was trained using 50 hours of speech from the BBN DARPA Babylon corpus. Pronunciation modeling was initially grapheme-based due to the absence of diacritic marks in transcriptions. Acoustic model parameters have been optimized including number of senones and Gaussians. In order to improve speech recognition accuracy, a cross-lingual hybrid acoustic and pronunciation modeling approach is proposed, where a MSA phoneme-based acoustic model is adapted using a small amount of LCA speech data. The adapted AM was then combined with the initial grapheme-based model to create a hybrid acoustic model.