Arabic automatic segmentation system and its application for Arabic speech recognition system

The paper presents an Arabic automatic segmentation system to be utilized in the development of an Arabic speech recognition system. The automatic segmentation system is used to label a speech database system that is used in the training of continuous, phoneme based speaker independent Hidden Markov Models. Our experiments showed that 7.8 % of the phoneme boundaries of automatic segmented data deviate from those that were manually segmented more than 30 milliseconds while 0.78 % deviate more than 70 milliseconds. Our experiments showed also that automatic segmentation led to improvement in speech recognition accuracy of 0.49 % for a 306 words bigram language model test and 0.14% for 1340 words bigram model