Automatic Spoken Customer Query Identification for Arabic Language

In this paper we propose an approach that aims to build an automated task-oriented Arabic dialogue system which is capable to determine the topic of spoken question asked by telecom provider customers. The system is based on an Arabic adapted CMU sphinx ASR. In addition to formal Arabic speech, our implemented Arabic ASR is capable to recognize some Palestinian Arabic dialectal words. The recognized text is used to determine the question category using supervised machine learning techniques in order to take desired action such as routing customer call to the appropriate destination. The best performance of proposed overall system is 76.4% accuracy with random forest classifier provided by Weka toolkit tested on 750 questions recorded by 30 speakers with Palestinian dialect.