A preliminary study of prosody-based detection of questions in arabic speech monologues

Prosody features have been widely used in many speech-related applications, including speaker and word recognition, emotion and accent identification, topic and sentence segmentation, and text-to-speech applications. Languages other than Arabic have received a lot of attention in this regard. An important application of prosodic features which is investigated here is that of identifying question sentences in Arabic monologue lectures. To our best knowledge, this is the first attempt at addressing question detection from spoken lectures in any language. To this end, we developed a small corpus made of 1028 utterances that were extracted from 15 Arabic spoken lectures. We approach this problem by first segmenting the continuous speech (recorded lectures) into sentences using both intensity and duration features. Prosodic features are, then, extracted from each sentence. These features are used as input to four different classifiers to classify each sentence into either a question or a non-question sentence. Our results suggest that questions are cued by more than one type of prosodic features in spontaneous Arabic speech. We classified questions with an accuracy of 77.43%. A feature-specific analysis further reveals that energy and fundamental frequency (F0) features are mainly responsible for discriminating between question and non-question sentences. In terms of classification, we found that a Bayes Network performs better than support vector machines, multi-layer perceptron neural networks, or decision trees on our dataset. Removal of correlated features through Correlation-based Feature Selection produced more efficient and accurate results than the complete feature set.