Numerical Learning-Based Method for the Automatic Processing of Complex Disfluencies in Spontaneous Arabic Speech

In this paper, we propose a numerical learning-based method for the automatic processing of complex disfluencies in spontaneous oral Arabic utterances. This method Allows, from a pretreated and semantically labeled utterance, to delimit and label the conceptual segments of a processed utterance. Also, it allows, from a segmented utterance, to detect and delimit the disfluent segments and then correct them. This work is a part of the realization of the Arabic vocal server SARF [2]. Thus, we implemented the complex disfluencies processing module (MTDC). The evaluation of the MTDC gave us satisfactory results with an F-measure equal to 91.9%. After integrating the MTDC into the SARF system, we achieved an improvement of 11.88% in acceptable understanding and a 3.77% decrease in error rate.