| Anger is potentially the most important human emotion to be detected in human-human dialogs, such as those found in call-centers and other similar fields. It measures the satisfaction, or lack of it, of a speaker directly from his or her voice. Recently, many software applications were built as a result of anger detection research work. In this thesis, we design a real-time framework to detect anger from spontaneous Arabic conversations. We construct a well-annotated corpus for anger and neutral emotion states from real-world Arabic speech dialogs for our experiments. We use a hybrid segmentation method of sentence length and fixed-length splitting techniques. The classification is based on acoustic sound features that are more appropriate for anger detection. Many acoustic features will be studied such as the fundamental frequency (f0), formants, energy and mel-frequency cepstral coefficients (MFCCs). Several classifiers are evaluated, and the experimental results show that support vector machine classifiers can yield 82.4% detection rate which is the best accuracy result as compared to other classifiers; in addition to being fast for online recognition. Moreover, a real-time software application is developed based on our proposed framework. |
