Call center is the department that is most relevant to audio data. One of its major tasks is to monitor customers' angry because it has a negative impact on the organization. Therefore, if customers' angry can be detected automatically, the call center monitoring process will be improved. This research aims at developing an approach to automatically classify anger voice in call center dialogs. The approach involves 4 processes: (1) Audio pre-processing, preparing the data especially separating customer's voice from the whole call. (2) Emotion annotation (3) Voice feature extraction, from each turn dialog, extracting 1,582 features that can be used to classify emotion. (4) Anger classification modeling, building classification model using 4 machine learning methods: Random Forest Classifier, Decision Tree, LinearSVC, and Logistic Regression. The number of turn dialogs used in the experiments is 1,345 turns of which 517 turns are labeled with “angry” and 828 turns are labeled with “not angry”. The experimental results showed that Random forest classifier has the highest recall and balanced accuracy which are 83.49% and 85.53%, respectively.