Anger detection in call center dialogues

We present a method to classify fixed-duration windows of speech as expressing anger or not, which does not require speech recognition, utterance segmentation, or separating the utterances of different speakers and can, thus, be easily applied to real-world recordings. We also introduce the task of ranking a set of spoken dialogues by decreasing percentage of anger duration, as a step towards helping call center supervisors and analysts identify conversations requiring further action. Our work is among the very few attempts to detect emotions in spontaneous human-human dialogues recorded in call centers, as opposed to acted studio recordings or human-machine dialogues. We show that despite the non-perfect performance (approx. 70% accuracy) of the window-level classifier, its decisions help produce a ranking of entire conversations by decreasing percentage of anger duration that is clearly better than a random ranking, which represents the case where supervisors and analysts randomly select conversations to inspect.