Modeling A-Priori Likelihoods for Angry User Turns with Hidden Markov Models

Current studies dealing with the detection of angry users in automated telephone-based speech applications take acoustic and sometimes linguistic information into account in order to classify the emotional state of the caller in single user turns. Angry user turns, however, don’t appear from nowhere and the likelihood of observing another angry turn rises substantially when anger has already been observed previously in the discourse. In this contribution we examine the context of angry user turns in two different telephone corpora. We then introduce Hidden Markov Models (HMM) as classifiers modeling the temporal aspects of anger across single turns. As additional information source, the HMMs improve our acoustic classifier serving as baseline substantially. Performance gains of 1-4 % can be reported by performing late fusion of the acoustic classifier and the HMMs.