Detection of Negative Emotional States in Real-World Scenario

In the present work we evaluate a detector of negative emotional states (DNES) that serves the purpose of enhancing a spoken dialogue system, which operates in smart-home environment. The DNES component is based on Gaussian mixture models (GMMs) and a set of commonly used speech features. In comprehensive performance evaluation we utilized a well-known acted speech database and real-world speech recordings. The real-world speech was collected during interaction of naive users with our smart-home spoken dialogue system. The experimental results show that the accuracy of recognizing negative emotions on the real- world data is lower than the one reported when testing on the acted speech database, though much promising, considering that, often, humans are unable to distinguish the emotion of other humans judging only from speech.