Affect awareness is important for improving human-computer interaction, but also facilitates the detection of atypical behaviours, danger, or crisis situations in surveillance and in human behaviour monitoring applications. The present work aims at the detection and recognition of specific affective states, such as panic, anger, happiness in close to real-world conditions. The affect recognition scheme investigated here relies on an utterance-level audio parameterization technique and a robust pattern recognition scheme based on the Gaussian Mixture Models with Universal Background Modelling (GMM-UBM) paradigm. We evaluate the applicability of the suggested architecture on the PROMETHEUS database, implemented in a number of indoor and outdoor conditions. The experimental results demonstrate the potential of the suggested architecture on the challenging task of affect recognition in real world conditions. However, further enhancement of the affect recognition performance would be needed before any deployment of practical applications.
展开▼
