Automatic emotion variation detection using multi-scaled sliding window

Emotion recognition from speech plays an important role in developing affective and intelligent Human Computer Interaction. The goal of this work is to build an Automatic Emotion Variation Detection (AEVD) system to determine each emotional salient segment in continuous speech. We focus on emotion detection in angry-neutral speech, which is common in recent studies of AEVD. This study proposes a novel framework for AEVD using Multi-scaled Sliding Window (MSW-AEVD) to assign an emotion class to each window-shift by fusion decisions of all the sliding windows containing the shift. Firstly, sliding window with fixed-length is introduced as the basic procedure, in which several different fusion methods are investigated. Then multi-scaled sliding window is employed to support multi-classifiers with different timescale features, in which another two fusion strategies are provided. Finally, a postprocessing is applied to refine the final outputs. Performance evaluation is carried out on the public Berlin database EMO-DB. Our experimental results show that proposed MSW-AEVD significantly outperforms the traditional HMM-based AEVD.