Impact of different voting strategies in CNN based speech emotion recognition

Automatic emotion recognition systems detect the emotional state of a person by analyzing expressions that can be recorded using different modalities. Among them, speech represents one of the major modalities that can be used independently or fused with other modalities. Due to the fact that emotional state may vary over a short time and that there are similarities among features obtained for different emotions, voting strategy has a key role when observing features collected over time. The focus of this paper is to examine the impact of different voting strategies in state-of-the-art speech emotion recognition systems based on convolutional neural networks on the example of the recently recorded Serbian Emotional Amateur Corpus.