An Ensemble Framework of Voice-Based Emotion Recognition System

Emotion recognition will improve the user experience of artificial intelligence (AI) product. Voice-based emotion recognition system has low hardware requirement and can be easily employed onto current AI product, e.g. Amazon Echo, Google Home and so on. The previous voicebased emotion recognition system mostly focused on the speech collected under controlled environment, e.g. studio environment. However, the practical scenario will be more complicated than controlled environment. Only focusing on acoustic characteristics of speech is not sufficient to model the emotion. In this paper, we propose a ensemble framework which can capture several aspects of characteristics related to emotion. The framework is evaluated on multimodal emotion challenge (MEC) 2017 corpus. The corpus is collected from Chinese films and TV programs, whose scenarios are close to real world. The proposed framework is able to outperform the best baseline system by 29.5% absolutely.