A Speech-Based Emotion Recognition Framework

Abstract

Human-computer interaction (HCI) has become one of the most challenging areas of research in the field of artificial intelligent (AI) at the present time. Speech emotion recognition (SER) introduces a new means of communication between humans and machines. Enabling a machine to understand human emotion renders it more capable of understanding the speech process. Despite the great progress and intensive research performed in this area, there is still a lack of naturalness in identifying emotions. There is a need to fill the gap between commercial interest and current performances. The key is to find significant speech emotion features that can map emotion correctly and efficiently. The previous works of SER extracted and selected different sets of acoustic features. However, the most significant features have not yet been found. These problem is addressed in this research by proposing a speech emotion recognition framework that provide an enhancement of features extraction technique and hybrid feature selection method respectively. The voice quality prosodic spectral-based feature extraction (VQPS) is implemented using prosodic and spectral features extraction technique in addition to new and traditional voice quality features extraction technique. At the same time, the balanced hybrid filter-based feature selection (BHFFS) consists of two layers: the balancing layer; and the hybrid filter-based layer. The proposed features extraction technique and selection method was successfully experimented through the use of EMO-DB dataset. The experimental results proved that using VQPS leads to performance improvement upon previous works. In addition, it demonstrates that the voice quality features are important in developing the SER system. In the same manner, BHFFS performance outperforms the previous work performance.