Transformer-Based Multilingual Speech Emotion Recognition Using Data Augmentation and Feature Fusion

In recent years data science has been applied in a variety of real-life applications such as human-computer interaction applications, computer gaming, mobile services, and emotion evaluation. Among the wide range of applications, speech emotion recognition (SER) is also an emerging and challenging research topic. For SER, recent studies used handcrafted features that provide the best results but failed to provide accuracy while applied in complex scenarios. Later, deep learning techniques were used for SER that automatically detect features from speech signals. Deep learning-based SER techniques overcome the issues of accuracy, yet there are still significant gaps in the reported methods. Studies using lightweight CNN failed to learn optimal features from composite acoustic signals. This study proposed a novel SER model to overcome the limitations mentioned earlier in this study. We focused on Arabic vocal emotions in particular because they received relatively little attention in research. The proposed model performs data augmentation before feature extraction. The 273 derived features were fed as input to the transformer model for emotion recognition. This model is applied to four datasets named BAVED, EMO-DB, SAVEE, and EMOVO. The experimental findings demonstrated the robust performance of the proposed model compared to existing techniques. The proposed SER model achieved 95.2%, 93.4%, 85.1%, and 91.7% accuracy on BAVED, EMO-DB, SAVEE, and EMOVO datasets respectively. The highest accuracy was obtained using BAVED dataset, indicating that the proposed model is well suited to Arabic vocal emotions.