Improving Performance of Speech Emotion Recognition Application using Extreme Learning Machine and Utterance-level

The recognition of emotions in speech is increasingly vital for numerous applications, with systems that identify human speech emotions playing a crucial role in decision-making systems. These systems can also assist individuals suffering from various diseases by monitoring their emotion. Therefore, the accuracy needs to be as high as possible to address this issue, where the determination of speech features is a key factor. In previous work, the ELM-based model could recognize the types of speech emotions with a maximum accuracy value of 92.55%. To increase accuracy, we proposed a feature-fusing strategy by leveraging utterance-level features and combining them with speech features. We utilize Toronto Emotional Speech Set (TESS) data, which comprises a total of 2,800 high-quality audio data centered on female voice to guarantee data dependability, and data split strategy at 80:20. Several speech features were extracted and fed into an Extreme Learning Machine (ELM) model to obtain emotion probability distribution to get the level of utterance features. The utterance-level feature was combined with speech features to form new features for the classification methods. The experimental findings suggest that the proposed method can enhance the value of accuracy by approximately 3.22% to 6.28% and can recognize the types of speech emotions with an accuracy value ranging from 95.55% to 100% using an ELM-based model with Extreme Gradient Boosting (XGBoost), Random Forest (RF), Relevance Vector Machine (RVM), Support Vector Machine (SVM) and Stochastic Gradient Descent (SGD) classification methods.