Enhancing Human Emotion Classification in Human-Robot Interaction

Speech Emotion Recognition (SER) is vital for enhancing Human-Robot Interaction (HRI). In the past decade, various models have been applied to diverse datasets. However, these datasets often fell short in representing real-world scenarios. This paper introduces an innovative acoustic feature set, based on Mel Frequency Cepstral Coefficients (MFCC) and Mel-Spectrograms, applied to a combined dataset comprising RAVDESS, TESS, and EmoDB. Our proposed model achieves an impressive accuracy of 85.05% on this combined dataset, demonstrating its suitability for real-world applications. This advancement holds significant promise for improving HRI systems, affective computing, and AI-driven applications, where understanding and responding to human emotions through speech is crucial. The robustness and high accuracy of our model provide valuable insights for researchers and practitioners seeking to implement SER technology in practical, real-world settings. This study offers two primary contributions. Firstly, it involves the compilation of a unified dataset from the three mentioned sources. Secondly, it achieved the highest accuracy of 85.05%.