Performance of Speech Emotion Recognition Utilizing Machine Learning Techniques

Speech Emotion Recognition is very crucial in enhancing human-computer interaction as it helps the systems to recognize the emotions and respond accordingly. The work is focused on the speech emotion recognition of human voices through the advanced extraction of features, which comprises Mel-Frequency Cepstral Coefficients, mel spectrograms, chroma, and zero-crossing rate. To develop the entire dataset, TESS and RAVDESS datasets were combined, so the ground for emotion classification is solid. Although previous studies using machine learning and data mining techniques have shown promising results, they often fail to integrate a comprehensive set of influencing factors. This study attempts to overcome these limitations by applying multiple machine learning algorithms that uncover complex patterns in the data, thus improving the accuracy and reliability of SER systems. The performance of the models was assessed and the following accuracies were obtained: Random Forest with 83.78%, KNN with 71.20%, SVM with 69.01%, Decision Tree with 55.29%, and Naive Bayes with 40.15%. This illustrates the strength of ensemble methods, particularly Random Forest, in modeling complex interactions within the data and achieving superior performance. With strong datasets and a variety of algorithms, this work further helps advance the state of art in SER. This outcome is valuable for insights in the design of more effective emotion recognition systems, therefore helping to enhance human and machine interaction in various applications.