Speech Emotion Recognition Using ML Models and Audio Features

The ability to recognize and respond to human emotions has become a crucial component in enhancing humancomputer interaction, with the development of speech emotion recognition technologies playing a pivotal role in this advancement. While deep learning models have driven significant progress in the field of speech emotion recognition, traditional machine learning algorithms offer a practical alternative, balancing performance with lower computational requirements. This study presents a comprehensive approach to speech emotion recognition using the Berlin Emotional Speech Database, classifying seven emotional states: anger, sadness, anxiety or fear, disgust, boredom, happiness, and neutrality. The study employs a range of acoustic features, including pitch, RMS energy, MFCCs, and formants, which are combined with 14 statistical descriptors and extracted using tools like OpenSMILE. Preprocessing steps, such as normalization, noise reduction, and silence removal, are applied to enhance the quality and reliability of the data. The performance of traditional machine learning models, including Support Vector Machine, Random Forest, and $k$-Nearest Neighbors, is evaluated on the processed dataset. The results demonstrate the effectiveness of these traditional models, with Support Vector Machine achieving the highest classification accuracy of 90.65 %, followed by Random Forest and $k$-Nearest Neighbors. The results of this study highlight the capacity of traditional machine learning techniques to effectively capture the complexities of emotional expression, while circumventing the computational burden associated with deep learning models. The practical relevance of this research extends to real-time applications across various domains, including healthcare, virtual assistants, and customer service, where the demand for efficient and reliable emotion recognition systems is paramount.