Enhancing Acoustic-based Speech Emotion Recognition Using CNN

Speech Emotion Recognition (SER) is a crucial field aiming to enhance machine understanding of human emotions. Among all the variety of data sources, we choose the acoustic features of the speech to explore both positive and negative aspects in feature extraction, aggressive data augmentation, and complex model architecture, highlighting its potential to inflate false confidence (overfitting) and hallucinations if not wielded with caution. Furthermore, our investigation explores the complex link between accuracy and feature set size, highlighting the need to extract rich emotional characteristics from speech. An intriguing finding of our analysis is: downscaling the filter size and learning rate of the neural network works better with small training dataset and the model can differentiate diverse emotions more accurately. This finding paves the way for improving the classification model, potentially leading to more robust emotion recognition systems. This multifaceted analysis goes beyond the conventional focus on accuracy, challenging assumptions and pinpointing areas for future exploration in the ever-evolving field of speech-based emotion recognition.