Emotional Resonance Unleashed by exploring Novel Audio Classification Techniques with Log- Melspectrogram Augmentation

This research work examines the domain of audio emotion categorization, employing a unique approach that incorporates log-Mel spectrogram with augmentation. The research demonstrates a commendable accuracy rate of $63 \%$, which is notably inferior to that of MFCC. Nevertheless, it also underscores the considerable capacity for growth. The study employs an advanced model, namely a 2D Convolutional Neural Network (CNN) that closely mimics the VGG19 architecture. In this approach, the acoustic characteristics are represented as a $30 \times 216$ pixel image. Specialised functions are employed to extract a diverse variety of attributes from datasets such as TESS, CREMA-D, SAVEE, and RAVDESS. The significance of swiftly recognising emotions through audio clues is emphasised. The 2D CNN model exhibits promising results, surpassing traditional methods. A custom class is employed to encapsulate outcome attributes to expedite prompt interpretation. This work not only advances the growing field of audio emotion classification but also facilitates accurate and efficient identification, bridging the gap between audio and image-based approaches in emotion analysis.