Audio Emotion Mapping using Enhanced Techniques with Augmented Log-Melspectrograms

Using a novel approach based on log-Mel spectrogram with augmentation, this research piece delves into the realm of audio emotion classification. Although it is much lower than MFCC, the research reveals a reasonable accuracy rate of 63%. On the other hand, it shows how much room there is for improvement. One complex model used in the research is a 2D convolutional neural network (CNN) based on the VGG19 architecture. This model treats the audio characteristics as an image with dimensions of 30x216 pixels. Specialised functions are employed to extract various attributes from datasets like TESS, CREMA-D, SAVEE, and RAVDESS. The significance of quickly recognising emotions through aural signals is emphasised. Results from the 2D CNN model are promising, and it has the potential to outperform more traditional methods. The results can be more quickly interpreted by storing them in a specific class. By bringing together audio and image-based approaches to emotion analysis, this work improves the growing field of audio emotion categorization and opens the door to more accurate and efficient identification.