Optimizing Speech Feature Extraction: A Comparative Study of MFCC, MFDWC, and MFWPC for Integration with Xception Block-Based CNN in Speech Emotion

The experiments conducted in this research examines the performance of various mel-frequency based feature extraction modules such as Mel-Frequency Cepstral Coefficient (MFCC), Mel-Frequency Discrete Wavelet Coefficient (MFDWC), and Mel-Frequency Wavelet Packet Coefficient (MFWPC) on the Speech Emotion Recognition task. Empirical studies proved that MFDWC performs better than MFCC and MFWPC. Besides that, in the emotion classification module, this study tested the performance of Xception block as the convolutional building blocks that requires lower number of learnable parameters. The model with Xception Block has better efficiency against Residual Block as the compared building block. The final stage of the research resulted in the output of an ideal Convolutional Network architecture and settings that had been trained in a single-corpus Speech Emotion Recognition scenario on each of the Ravdess, EMOVO, and Emo-DB datasets. This study evaluates the performance of MFCC, MFDWC, and MFWPC as feature extraction modules for Speech Emotion Recognition, demonstrating that MFDWC outperforms the others