A Comprehensive Approach to Speech Emotion Recognition: Combining MFCC, Multifractal, Prosodic Features, and Volterra Filtering

Abstract

This study proposes a new approach to speech emotion recognition by integrating traditional Mel-Frequency Cepstral Coefficients (MFCC) features with multifractal features derived from speech signals. Then a Sparse Adaptive Volterra Filter was used to improve these features. A Convolutional Neural Networks – Bidirectional Long Short- Term Memory (CNN-BiLSTM) was used to classify the emotions. Experiments were conducted on a self-built Malayalam emotional dataset of 3120 samples. Accuracy increased from 89.10% with the MFCC-only model to 93.43% when combined with multifractal features. This is an absolute improvement of 4.33%. Adding the Volterra filter raised the final accuracy to 95.51%, a total improvement of 6.41% over the MFCC-only baseline. Including pitch and formant features raised the accuracy further to 96.15%. For further validation, the method was also tested on the RAVDESS, a widely used benchmark dataset. Overall, the proposed framework gave high accuracy compared to baseline methods. These results show the potential of this methodology for practical emotion recognition applications, and points to the model’s robustness and effectiveness in real-world scenarios.