Effect of Feature Dimension on Classification of Speech Emotions

This paper analyses both the static and temporal dynamics of the spectral features in classifying speech emotions. Initially, different frame-level spectral techniques such as the Linear Prediction Cepstral Coefficients (LPCC), Perceptual LP coefficients (PLP), and Mel-Frequency Cepstral Coefficients (MFCC) have been examined. Further, these spectral features are extracted using Wavelet Analysis (WA) for a better emotional portrayal. The extracted feature sets remain high-dimensional and overload the recognizer with redundant features, large memory space, and slower response. To alleviate these issues and fetch more discriminating parameters, the applicability of Vector Quantization in clustering the data has been explored. Machine learning algorithms such as the Gaussian Mixture Model (GMM), the Probabilistic Neural Network (PNN), and the Multilayer Perceptron (MLP) have been simulated with the derived feature sets for their effectiveness in classifying speech emotions. While the GMM has been efficient in classifying the frame-level feature dimension, the NN-based classifiers outperform the GMM for low feature dimensions as revealed from our results.