Novel Sub-band Spectral Centroid Weighted Wavelet Packet Features with Importance-Weighted Support Vector Machines for Robust Speech Emotion Recognition

In this paper, we propose novel sub-band spectral centroid weighted wavelet packet cepstral coefficients (W-WPCC) for robust speech emotion recognition. Wavelet packet transform (WPT), as an effective tool for non-stationary signal analysis, is applied for speech analysis with a human auditory perception based WP filterbank structure. For each sub-band, the spectral centroid, which has been proved to be noise-robust, is calculated. On this basis, the W-WPCC feature is computed by combining the sub-band energies with sub-band spectral centroids via a weighting scheme to generate noise-robust acoustic features. The importance-weighted support vector machine (IW-SVM) is proposed to improve the robustness of classifier to the noises, while the important weight is utilized to compensate the covariate shift between test dataset and training dataset. Clean speech environments while demonstrates better noise-robustness in noisy environments and the IW-SVM improves the robustness to white Gaussian noise in speech emotion recognition compared with conventional classifiers.