New features for emotional speech recognition

Bio-medical research extends towards human voice and auditory systems day by day. Similarly it helps for the security issues. Emotion analysis and recognition for such purpose is a challenging task. To analyze and recognize, the emotions has been attempted in this piece of work. Initially, Sub-band spectral features have been extracted to characterize high arousal angry, happy, fear, surprise and neutral speech emotions. Further, two new features as spectrum and cepstrum with vector quantization has been used. Finally, simulations for robust feature have been applied to Probabilistic Neural Network (PNN) classifier for recognition and the performance. Classifier performance degrades in presence of large unknown data due to data dependency smoothing parameter over sensitization of training data. As Vector Quantization (VQ) has the ability to reduce the feature size, the modified feature sets are developed to improve the classifier robustness. Promising results have been manifested in result section.