Neural network based classification of stressed speech using nonlinear spectral and cepstral features

The goal of this research work is to propose two spectral features, namely, the `Bark band spectral energy´ and the `significant spectral energy´, for the task of stressed speech classification and compare the result with mel frequency cepstrum coefficients (MFCCs) features. It is shown that these two spectral features outperform traditional cepstral (MFCC) features. Spectral energy in 17 bands of frequencies on Bark scale as well as 16 mel-scale warped cepstral coefficients were used independently for classifying stressed speech. The proposed features employ a neural network model based on the Levenberg-Marquardt algorithm. Their observed performance demonstrates the viability of the Bark spectral energy set in stressed speech detection experiment and classifies angry, question, and clear stressed speech conditions. Preliminary results of matching features for a small set of utterances showed correct detection of speech condition in better than 83% of the cases using each set of features from Bark band energy. Significant spectral energy and MFCC features, on the other hand, showed close to 75% correct detection for the same utterances.