The automatic segmentation of the vocal signal using predictive neural network

The automatic segmentation of the vocal signal precedes the features extraction stages, respectively the emotion recognition/classification. The extraction of the prosodic parameters as fundamental frequency (F0) and formants (F1-F4) cepstral coefficients LPCC and MFCC are made only on the vowel areas. The analysis tools from the SROL corpus are using a hybrid hierarchical system with four segmentation methods based on the autocorrelation function, AMDF method, the cepstral analysis and HPS method. Since the performance of this instrument has not been yet satisfactory, we analyzed other segmentation possibilities in order to obtain the best possible accuracy in segmentation. The predictive neural network used in this paper is in fact a simple perceptron which can approximate with high accuracy the quasi-periodic signals such as the vowels. The consonants have noisy properties and are complicated transition processes. The prediction error for the consonants comparing with the vowels is higher when it is used a sample neural network architecture.