UNIVERSITY OF WEST BOHEMIA IN PILSEN DEPARTMENT OF CYBERNETIC Optimization of Features for Robust Speaker Recognition

Currently, the old feature extraction method, which was used early for speech recognition, is used in speaker recognition in our speaker recognition group. Standard Mell Frequency Cepstral Coefficients (MFCC) features are used. They can be extended by delta and acceleration coefficients eventually. Whereas features for speech recognition has been evolved and optimized until now, features for speaker recognition remains same. These outdated features suffer from various deficiencies, regarding low robustness in particular. It can be said that these features are unsuitable in practical application. This study is aimed to examine possibilities of improving of the features. In conclusion then came up with suggestion of appropriate features extraction technique, which have been combined from examined method on the basis of the before explored methods. Main emphasis is placed on the robustness, i.e. noisy test data and/or channel disturbances (e.g. microphone mismatch). The study can be divided into several parts. At first, standard MFCC and Perceptual Linear Prediction (PLP) feature sets were optimized, i.e. the optimal numbers of the band filters and of the cepstral coefficients were examined. Next, the influence of delta and acceleration coefficients was discussed. Then, the channel normalization techniques were employed. Next, the possibilities of the linear transformations Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) were investigated. Then, the smoothing of spectrum or cepstrum in time was examined. Finally, several proposed combinations of above described approaches were tested. The new proposed features allow us to decrease the recognition error rate by 35-50%.