Forensic identification for electronic disguised voice based on supervector and statistical analysis

This paper proposed a novel algorithm for forensic identification of electronic disguised voice based on supervector and statistical analysis. The supervector was stacked by mixtures of mean vector of Gaussian mixture model. SVM classifier was used to identify whether a testing voice was disguised or not. By comparing the difference of Mel-cepstrum coefficients statistical characteristics between normal and disguised voice, we studied the variation of voice parameters. Experimental results showed that we can get good detection performance with error identified rate lower than 7%.