Identification Of Electronic Disguised Voice Based On Gaussian Supervector And SVM

With the development of digital multimedia technology, digital information has been widely used and spread, which brings great convenience to human communication. Speech signal processing technology gradually become mature, and the application of this technology are more and more extensive.This kind of information is easy to be modified, so that it is faced with a serious crisis of malicious tampering. In recent years, a large number of softwares appear in mobile phone application store,such as “Wechat Voice Changer”, “Super Voice Changer” and so on, which can change the speaker’s voice a lot. As a result, the hearer cannot identify the speaker’s age and sex,even the hearer is familiar with the speaker.By comparing the difference of voice mel-cepstrum coefficients statistical characteristics between nature and disguised voice, we studied the variation of voice parameters. This paper proposed a novel algorithm for identification of electronic disguised voice based on supervector combined by mean vectors of gaussian mixture model and SVM classifier for training and identification. Experimental results showed that the identification rate can reach 80%.What’s more, according to the characteristics of electronic disguised voice often contain noise,we introduce a kind of speech feature parameters which is based on Gammatone filter bank, and we propose a method to apply this speech feature parameters to the electronic disguised voice identification system. The experimental results show that the improved electronic disguised voice identification system has a better robustness.