Identification Of Computer-generated Spoofed Speech And Natural Speech

Speaker recognition is a kind of biometric authentications.Because of speech's advantages in efficient acquirement,transmission and storage,speaker recognition has a promising application in future identity authentication.With the development of speech synthesis technique,a computergenerated speech which sounds like a target speaker can be easily obtained.Currently,most of speaker recognitions can hardly discriminate the computer-generated spoofed speech from human speech.That may reduce the reliability of speaker recognition system and restrict its application.Two spoofing countermeasures,for speeches generated by two typical speech synthesis techniques(unit selection and statistical parametric speech synthesis),have been proposed in this thesis.Besides the two techniques mentioned above,there are many other techniques which also can produce computer-generated spoofed speech,such as voice conversion.Therefore,the proposed countermeasures are also evaluated whether they are qualified to be generalized countermeasures for kinds of spoofing attacks or not.The work of this thesis can be divided into the following three aspects:(1)The spoofing countermeasures in recent years have been reviewed and summarized.Databases for identification of computer-generated spoofed speech have been carefully studied,especially for SAS.We focus on the subsets and the protocol in SAS,and the researches based on it.The experiments in the thesis are also based on SAS.(2)By wavelet decomposition on speech generated by HMM-based speech synthesis and natural speech,a difference has been noticed in their low frequency wavelet coefficients.A spoofing countermeasure based on the difference has been proposed.In the countermeasure,the detection feature is extracted from the low frequency wavelet coefficients and support vector machine is chosen as the classifier.The results show that the proposed countermeasure can successfully detect the speech generated by statistical parametric speech synthesis in SAS,with the average accuracy of 99.5% and the average equal error rate of 0.2%.Meanwhile,it has been evaluated whether it is qualified to be a generalized countermeasure for other kinds of spoofing attacks or not.(3)For speech generated by unit selection,a difference in high-frequency amount of information between the speech and natural speech has been found.A spoofing countermeasure based on the difference has been proposed.In the countermeasure,the detection feature has been extracted from the high-frequency information.Support vector machine and Gaussian mixture model are used as classifiers.The proposed countermeasure can successfully detect the speech generated by unit selection in SAS,with the average accuracy of 97.1% and the average equal error rate of 4.3%.Meanwhile,it has been evaluated whether it is qualified to be a generalized countermeasure for other kinds of spoofing attacks or not.The experimental results show that,the detection feature based on amount of information can efficiently discriminate all kinds of spoofed speech in SAS from natural speech.And it outperforms other features related to the research.