Research Of The Key Techniques For Voice Playback Detection

With the development of society,the application of biometric technology is becoming more and more common.Voiceprint recognition,fingerprint recognition,iris recognition,face recognition and other living detection technology products have been available and sought after by people.Such technologies not only improve the efficiency of the relevant personnel,but also greatly meet the people's information security requirements.Among them,the voiceprint recognition,also known as the speaker recognition technology,according to the scope of its use is divided into the speaker identification and speaker confirmation of two categories.Speaker recognition technology,with the advantage of obtain convenient,the low price of pickup equipment,and support the site identification,has been gradually used in the fields of finance,social security,the common life.With the popularity of high fidelity and portable recording device,it becomes very easy for the attacker to get the voice by recording device.When the attacker obtained the specific voice,it becomes very easy for the attacker to spoofing speaker verification system by voice playback.In order to avoid such security problem,the system must prevent an attacker get permission by using the playback voice.At present,the research team has proposed detection algorithm for voice playback.The existing algorithm,however,neglected the actual attack possibilities,Such as the diversity of eavesdropping devices,playback devices,only consider a single device.So in this paper,the key technologies and research achievements of current playback speech detection algorithm were summarized,and carry out the related research work on the basis of the above-mentioned problems and the diversity of the development of science and technology.The aim of the paper is to provide voice playback detection algorithm for relevant technical.The paper mainly carry out the research work in three aspects,the performance of voice playback,its concept,the construction of database and the voice playback detection algorithm.Firstly,in this paper,the playback speech is tested on GMM-UBM system and OPPOfind7 platform respectively.From the results,we can see that the playback voice is easy to attack such listing of the voiceprint recognition system;the results of this experiment further prove the importance and necessity of studying the voice playback detection.In addition,this paper also explains the playback voice in the aspects of waveform and speech spectrum,and has reached the preliminary understanding of the playback speech and paves the way for the corresponding playback speech detection algorithm.Secondly,On the one hand there is no recognized correlation database so far.Another important factor is to improve the limitations of existing literature on sample setup.According to the research needs of this paper,the required database will involve a variety of different recording devices,playback devices,and eavesdropping devices.So in the actual research process,we spend a lot of time and effort in the construction of the database.The constructed database contains two types of acquisition devices three playback device or five eavesdropping devices,formed a total of two kinds and 30 kinds of original speech playback voice different sources.Thirdly,in this paper,two kinds of statistical features are proposed based on the change of frequency distribution and coding parameters.In this paper,through the in-depth analysis of the production process of playback voice,that eavesdropping devices and playback devices will inevitably lead to quantitative distortion and coding distortion,so that playback of voice-related equipment to carry information.If the measured wav formatted voice is encoded by MP3,it is found that there is a difference between the original voice and the playback speech in the long window scale factor,and a detection algorithm based on the long window scale factor is proposed.The experimental results show that the algorithm can distinguish between playback voice and original voice,the ACC is more than 99%,when the code rate is 64 kbps,classifier using Libsvm.According to the difference between the playback voice and the original voice at different frequencies,a detection algorithm based on the information quantity distribution is proposed.The experimental results show that the algorithm can also achieve a good classification of the original voice and playback voice.When the Fourier transform points are 512,the detection result is more than 99%.Loaded on the GMM-UBM system,the proposed algorithm greatly improves the system's ability to resist the playback of voice attacks.