Research On Speech Voiceprint Password Authentication Technology

Speech voiceprint password authentication is a technique which uses text information and speaker information of the speech segments for user account information encryption. So it has a good safety and convenience which can be applied in many domains like banking, police, smart home, and so on. But in the real application, speech voiceprint password authentication is still faced the challenge of password leak, feature of redundancy and low ability to resist anti-noise.The traditional speech voiceprint password authentication technology belongs to the text-dependent voiceprint recognition task. Due to the password text is fixed, password is easy to forget and leaking, so security is not good. To solve these shortcomings, this article proposed the text-prompt voiceprint password identification. Every time when the user logins the system, the system prompts password text dynamically, then user speaks password out according to prompt text. Although dynamic password has high security, dynamic voiceprint password recognition belongs to text-independent voiceprint verification task, which has relatively low recognition performance. In view of the above problems, this article improves the text-prompt voiceprint password recognition performance from the following several aspects mainly.First of all, the speech voiceprint password recognition contains two parts: speech password verification and voiceprint verification. So, the front of system needs a relatively high recognition rate of password verification system to verify the user speech password is correct or not. The recognition performance of speech recognition system based on GMM-HMM is relatively low, which is difficult to meet the security requirements. So, this article uses the DNN-HMM speech recognition system which has higher recognition rate as the speech password identification system.Second, the acoustic features (such as MFCC, PLP, etc.) in traditional voiceprint verification mainly include text information and channel information, speaker information belongs to the weak information. The performance of voiceprint verification is easily affected by text information, the channel information and noise interference in the speech signals. To remove these interference informations in speech feature, this paper proposes a method of speaker information extraction based on deep neural network for ASR, which takes advantage of the feature extraction ability of deep neural network. The speaker information extracted by this method has better speaker discrimination ability than the traditional acoustic features.Third, to remove the redundant information in acoustic feature, this article further uses the acoustic factor analysis method to remove the redundancy in the acoustic feature. The traditional acoustic factor analysis uses factor analysis method for dimension reduction on each component of the GMM. But the GMM belongs to unsupervised clustering algorithm, each Gaussian component of GMM has no defining acoustic meaning. So, this article replaces GMM with DNN of acoustic model in acoustic factor analysis and derives a phoneme dependent dimensionality reduction of feature in every phoneme subspace to extract speaker information which is used to extract the DNN i-vector. Furthermore, in the speaker information extraction based on DNN in the chapter 3, this article uses acoustic factor analysis based on DNN to replace LDA for dimension reduction of the hidden layer output supvector.Last, according to the characteristic of text-prompt voiceprint password identification, this article puts forward the digital modeling voiceprint identification system. For each number in digital speech segments trains a voiceprint recognition model respectively. Matching the digital appearing at the same time in enrollment speech and test speech when testing, this method converts the text-independent voiceprint recognition task into text-dependent voiceprint recognition task.The speech database this article used is RSR2015 speech database. Through the experiments on the RSR2015 speech database demonstrate the effectiveness of the above method.