Research On Voiceprint Verification Algorithms Based On Different Spoken Ways

Voiceprint recognition is an important research direction in the field of speech signal processing.At present,the most important application is to obtain speaker identity in speech,which is widely used in banking,public security systems and smart homes.Although voiceprint recognition technology has been developed rapidly,more and more voiceprint recognition technology is being continuously applied to actual intelligent voice-related products and services,but in practical applications,the existing technology still faces many challenges.Attacking the voiceprint recognition system such as synthesized speech and recorded replayed speech is one of the most important challenges.On the other hand,the phenomenon of voiceprint drift that does not match the acoustic environment of the registered and test voices will also seriously degrade the performance of the voiceprint recognition system.Based on the above problems,this paper focuses on the task of voiceprint recognition in the field of voiceprint recognition,and explores the robustness of voiceprint features from the perspective of the speaker's different utterance methods.First,a voiceprint database of different utterance methods is constructed,and then based on We explored the performance of voiceprint verification system based on traditional Gaussian mixture model,dynamic time warping and deep learning.The main work and innovations are as follows:First,corpus of voiceprint verification based on different utterance methods was constructed.Because of the lack of the current open source corpus,there is no corpus with abnormal utterance mode speech for the construction of voiceprint verification system,so the construction of voiceprint verification corpus with diverse utterance modes is the basis of this article.Based on this corpus,the distinguishability of speech with different utterance methods for extracting speaker identity information is studied in different acoustic feature parameter space.The results show that,to a certain extent,the speech in the abnormal way of speaking is more distinguishable than the normal speech in confirming the speaker's identity information.Secondly,a voiceprint modeling system based on the traditional model is constructed.The feature extraction,feature preprocessing,dynamic time warping algorithm,general background model training method,adaptive process of maximum posterior likelihood estimation and its back-end scoring algorithm are introduced in detail.This paper discusses the performance of text-related and text-independent voiceprint verification tasks in traditional voiceprint recognition systems.Experimental results show that the dynamic time warping algorithm and Gaussian mixture model performs well in related voiceprint verification tasks.Thirdly,voiceprint verification system based on deep learning is constructed.The method of using delay neural network and intermediate statistical layer to directly extract the statistics of complete speech as sentence-level input is used to characterize the voiceprint information contained in speech.The experimental results show that the performance of this method surpasses the traditional voiceprint recognition system.Specifically,the system studied in this paper can achieve the same error rate as the normal speaking method on the corpus of utterances in various utterances.Even to a certain extent,the voice with abnormal utterance is more suitable for voiceprint modeling than the normal utterance.Based on the traditional theory of voiceprint verification,this paper proposes a study of voiceprint verification based on diversified utterance speech.The purpose of this study is to address the weakness of the voiceprint verification system that is vulnerable to attack under normal speaking speech conditions and to explore the effect of different acoustic utterances on the voiceprint system.The research content provides a reference for the diversity of the subsequent research on voiceprint verification,and solves some problems in voiceprint verification from a new perspective.