Text-independent Speaker Recognition Method And System Based On Spatial Distribution Of Speech Features

With the widespread application of smart devices in life,the natural and smooth human-computer interaction has attracted people's attention.The ways of identification have gradually changed from passwords and other methods to biometric technique.Speaker recognition is a type of biometric recognition which uses voice for identity recognition Speaker recognition has the characteristics of convenient data collection,good user experience and high recognition accuracy.It is widely used in judicial identification,remote identity recognition and smart home scenariosIn speaker recognition,the registration process extracts the acoustic features of speech to construct voiceprint models,and the recognition process makes judgments based on the matching scores of the features of the sample to be recognized and the voiceprint models.It is difficult to construct an efficient voiceprint model with a small amount of registered speech In the case of text-independent speaker recognition,the difference in the text content of speech will also affect the recognition performance.In the thesis,under the framework of speaker recognition based on the spatial distribution of speech features,the methods of reducing the impact of registration and recognition speech text differences on recognition performance are studied.The main work and contributions are as follows1.A speaker recognition method based on the spatial distribution of speech features is proposed.The method includes three steps:speech feature space construction,speaker registration(using the speech feature space to calculate the spatial distribution of acoustic feature sequences)and speaker recognition(using the speech feature space to calculate the acoustic feature distribution of the samples to be recognized and makes a score decision).The method is based on the idea of feature space localization and uses the spatial distribution of features to model the voiceprint,which realizes lightweight sample modeling and speaker differentiation information expression.And the training data is easy to collect2.Verified the feasibility of the method.The method can achieve a recognition accuracy rate of 0.90 on the corpus of 400 speakers.Theoretical analysis and experimental exploration of steps and advantages of the method have been done,including the construction plan of the speech feature space,the selection of feature neighborhoods,the definition of the degree of association,the impact of speech duration on performance and the sharing of the speech feature space3.Two strategies are proposed to reduce the effect of registration and recognition speech text differences on speaker recognition performance.One is using the linear discriminant analysis to transform the feature distribution vector to reduce the influence of the speech text content on the speaker characteristics.On the corpus of 400 speakers,the recognition accuracy rate is improved to 0.95.The second is designing a minimum registered text that includes all finals,to improve the spatial coverage of the voiceprint feature of the registered text.It reduces the probability of spatial mismatch between the recognition sample and the registration sample and achieves the purpose of improving speaker recognition performance Combining the two strategies for experimental research4.A text-independent speaker recognition system is implemented.The system includes speaker registration,voiceprint model storage and speaker recognition.It provides speaker registration and speaker recognition module functions for secondary development.Through the measurement of the system,the speaker recognition system has a higher recognition accuracy rate for registered users and a higher detection rate for unregistered users Instructions for using the system are provided for users.