Speaker Recognition With Short Utterances Based On Multi-core SVM-i-vector And Its Application In The Inspection System

As a kind of biometric technology,speaker recognition technology is widely used in the field of biometrics because of its convenience,non-interactivity and accuracy.In practical application,the recognition accuracy of existing speaker recognition algorithms is often greatly affected,due to different factors such as channel interference,environmental noise and speaker speech length.For the speaker application scenario with short utterances,this thesis proposed using extracted speaker i-vector to classify and judge with multi-core SVM,in order to improve the recognition accuracy of the short utterances,and apply it to the inspection system for testing.First of all,this thesis researched speech signal processing technology,it selected MFCC feature parameters which is matching human ear characteristics.In addition,it did in-depth research and experimentation on i-vector voiceprint recognition technology,it trained the GMM-UBM model and got the i-vector of speaker from the model,it compensated channel by LDA.Second,this thesis researched on multi-core SVM technology,designed a speaker recognition system with short utterances based on multi-core SVM-i-vector,experimented with different kernel functions,got the parameters of each kernel function when they make the system recognition rate the highest,and performed a variety of linear combinations of kernel functions based on these parameters,selected a combination by experiment which makes the system recognition rate highest,compared the system recognition rate of this combination with the system recognition rate which use using the single-core SVM and not using the SVM,and experimented with different speech durations and noise environments.The experimental results show that using multi-core SVM to classify i-vector feature vectors can obtain better system recognition rate,and the shorter the speech duration,the more obvious the recognition rate is improved,shows that it is better suited for the short utterances application scenario.Then,this thesis applied speaker recognition technology with short utterances based on multi-core SVM-i-vector to the inspection system,designed system overall architecture and system flow,and tested the system through experiments.The test results show that the system has good stability and feasibility.Finally,this thesis summarized the tasks of this paper,and analyzed the shortcomings of the system,look forward to the follow-up research direction.