Research On Way Of Speaking Reliability In Voiceprint Recognition

This paper studies the influence of vocal effort on the reliability of voiceprint recognition system from two aspects: speaker model and feature.Major efforts and contribution are:1.Through the analysis of speech at this stage of the study,select the vocal effort of speech in a representative as the main object of this article.Established a vocal effort database which suitable for the reliability voiceprint recognition.The vocal of speech had been proposed since 2010,vocal effort gradually got the attention of the researchers,a large number of literature research shows that the vocal effort has great influence on the voiceprint recognition system performance,vocal effort has research value.Reference to the 2010 National Institute of Standards and Technology(NIST)voiceprint recognition evaluation of the organizers to the participating units of the development library Tarball database settings,the database at the time of recording only consider the same person who speaker whisper,Shouted,normal three orders of magnitude.This database of 30 people involved in the recording,including 15 men,15 women.2.The acoustic characteristics of speech signals and the visualization of the model show that the vocal effort has the possibility of influencing the reliability of the voiceprint recognition system,and it is proposed that the phonetic characteristics of different vocal effort are a special independent subspace Hypothesis.Firstly,this paper analyzes the formant,fundamental frequency and spectrum of the speech under different vocal effort of the same speaker,and finds that the acoustic characteristics of the speech signal under different vocal effort are obviously different.Secondly,this paper,through the analyzes the position distribution of the model mean vector and the displacement of the relative position of the speech under the different effort vocal from same speaker.Founding that the model mean vector which belong to different effort vocal from same speaker are mixed with each obviously.Based on the above two points,it is proposed that the speech feature of different vocal effort is a special independent subspace.The hypothesis is also verified in the related experiment.3.Using the Maximum Likelihood Linear Regression(MLLR)model projection and Constrained Maximum Likelihood Linear Regression(CMLLR)feature space projection method projection transforming model features to improve the reliability of the voiceprint recognition system.When tested in reserve for speech,without reserve the speech under the different vocal effort,In this paper,we propose to use the MLLR model projection method and the CMLLR feature space projection method to training projection matrix in the development set of database.If the matrix training is effective,the MLLR model projection method will make the speaker model learn the distinguishing information of different vocal effort speech;The CMLLR feature space projection method will make the different vocal intensity information of the speaker test speech be weakened.Experiments show that the above two methods are effective to improve mismatch between the training speech and test speech.4.Based on the Maximum Posteriori Probability(MAP)adaptive method to update the speaker model the Maximum Posteriori Probability +Constraint Maximum Likelihood Linear Regression(MAP + CMLLR)method to update the speaker model,and projection transforming speaker characteristics The In this paper,we propose a method to update the speaker model by using the MAP adaptive method,so that the speaker model can learn the distinguishing information of different vocal effort speech.MAP + CMLLR method is to use the MAP adaptive method to update the speaker model at the same time,the use of CMLLR feature space projection method projection transform test speech,so that the speaker model in the study of different vocal effort of speech discrimination information at the same time,weaken the test speech in different vocal effort of speech discrimination information.The MAP + CMLLR toward the weakening and learning whisper,shouted speech separation of information in the middle of a balance point closer,when the balance point is reached,the two checks and balances,speaker distinction information will be prominent,the voiceprint recognition system performance can be improved,thereby enhancing the vocal effort under the influence of voiceprint identification system reliability.