一种简单有效的说话人身份认证方法

采用生物的特征识别技术,对说话人识别中说话人确认与说话人辨认的传统方法与分类进行了讨论,对现在使用的各种说话人识别算法进行了综合分析。以LPCC(Linear Prediction coding Coefficient)和MFCC(MEL Frequency Cestrum Coefficient)两种特征参数提取为基础,对GMM(Gaussian MixtureModels),VQ(vector Quantization),DHMM(Discrete Hidden Markov Model),CHMM(Concatenation HiddenMarkov Model)等几种识别方法进行了实现,做到了真正的与文本无关。特别以GMM识别方法的部分实验结果为例进行分析,解决了系统中的阈值设置问题,改进了该方法的决策手段。根据实验数据,在各种方法中,说话人确认的错误拒识率和误识率相对说话人辨认总是较高,本文把说话人辨认的阈值选择方法应用于说话人确认,以多模板匹配方式为辅助,使得说话人确认的错误拒识率和误识率大大降低,并通过实验证明了该种改进方法的有效性。

Efficacious method of speaker verification

Speaker recognition, as a special form of biometric identification technology, is presented in the paper about it's two aspects, SI (Speaker Identification) and SV(Speaker Verification). We can see from the experiment data that the rate of the SV'FRR and FAR is always higher than that of SI's. Also, this paper analyses and implements algorithms in common use about SR ?such as GMM,VQ,DHMM,CHMM, and so on. And expounds the valve-setting way (the SV system uses the decision-making artifice of the SI system), declining the SV system's FRR(False rejection Rate) and FAR(False Acceptance Rate). With the new way and a help of the multi-model method, we have implemented a system, which is really of text independent. It is efficacious means through experiment.