论文部分内容阅读
该文以掌上电脑录音的语音数据库为对象进行与文本无关的说话人识别实验,通过对识别算法的改进,探索有效消除强噪声所带来的影响,提高说话人识别系统的鲁棒性、稳定性与准确性的方法.1、在说话人辨认系统中,以LPC倒谱参数为特征矢量来描述说话人语音和个性特征,用高斯混合模型(GMM,Gaussian Mixture Model)描述特征矢量的分布,利用EM算法估计高斯混合模型(GM)参数,比较了特征参数的协方差矩阵取为不同类型时,EM算法估计高斯混合模型参数的有效性及其对系统识别率的影响.实验表明,当特征矢量协方差矩阵取为满矩阵时,可以对模型参数进行更有效估计,从而有效抵消噪声影响,提高说话人辨认的正确识别率.2、在说话人确认系统中,用LPC倒谱参数为特征矢量表征说话人语音和个性特征,采取两种不同的识别算法进行实验.1)考虑在不同的环境及声道特性下,语音特征具有一定的非平稳性,因此使用假冒说话人集标准化法,首先计算被测试者的标准化似然分值,再根据所给定的门限阈值,由说话人确认系统作出是与否的判决,以此研究与文本无关的说话人确认系统的有效性.2)摒取庞大的假冒说话人背景集,以假设检验理论为基础,在训练阶段由说话人确认系统依据训练集自行确定最佳判决阈值.在此基础上,研究在强噪声背景下,如何实现与文本无关的说话人确认系统的低错误拒绝率(FRR,False Rejection Ratio)和低错误接受率(FAR,False Acceptance Ratio).实验结果显示,以假设检验分析为基础的识别算法比使用假冒说话人集标准化法的识别算法更加有效,两者相比,使用基于假设检验的识别算法,说话人确认系统可以同时获得低错误拒绝率(FRR)和低错误接受率(FAR).有效提高系统的抗噪声性能.该文所述的方法均是在Microsoft Windows 2000和Microsoft Windows ME平台下使用Matlab 6.1编程实现,除估算高斯混合模型(GMM)参数的EM算法外,文中所用的算法均为作者在导师的指导下使用Matlab 6.1编写.关键词:说话人识别 线性预测分析 高斯混合模型 EM算法假冒说话人集 假设检验Abstract Speaker recognition (SR) is a method of identifying and verifying someoneidentity by his speech. It is an important part of speech signal processing. SR isapplied in many realm, including voice dialing, telephone bank, telephone shopping,database speech accessing, information service, speech E-mail, safety control and computer long-distance login. The hand-held PC (PDA, Personal Digital Assistant) is a convenient and movable high-tech electronic device. Its functions include information inputting,accessing, managing and delivering. In addition, it has general computer function,business affair handling, amusement and movable communication. Due to the limitation of PDA volume and hardware configuration, there is much strong different noise mixing among recorded speech or voice compared with general computer or special recording devices. Therefore, a challenging problem is how to remove the strong noise among speech when we process speech or voice recorded by PDA. In this paper, text independent speaker recognition experiment is performed based on PDA speech database. By improving recognition algorithm, we try to find a way to eliminate the strong noise and to improve speaker recognition system robustness, stability and veracity. Firstly, LPC cepstmm feature vector parameters and GMM are adopted in speaker identification system. GMM parameters are estimated by EM algorithm. And its recognition results are compared by adopting different feature parameter covariance matrix. The results indicate when the type of feature parameter covariance matrix is full, GMM parameters estimation is more effective. It can eliminate more noise influence and improve speaker recognition ratio. Secondly, LPC cepstrum feature vector parameters and GMM are adopted in speaker verification system. Two recognition algorithms are used in the experiments.The first method uses impostor speaker sets normalization, considering different environment, acoustic track feature and speech random speciality. To begin with, a judgment of speaker verification system is made, according to testers normalized probability value and given threshold value. And followed with speaker verification.The second method is based on hypothesis test theory. In training phase, the system produce an optimal threshold value automatically. Then we further improve system to achieve low false rejection ratio and low false acceptance ratio in the strong noise background. The experiments indicate the latter recognition method is more effective than the former. Moreover low false rejection ratio and low false acceptance ratio are achieved at the same time by the second method.All experiments discussed in this paper performed by Matlab 6.1 toolkit under Microsoft Windows 2000 and Microsoft Windows ME environment. Except EM algorithm, other algorithms in the paper are programmed by matlab 6.1 toolkit with the tutors direction.Keywords:Speaker recognition, Linear predictive analysis, GMM,EM algorithm, Impostor speakers sets, Hypothesis test