频率规整线性预测系数的实时提取算法研究  

Research on real-time extraction of warped linear predictive coding coefficients

作　　者：侯传宇[1] 吴小培[2] 

机构地区：[1]宿州学院信息工程学院,安徽宿州234000 [2]安徽大学计算智能与信号处理教育部重点实验室,合肥230039

出　　处：《计算机工程与应用》2014年第9期214-218,共5页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61271352);安徽省高校优秀青年基金项目(No.2012SQRL199ZD);安徽省教育厅自然科学基金项目(No.KJ2013B285;No.KJ2012Z400;No.KJ2012Z401)

摘　　要：在语音识别特征提取过程中,为克服传统自相关法在计算特征参数时实时性较差的缺点,提出一种用于提取频率规整线性预测系数(WLPC)的自适应最小均方误差(LMS)算法。该方法通过自适应LMS技术,不仅能提取出符合人耳的听觉特性的特征参数,而且实现了对WLPC系数的实时提取。实验采用DTW(动态时间规整)算法,对比了自相关法WLPC预测误差和自适应法WLPC两种特征参数对孤立词识别率的影响结果和预测误差,结果证明了采用该算法具有较高的分类准确率和良好的时间性能。To overcome the disadvantages that the traditional autocorrelation algorithm has a poor performance in real-time extraction, an adaptive Least Mean Square（LMS）algorithm which is used to extract WLPC coefficients is presented. Based on an adaptive LMS algorithm, the proposed algorithm not only realizes the real-time extraction of the feature parameters which accord with characteristics of human hearing, but also extracts WLPC coefficients in real-time. A speech recognition model based on DTW algorithm is used to estimate the performance of autocorrelation algorithm and adaptive LMS algorithm. The experimental results demonstrate that the proposed algorithm has high classification accuracy and good real-time performance.

关 键 词：频率规整线性预测(WLPC) 最小均方误差(LMS)算法 自相关 实时 

分 类 号：TN912.3[电子电信—通信与信息系统]