摘要
在说话人识别的建模过程中,为传统矢量量化模型的码字增加了方差分量,形成了一种新的连续码字分布的矢量量化模型。同时采用美尔倒谱系数及其差分和线性预测倒谱系数及其差分相结合作为识别的特征参数,来进行与文本有关的说话人识别。通过与动态时间规整算法和传统的矢量量化方法进行比较表明,在系统响应时间并未明显增加的基础上,该模型识别率有一定提高。
In the process of feature extraction of a text-dependent speaker recognition system, the difference of Mel Frequency Cepstrum Coefficient(MFCC) and Linear Prediction Cepstrum Coefficient(LPCC) was chosen to be the speech characteristic parameters, and in the process of speech modeling, a variance was added to the code word of Vector Qantization (VQ) and got continuous vector quantization, then compared it with Dynamic Time Warping(DTW) method and VQ method in text-dependent speaker recognition experiment. The results of identification show that the recognition efficiency is proved without any obvious increasing of responds time.
出处
《计算机应用》
CSCD
北大核心
2006年第4期883-885,共3页
journal of Computer Applications
基金
河北省教育厅博士基金资助项目(B2003202)
关键词
说话人识别
线性预测倒谱系数
美尔倒谱系数
矢量量化
动态时间规整
speaker recognition
Linear Prediction Cepstrum Coefficient(LPCC)
Mel Frequency Cepstrum Coefficient (MFCC)
Vector Qantization(VQ)
Dynamic Time Warping(DTW)