噪声环境下稳健的说话人识别特征研究  

A study of robust speaker recognition feature under noisy environment

作　　者：程小伟[1] 王健[1] 曾庆宁[1] 谢先明[1] 龙超[1] CHENG Xiao-wei WANG Jian ZENG Qing-ning XIE Xian-ming LONG Chao(School of Information and Communication, Guilin University of Electronic Technology, Guilin 541004, Guangxi, China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《声学技术》2017年第5期479-483,共5页Technical Acoustics

基　　金：国家自然科学基金项目61461011;教育部重点实验室2016年主任基金项目资助CRKL160107;广西自然科学基金2014GXNSFBA118273项目

摘　　要：针对噪声环境下说话人识别率较低的问题,提出一种基于正规化线性预测功率谱的说话人识别特征。首先对语音信号线性预测分析和正规化处理求出语音频谱包络,然后通过伽马通滤波器组得到对数子带能量,最后对特征参数进行离散余弦变换,得到了一种说话人识别特征正规化线性预测伽马通滤波器倒谱系数(Regularized Linear Prediction Gammatone Filter Cepstral Coefficient,RLP-GFCC)。仿真结果表明,在噪声环境说话人辨认试验中,相比传统特征美尔频率倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)和伽马通滤波器倒谱系数(Gammatone Filter Cepstral Coefficient,GFCC)的系统识别率得到了明显提高,对噪声环境的鲁棒性得到了增强。In order to solve the problem that speaker recognition rate is low under noisy environment, a speaker recognition feature based on regularized linear predictive power spectrum is proposed. The method uses linear prediction analysis and regularization of speech signal to get speech spectral envelope and then to get logarithmic subband energy through the Gammatone filter group, and finally uses discrete cosine transform to compute feature parameters to get a kind of speaker recognition feature named regularized linear predicted Gammatone filter cepstral coefficients RLPGFCC . The simulation results show that the recognition rate of the system is significantly improved in comparison with the systems of traditional feature MFCC and GFCC under noisy environment, and the robustness of the system to noise environment is improved.

关 键 词：线性预测 正规化 说话人识别 伽马通滤波器组 鲁棒性 

分 类 号：TN912.3[电子电信—通信与信息系统]