噪声环境下基于特征信息融合的说话人识别  

Improve Speaker Identification Performance by Integrating Characters under Noisy Conditions

作　　者：叶寒生[1] 陶进绪[1] 张东文[1] 余斌[1] 

机构地区：[1]中国科学技术大学电子工程与信息科学系,安徽合肥230027

出　　处：《计算机仿真》2009年第3期325-328,共4页Computer Simulation

摘　　要：针对在干净的语音环境下说话人识别率很高,但噪声环境下说话人识别率急剧下降的问题,提出了一种在噪声环境下,利用信噪比权重对说话人的特征信息MFCC系数和基音周期进行非线性融合,同时对MFCC特征参数进行基于帧信噪比权重得分,并同传统的高斯混合模型算法和基于F0-MFCC联合分布的特征融合方法,在噪声环境下分别进行了说话人识别的性能比较,同时对提出的融合算法进行了仿真实现。实验结果表明:在噪声的环境下方法相比上述传统说话人识别方法,性能有了明显的提高,在干净的语音环境下性能相当。The performance of speaker identification in clean speech circumstance is excellent, but it is degraded very rapidly under noisy conditions. To deal with the problem, this paper integrates the characters of MFCC coefficients and pitches by the signal noise ratios （SNR） of the speech, at the same time, weights the MFCC likelihood scores for different observation vectors based on the signal noise ratio of corresponding speech frames. The paper also compares this approach with traditional combination of MFCC and pitches. Simulation and experimental results indicate that, under the noisy conditions, the proposed scheme is superior to the traditional methods, and in clean circumstance, the performance is comparable.

关 键 词：说话人识别 信噪比 美尔倒谱系数 基音周期 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]