论文部分内容阅读
鉴别性Mel频率倒谱系数(DMFCC)是一种修正的Mel频率倒谱系数(MFCC),其更加强调语音频谱各个子带携带的鉴别性信息,采用自适应的非均匀的滤波器组设置。在宽带信号应用中,DMFCC的作用和效果已经被证明;但在窄带信号应用中,DMFCC还鲜见有成功应用的例子。该文在电话信道下对应用DMFCC进行说话人识别研究,在美国国家标准技术研究院(NIST)2006年说话人识别评测Female核心测试集上,以MFCC作为特征参数的系统的等错误率为7.57%,以DMFCC作为特征参数的系统的等错误率为7.25%,而采用基于逻辑自回归的线性融合方法把基于两种不同特征的系统在分数域进行融合后系统的等错误率可达到6.31%,相对于基于MFCC的系统等错误率下降16.6%。实验表明,在电话信道下直接应用DMFCC可小幅度提高性能;理论分析以及实验结果表明:二者存在一定的互补性,即把DMFCC和MFCC融合应用能够大幅度提高电话信道下说话人识别的性能。
Discriminant Mel Frequency Cepstral Coefficients (DMFCC) are a modified Mel Frequency Cepstral Coefficients (MFCC) that emphasize more discriminative information carried by the individual subbands of the speech spectrum, using an adaptive, non-uniform set of filters. The role and effect of DMFCC have been demonstrated in wideband signal applications; however, there are few examples of successful applications of DMFCC in narrowband signal applications. In this paper, speaker identification based on DMFCC is applied under the telephone channel. At the National Institute of Standards and Technology (NIST) speaker evaluation test Female Core Test Set in 2006, the system with MFCC as the characteristic parameter has an equal error rate of 7.57 % And DMFCC as the characteristic parameter is 7.25%, while the linear fusion method based on logic autoregressive method converges the system based on two different features in the fractional domain with the same error rate of 6.31% , Compared with MFCC-based systems such as the error rate decreased by 16.6%. Experiments show that the application of DMFCC directly on the telephone channel can improve the performance slightly. The theoretical analysis and experimental results show that there is a certain complementarity between the two, that is, the application of DMFCC and MFCC can significantly improve the performance of speaker recognition under the telephone channel .