基于改进MFD的I-Vector说话人识别  

Improving MFD technique for I-Vector based speaker recognition

作　　者：李湾湾 范承志[1] 祁才君[1] 

机构地区：[1]浙江大学电气工程学院,浙江杭州310027

出　　处：《电声技术》2016年第12期43-48,共6页Audio Engineering

摘　　要：基于I-Vector的说话人识别系统通常采用LDA进行信道补偿和特征降维,在开发集样本有限的情况下,LDA的区分性不强。基于此,提出一种改进I-Vector说话人确认算法。在话者样本数较少的情况下,以中值i向量代替均值i向量作为集中统计量可以减少区分信息的丢失。随着样本数量增加,改进中值分类器,采用去最大最小值后求均值的方法作为i向量的集中趋势。用此方法计算类间与类内离散度矩阵后,对i向量进行信道补偿和降维。结合高斯PLDA模型,以LDA和WCCN为基线系统进行仿真对比。实验结果表明,提出的算法具有良好的区分性能,在有限的话者语音样本数量范围内,与基线相比能提升总和约3%的性能。I-Vector based speaker recognition system usually uses LDA technique for channel compensation and feature di- mension reduction, in the case of limited session development data, the separation ability of LDA is not obvious. Aiming at this problem, an improving I-Vector speaker verification algorithm is proposed using MFD technique. By taking the median as the estimator for the central tendency, instead of the mean, the MFD approach helps to attenuate the loss. And then get rid of the maximum and minimum values for average in the case of more samples. This improved MFD estimation is performed by calculating the between-and within-class scatter estimations for channel compensation and dimension reduction. Combining PLDA model, the experiments on different session development data using the techniques mentioned above are conducted compared with that of LDA method. The results show obvious improvement in separating different speakers, and the EER has a 3% reduction within the overall development data.

关 键 词：说话人识别 i向量 PLDA模型 线性区分性分析 改进中值分类器 信道补偿 

分 类 号：TN912.3[电子电信—通信与信息系统]