基于改进MFD的I-Vector说话人识别

发布时间：2017-12-28 20:43

本文关键词：基于改进MFD的I-Vector说话人识别　出处：《电声技术》2016年12期 　论文类型：期刊论文

【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。
[Abstract]:The speaker recognition system based on I-Vector usually uses LDA for channel compensation and feature reduction. In the case of limited sample development set, the LDA is not strong. Based on this, an improved I-Vector speaker recognition algorithm is proposed. When the number of speakers is less, the median I vector instead of the mean I vector can reduce the loss of the discriminative information. As the number of samples increases, the median classifier is improved, and the method of seeking the mean after the maximum minimum value is used as the centralization trend of the I vector. After calculating the inter class and intra class dispersion matrix by this method, the channel compensation and dimensionality reduction for the I vector are carried out. Combined with the Gauss PLDA model, LDA and WCCN are used as the baseline system for simulation and comparison. The experimental results show that the proposed algorithm has good discriminative performance, and the performance of the proposed algorithm can be increased by about 3% compared with baseline in the limited number of speakers.
【作者单位】：浙江大学电气工程学院;
【分类号】：TN912.34
【正文快照】： 与二次验证码校验比较,生物特征验证更加安全。相对指纹认证、面部轮廓(人脸)认证、虹膜认证以及DNA认证等技术,语音信号获取容易,采集设备成本较低,说话人识别技术更加容易普及。说话人确认是说话人识别研究的一个重要分支,在语音门禁、电话银行、数据库访问等系统中具有广泛

【相似文献】