基于改进信道补偿的I-vector说话人识别

Improving channel compensation for I-vector based speaker recogniton

说话人识别算法的准确率受多种因素影响,其中受到说话人语音信道影响较为明显,常用说话人识别方法通过使用MFCC提取说话人语音信息的特征参数,并通过端点检测技术处理说话人特征,建立说话人通用背景模型,即GMM-UBM模型。使用因子分析技术将模型中的语音特征高维超向量映射成固定长度低维矢量,得到身份认证矢量I-vector。利用信道补偿技术处理I-vector,可以提高识别准确率。文中提出一种改进的信道补偿技术,通过对身份认证矢量进行信道补偿和特征参数降维,在样本数较多的情况下,通过补充缺失类内类间信息,去除信道因子的影响,能够提高系统对说话人的区分度,并提高识别准确率。实验结果表明,改进的信道补偿算法和I-vector模型结合后,可以获得更好的识别准确率,较传统信道补偿技术准确率提高3%~5%。

The accuracy of speaker recognition algorithm is affected by many factors,among which,it is significantly affected by the speaker voice channel.The commonly used speaker recognition method extracts the characteristic parameters of the speaker voice information by using MFCC,and processes the speaker characteristics by endpoint detection technology,so as to establish the universal background model of the speaker,namely GMM-UBM model.Factor analysis was used to map the high-dimensional hypervectors of speech features in the model to fixed-length low-dimensional vectors,and the identity authentication vector I-vector was obtained.In this paper,an improved channel compensation technology is proposed.In the case of a large number of samples,channel compensation and characteristic parameter dimensionality reduction are carried out for I-vector vectors.By supplementing the inter-class information within the missing class and removing the influence of channel factors,the speaker discrimination degree of the system can be improved and the recognition accuracy can be improved.Experimental results show that the improved channel compensation algorithm combined with I-vector model can achieve better recognition accuracy,which is 3~5 percent higher than that of traditional channel compensation technology.