基于极大似然线性回归的模型合成和特征映射进行说话人确认

Speaker verification using speaker model synthesis and feature mapping based on maximum-likelihood linear regression

提出了基于极大似然线性回归(MLLR)调整的说话人模型合成和特征映射方法。MAP调整事后确定相应模型间线性关系,变换参数人为确定;而MLLR调整首先确定相应模型间线性关系,变换参数由训练数据确定,并且可以只调整均值向量。模型合成时,MLLR调整指定通用信道背景模型参数间的线性变换;特征映射时,MLLR调整指定Root GMM-UBM与通用信道背景模型参数间的线性变换。通过对模型参数进行分组调整,可以在训练数据和参数数目间达成平衡。实验结果表明,合适选取MLLR回归类,可以取得比相应MAP调整方法更好的识别效果。

This paper proposes new methods of speaker verification,which use speaker model synthesis（SMS） and feature mapping based on maximum-likelihood linear regression.MAP method determines a linear relationship among the corresponding models after adjustment and transformation parameters are determined artificially,while MLLR first identify a linear relationship among the corresponding models and transformation parameters are determined from the training data,also it can only adjust the mean vectors.In SMS,MLLR determines transformation parameters among different channel UBMs.In feature mapping,MLLR determines transformation parameters between Root GMM-UBM and the channel UBM.By grouping to the model parameters,it can reach a balance between the training data and the number of parameters.The experimental results show that MLLR adjustment can achieve better verification effect than MAP adjustment by selecting the appropriate classes of regression.