基于i-vector和深度学习的说话人识别

摘要

为了提高说话人识别系统的性能,在研究基础上提出了一种将深度神经网络(Deep Neural Nerwork,DNN)模型成果与i-vector模型相结合的新方案.该方案通过有效的神经网络构建,准确地提取了说话人语音里的隐藏信息.尽管DNN模型可以帮助挖掘很多信息,但是i-vector特征并没有完全覆盖住声纹特征的所有维度.为此,在i-vector特征的基础上继续提取维数更高的i-supervector特征,有效地避免了信息的不必要损失.为证明提出方案的可行性,采用对TIMIT等语音数据库630个说话人的语音进行了训练、验证和测试.验证实验结果表明,在提取i-vector特征的基础上提取i-supervector特征的说话人识别同等错误率有30%的降低,是一种有效的识别方法.%To improve the performance of speaker recognition systems,a novel scheme combined DNN (Deep Neural Network) model with the i-vector model has been proposed.Via construction of network,the hidden information in the voice of speakers has been extracted accurately.Although DNN model can help dig a lot of information,the i-vector features have not completely cover all dimensions of voiceprint.Thus i-supervector characteristics of higher dimension have been drawn with the i-vector features,which have effectively avoided the unnecessary loss of information.Experiments on TIMIT and other speech databases which contain 630 the speaker''s voices for training,validation and testing have been conducted to verify the proposed scheme.The results illustrate that the i-supervector features with i-vector features for speaker recognition have achieved 30% reduction of equal error rate that implies effectiveness of the identification method proposed.