深度神经网络自适应中基于身份认证向量的归一化方法

发布时间：2018-01-18 00:17

本文关键词：深度神经网络自适应中基于身份认证向量的归一化方法　出处：《中国科学院大学学报》2017年05期 　论文类型：期刊论文

【摘要】：深度神经网络是近年来非常流行的一种语音识别声学建模技术,其性能比之前主流的高斯混合模型有显著提高,但是深度神经网络的说话人自适应技术一直没有很好地解决。利用身份认证向量对深度神经网络进行自适应,并研究身份认证向量归一化对系统的影响,提出一种新的max-min线性归一化技术。实验表明在TIMIT数据集上该技术可使字错误率比传统方法相对下降5.10%。
[Abstract]:Depth neural network is a very popular acoustic modeling technology for speech recognition in recent years, its performance is significantly improved than the former mainstream Gao Si hybrid model. But the speaker adaptation technology of depth neural network has not been well solved. The identity authentication vector is used to adapt the depth neural network, and the effect of identity authentication vector normalization on the system is studied. In this paper, a new max-min linear normalization technique is proposed. The experimental results show that this technique can reduce the error rate by 5.10% on the TIMIT data set.
【作者单位】：清华大学电子工程系;
【基金】：国家自然科学基金(61370034,61403224)资助
【分类号】：TN912.34
【正文快照】： Yang J B,Zhang W Q,Liu J.Investigation of normalization methods in speaker adaptation of deep neural networkusing i-vector[J].Journal of University of Chinese Academy of Sciences,2017,34(5):633-639.近几年来,深度神经网络(deep neuralnetwork,DNN)在语音识别

【相似文献】