基于深度神经网络的说话人自适应方法研究  

Study of speaker adaptation method based on deep neural network

作　　者：古典 李辉[1] Gu Dian;Li Hui(Department of Electronic Science and Technology, University of Science and Technology of China, Hefei 230027, Chin)

机构地区：[1]中国科学技术大学电子科学与技术系,安徽合肥230027

出　　处：《信息技术与网络安全》2018年第4期60-64,共5页Information Technology and Network Security

摘　　要：在说话人自适应领域,基于深度神经网络(DNN)的说话人自适应方法已成为主流方法。学习性隐层单元贡献(LHUC)作为一种有效的无监督自适应方法比其他自适应方法有诸多优势,而自适应数据的稀疏问题却导致识别性能受到影响。为了降低数据稀疏所带来的影响,在原有LHUC技术的基础上,加入i-vector及多任务学习这两种方法分别与其相结合。在TEDLIUM语料库上进行实验,实验结果表明,两种融合方法都比原系统表现出了更好的性能,单词错误率(WER)分别相对降低了2.5%和1.9%。The approaches based on deep neural network（ DNN） have become popular in speaker adaptation field. As an effective unsupervised speaker adaptation approach,learning hidden unit contributions（ LHUC） has many advantages over other approaches,but the sparse adaptation data may cause the recognition performance degradation. In order to reduce the influence of this problem,i-vector based approach and multitask learning are added to combine with LHUC respectively. Experiments are carried out on TED-LIUM corpus and results show that the two fusion approaches perform better than original system. The reduction of word error rate（ WER） is 2. 5% and 1. 9% respectively.

关 键 词：说话人自适应 LHUC i-vector 多任务学习 

分 类 号：TN912.34[电子电信—通信与信息系统]