一种基于深度神经网络的话者确认方法

A SPEAKER VERIFICATION METHOD BASED ON DEEP NEURAL NETWORK

主要研究基于深度神经网络的话者确认方法。在训练阶段,以语音倒谱特征参数作为输入,说话人标签作为输出有监督的训练DNN;在话者注册阶段,从已训练的DNN最后一个隐藏层抽取与说话人相关的特征矢量,称为d-vector,作为话者模型;在测试阶段,从测试语音中抽取其d-vector与注册的话者模型相比较然后做出判决。实验结果表明,基于DNN的话者确认方法是可行的,并且在噪声环境及低的错误拒绝率的条件下,基于DNN的话者确认系统性能比i-vector基线系统性能更优。最后,将两个系统进行融合,融合后的系统相对于i-vector基线系统在干净语音和噪声语音条件下等误识率(EER)分别下降了13%和27%。

In this paper we mainly investigate the method of using deep neural network ( DNN) for speaker verification. At the stage of training, the DNN is trained under supervision using the feature parameter of speech cepstrum as input and the label of speaker as output. At the stage of speaker registration, an eigenvector correlated to the speaker, namely d-vector, is extracted from the last hidden layer of the trained DNN and is used as the model of speaker. At test stage, from testing speech a d-vector is extracted to compare it with the model of the registered speaker and then to make the verification decision. Experimental results show that the DNN-based speaker verification method is feasible. Moreover, under the condition of noisy environment and low error-rejection rate, the DNN-based speaker verification system outperforms the i-vector base line system in performance. Finally, we integrate these two systems, relative to the i-vector base line system,the integrated system reduces the equal error rate (EER) by 1 3 % and 2 7 % for clean speech and noisy speck conditions respectively.