期刊论文

i-TDNN:An improved noise speaker recognition method based on TDNN

中文

长江信息通信

2096-9759

2023

36

2

27-30

10.3969/j.issn.1673-1131.2023.02.008

湖北省教育厅科技项目(B2020061)。

本校为第一机构

针对声纹识别任务在含噪背景下鲁棒性欠佳的问题,文章提出了一种基于TDNN改进的含噪声纹识别方法。该算法先提取说话人音频的梅尔频谱,利用自注意力机制(SE)使得网络更加聚焦于重要特征,引入残差连接(Res)修正梅尔频谱与输...展开更多 针对声纹识别任务在含噪背景下鲁棒性欠佳的问题,文章提出了一种基于TDNN改进的含噪声纹识别方法。该算法先提取说话人音频的梅尔频谱,利用自注意力机制(SE)使得网络更加聚焦于重要特征,引入残差连接(Res)修正梅尔频谱与输出层的特征损失信息,一定程度缓解神经网络退化的问题,使用多层特征聚合(MFA)密集连接输出特征,生成关注统计池的特征,最终生成一种强鲁棒性的...

To solve the problem that voice print recognition is not robust under background noise,this paper proposes an end-to-end Speaker Vector based on TDNN.Firstly,the algorithm extracts the Mahr spectrum of the speaker audio,and corrects the feature loss information of...MORE To solve the problem that voice print recognition is not robust under background noise,this paper proposes an end-to-end Speaker Vector based on TDNN.Firstly,the algorithm extracts the Mahr spectrum of the speaker audio,and corrects the feature loss information of the Mahr spectrum and the output layer with the residuals conne...