跨域注意力特征融合的说话人确认方法

Speaker verification method based on cross-domain attentive feature fusion

针对目前说话人确认系统中前端特征的语音信号样点间结构信息缺失问题,提出了跨域注意力特征融合的说话人确认方法。首先,提出了一种基于图信号处理的图频域特征提取方法来有效利用语音信号的结构信息,将语音信号帧的每个样点作为图节点,构建语音图信号,通过图傅里叶变换以及滤波器组提取图频域特征。其次,提出了一种由残差模块与挤压-激励模块构成的注意力特征融合网络,对传统时频域特征与图频域特征进行跨域融合,来提升说话人确认系统的性能。最后,在VoxCeleb、SITW和CN-Celeb数据集上进行实验。实验结果表明,所提方法在等错误率以及最小检测代价函数的评价指标上,优于基线模型ECAPA-TDNN。

Aiming at the problem that the lack of structure information among speech signal sample in the front-end acoustic features of speaker verification system,a speaker verification method based on cross-domain attentive feature fusion was proposed.Firstly,a feature extraction method based on the graph signal processing(GSP)was proposed to extract the structural information of speech signals,each sample point in a speech signal frame was regarded as a graph node to construct the speech graph signal and the graph frequency information of the speech signal was extracted through the graph Fourier transform and filter banks.Then,an attentive feature fusion network with the residual neural network and the squeeze-and-excitation block was proposed to fuse the features in the traditional time-frequency domain and those in the graph frequency domain to promote the speaker verification system performance.Finally,the experiment was carried out on the VoxCeleb,SITW,and CN-Celeb datasets.The experimental results show that the proposed method performs better than the baseline ECAPA-TDNN model in terms of equal error rate(EER)and minimum detection cost function(min-DCF).