关键词: 关键词: 说话人验证, 数据增强, 频率掩蔽, 注意力机制, 欺骗语音检测

Abstract: Abstract: Aiming at the problem that the convolutional model-based anti-spoofing speaker verification method cannot capture global feature dependency well, an anti-spoofing speaker verification method utilizing global-local feature dependency was proposed. First, for the spoofing speech detection module, two filter combinations were designed to filter the original speech, and sample augmentation was achieved by masking the frequency sub-bands. Second, a multidimensional global attention mechanism was proposed, where the global dependencies of each dimension were obtained by pooling the channel dimension, frequency dimension, and time dimension, respectively, and the global information was fused with the original features by weighting. Finally, for the speaker verification part, a Statistical Pyramid Dense Time Delay Neural Network (SPD-TDNN) was introduced to compute the standard deviation of the features and join the global information while obtaining the multi-scale time-frequency features. The results show that the proposed spoofing speech detection method reduces the equal error rate by 53% on the ASVspoof2019 dataset compared to the Audio Anti-Spoofing using Integrated Spectro-Temporal graph attention network (AASIST). The proposed anti-spoofing speaker verification method reduces the equal error rate by 23 percentage points compared to the statistical pyramid dense time delay neural network method. It has been verified that the proposed method achieves better classification results with the help of global feature dependency.

Key words: Keywords: speaker verification, data augmentation, frequency masking, attention mechanism, synthetic speech detection