针对声纹识别领域中存在信道失配与对短语音或噪声条件下声纹特征获取不完全的问题，提出一种将传统方法与深度学习相结合的方法，以I-Vector模型作为教师模型对学生模型ResNet进行知识蒸馏。构建基于度量学习的ResNet网络，引入注意力统计池化层，捕获并强调声纹特征的重要信息，提高声纹特征的可区分性。设计联合训练损失函数，将均方根误差(Mean Square Error，MSE)与基于度量学习的损失相结合，降低计算复杂度，增强模型学习能力。最后，利用训练完成的模型进行声纹识别测试，并且与多种深度学习方法下的声纹识别模型相比，等错误率（Equal Error Rate，EER）是最低的，达到了3.229%，表明该模型能够更有效地进行声纹识别。

Channel mismatch and incomplete acquisition of voiceprint features under short speech or noise conditions are two thorny problems for voiceprint recognition. This paper proposes a solution that combines traditional techniques with deep learning: A I-Vector model was used as the teacher model to conduct knowledge distillation of the student model ResNet, a ResNet network based on metric learning was constructed, including an attentive statistics pooling layer to capture and emphasize the critical information of voiceprint features and improve the distinguishability of voiceprint features, and the mean square error (MSE) was combined with the loss based on metric learning to reduce computational complexity and enhance model learning capabilities. The trained model was then used for the voiceprint recognition test. Compared with the voiceprint recognition model under various deep learning methods, the equal error rate (EER) was the lowest, and the equal error rate reached 3.229%, indicating that the model can perform voiceprint recognition more effectively.

TP751

教育部-中国移动科研基金资助项目(MCM20180404)。