Abstract: In the text-dependent speaker recognition, the content of training and testing speech samples are same. Because of those speeches are same, the traditional model cannot resist the attacking of synthetic voice. This paper proposed an asymmetric text-dependent speaker recognition model. In the model, the training speeches are open access, but the testing speeches are secret for public and only the user know them. In this way, testing speeches cannot be synthesized by attackers. For improving the recognition performance, this model combines the wavelet and supervector. The wavelet can effectively analyze the non-stationary signal such as speech signal, and the supervector can improve the discrimination between different feature vectors. The experimental result shown that the proposed model can improve the recognition accuracy compared with the traditional models and can resist the attacking from the synthetic voice.

Key words: wavelet analysis, supervector, text-dependent speaker recognition, support vector machine, signal processing

摘要： 在文本相关的说话人识别模型中，训练语音和测试语音内容固定且相同。由于语音内容相同，这种模型将无法有效抵御“合成语音攻击”。本论文提出一种非对称文本相关识别模型。该模型中，训练语音和测试语音内容不同，且训练语音内容公开而测试语音内容由用户保留且不公开。这样就避免了测试语音被攻击者合成。同时，为了提高识别性能，小波分析和超级向量被引入该模型。小波分析能有效分析语音这种非平稳信号，而超级向量能有效提高不同特征向量间的区分度。实验结果表明新模型和传统模型相比可以提高识别性。

关键词: 小波分析, 超级向量, 文本相关说话人识别, 支持向量机, 信号处理