| Summary: | 碩士 === 國立中山大學 === 資訊工程學系研究所 === 107 === In this paper, we implement a speaker verification system that can detect converted speech attack through combining representation learning and neural networks. The system is divided into two subsystems: the countermeasure system and the verification system. The countermeasure system is responsible for detecting whether the speech is a spoofing speech generated by voice conversion or speech synthesis. The verification system is able to verify whether the speech is consistent with the identity claimed by the speaker through the voiceprint feature. In the countermeasure system, we use the method of representation learning and transfer learning to let the neural network can learn various spoofing speech features. First we use multiple labels of data for training, then use two labels of data fine-tuning models to learn the representation vectors of bona fide and spoofing speech, and finally use the support vector machine to classify. On the ASVspoof 2019 evaluation set, our system achieves a minimum tandem decision cost of 0.1782, and an equal error rate (EER) of 7.62%. In the speaker verification system, we apply a large training data to learn the speaker characterization, and use the learned speaker representation to enrollment and verification. We focus on the text-dependent task, and we evaluate our system on the real environment of 20 testers can achieve the 99% accuracy.
|
|---|
