Robust Speaker Verification in Real Application Scenarios

点击翻译按钮获取中文摘要

Spoken language is the most natural way we human communicate with each other. There is rich information conveyed in speech signal including language information, speaker information, environmental information and so on. Speaker verification refers to the problem of verifying the identity of a person from his/her voice using the characteristic vocal information. The necessity to legitimate an individuals' identity arises in several situations, including access control and authorization of financial transactions. Recent developments in speech-based technologies have led to speech being touted as becoming the primary means of communication between humans and technology in the future. As the use of speech becomes more ubiquitous, there is a need for improvement and innovation in voice-based verification technologies, specifically speaker verification, so that these methods work reliably in real world scenarios. To incorporate voice biometrics into real-world applications, it is important to ensure that verification performance can still be maintained even if the speakers are speaking in an adverse environment that the system has not confronted during training. Adverseness can be caused by background noise, reverberation, channel mismatch, language mismatch, and accent. Apart from domain robustness, a major concern with deploying speaker verification in real-world applications is the system's robustness against fraudulent attacks. This is due to the vulnerability of speaker verification systems to spoofing attacks. This proposal focuses on building speaker verification systems which are domain-invariant and are robust to spoofing attacks. To tackle the domain mismatch problem, our goal is to learn domain-invariant speaker embeddings using domain adversarial training for robust speaker verification. We propose the use of deep learning architectures trained to both classify speakers and the domain. The key insight to this approach is that while network gets better at classifying speakers but gets worse at domain classification. As a result, the network leads to domain-invariant speaker representations. We also propose employing some novel unsupervised domain adaptation approaches to bridge the source and target domains. For improving performance further, we also consider combining these approaches with data augmentation and unsupervised PLDA adaptation methods. Finally, to make ASV technology robust against fraudulent attacks, we propose a method for blind automatic detection of spoofing attacks which does not require any prior knowledge about the type of spoofing attacks. Here, convolution neural network based deep countermeasures are proposed for anti-spoofing. Novelty of the proposed research includes to take the benefit of deep learning for domain-invariant representation learning, domain adaptation and deep spoofing countermeasures. The expected results will have a certain impact on speaker recognition research and commercial communities.