Year: 2023
Type: dissertation
Abstract: Automatic speaker verification has numerous studies widely used in applications such as user authentication, access control, and smart assistants.Empowered by stronger hardware utilities, modern speaker verification systems based on deep neural networks have been moving away from conventional statistical models.However, their generalizability in realistic application scenarios remains challenging due to a number of nuisance variations resulting from both the environment and the speakers themselves.In addition, the increasing demands for offline usage on limited computational conditions and security need to be addressed as well.This dissertation advances modern text-independent speaker verification systems on three fronts: robustness, portability, and security.First, regarding robustness, this dissertation focuses on the feature extraction of speaker verification systems.The investigation begins with re-assessing existing feature extractors, followed by several proposals on data-driven acoustic feature extractors trained jointly with a speaker embedding extractor.This covers conventional mel spectral features based on conventional and multi-taper windowing functions and power-normalized cepstral coefficients.These enhancements are shown to be helpful in improving robustness against nuisance variations, leading to improvements in automatic speaker verification performance.Second, regarding portability, the author proposes a simple yet effective knowledge distillation framework to develop a speaker embedding extractor with a small footprint and simple network architecture.The speaker embedding vectors extracted from a sizeable pre-trained model are used as the targets for the small model to learn at the training stage.This strategy substantially reduces the size of the model, with comparable speaker verification ability.Finally, regarding security, this dissertation addresses the problem of improving the speaker verification module against spoofing attacks (such as text-to-speech and voice conversion) without the primary reliance on an additional spoofing countermeasure module.The investigation starts by studying the feasibility of a simple, unsupervised domain adaptation approach that uses a small quantity of spoofing data to enhance the system.The work is then furthered with a simple backend classifier based on a fully-connected 3-layer neural network, trained in a supervised manner.This involves an extensive comparison of different strategies using additional spoofed and non-spoofed training data.In summary, this dissertation proposes advances in speaker verification enhanced by deep learning.It introduces multiple methods to make the speaker verification systems more robust to variations, usable with limited computational resources, and secure against audio spoofing attacks.The results can be utilized both in basic research and practical applications of speaker verification technology.v
Author: Liu, Xuechen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: speaker verification, text-independent speaker verification, speaker embedding, audio feature extraction, power normalized cepstral coefficients, knowledge distillation, model compression, spoofing attacks, voice spoofing detection, unsupervised domain adaptation, voice conversion, text-to-speech, speaker verification security
Open Access status: green
OpenAlex ID: W7215149501
