The doctoral dissertation in the fields of Engineering and Technology will be examined at the Faculty of Science, Forestry and Technology, Joensuu Campus and online.
This PhD program is jointly funded by Inria Nancy Grand Est, France and UEF. Xuechen Liu, MSc, has been a visiting PhD student at the French site from March 2020 to August 2021. He is co-supervised by Dr. Md Sahidullah. He also has spent time at A*STAR, Singapore from May to August 2022.
What is the topic of your doctoral research? Why is it important to study the topic?
Automatic speaker verifiers have been extensively researched and are commonly applied in various areas like user authentication, access control, and intelligent personal assistants. However, their ability to perform well in practical application scenarios remains a challenge, primarily due to various factors that introduce unwanted variations, stemming from both environmental conditions and individual speakers. Furthermore, there is a growing demand for offline use under restricted computational resources, along with a heightened need for security considerations that must also be addressed.
What are the key findings or observations of your doctoral research?
We have addressed several issues to improve modern deep speaker verification systems by crafting the modules. Several interesting findings can be useful:
1. While simple data-driven features offer specialized advantages under practical scenarios, their effectiveness may not be as pronounced in typical scenarios.
2. Small speaker verifiers can be developed via distilling knowledge from a large, non-deployable model to a very small one. But it is important to note that "segment-level embedding" is not the sole approach for knowledge distillation.
3. Moreover, in order to improve the security of a speaker verifier, apart from separately utilizing a model, the incorporation of countermeasures can enhance performance during the training phase, with their necessity being able to diminish outside of that context.
What are the key research methods and materials used in your doctoral research?
In terms of research methods, I mainly have done research by utilising effective deep neural network architectures. I have exercised statistical methods for processing input data effectively for training the neural network models, at both large and small scales.
I also have implemented domain adaptation and multi-task learning to enhance the security of neural network speaker verifiers. I have evaluated my methods on both commonly used and specialized (and accessible) datasets, most of which are familiar and considered as practical/"in-the-wild" for the research community. In terms of research process, I mainly have developed algorithms via extending conventional frameworks and run experiments on CPU and GPU resources provided.
The doctoral dissertation of Xuechen Liu, MSc, entitled Advances in Deep Speaker Verification: a study on robustness, portability, and security will be examined at the Faculty of Science, Forestry and Technology, Joensuu Campus and online. The opponent will be Assistant Professor Lauri Juvela, Aalto University, and the custos will be Professor Tomi Kinnunen, University of Eastern Finland. Language of the public defence is English.
For more information, please contact:
Xuechen Liu, xuechen.liu@uef.fi
