Self-supervised learning for automatic speech recognition In low-resource environments
Publication Date
July 23, 2024
Creator
Abstract
Supervised deep neural networks trained with substantial amounts of annotated speech data have demonstrated impressive performance across a spectrum of spoken language processing applications, frequently establishing themselves as the leading models in respective competitions. Nonetheless, a significant challenge arises from the heavy reliance on extensive annotated data for training these systems. This reliance poses a significant scalability limitation, hindering the continual enhancement of state-of-the-art performance. Moreover, it presents a more fundamental obstacle for deploying deep neural networks in speech-related domains where acquiring labeled data is inherently arduous, expensive, or time-intensive, which are considered as low-resource ASR problems in this thesis.
Unlike annotated speech data, collecting untranscribed audio is typically more cost-effective. In this thesis, we investigate the application of self-supervised learning in low-resource tasks, a learning approach where the learning objective is derived directly from the input data itself. We employ this method to harness the scalability and affordability of untranscribed audio resources in problems where we do not have enough training data, with the goal of enhancing the performance of spoken language technology. In particular, we propose three self-supervised methodologies. One model is based on the concept of two-fine-tuning steps, while the other two revolve around the notion of identifying an improved hidden unit. These approaches are designed to learn contextualized speech representations from speech data lacking annotations. We demonstrate the capacity of our self-supervised techniques to learn representations that convert the higher-level characteristics of speech signals more effectively than conventional acoustic features. Additionally, we present how these representations enhance the performance of deep neural networks on ASR tasks with limited resources. Beyond introducing novel learning algorithms, we conduct in-depth analyses to comprehend the properties of the acquired self-supervised representations and elucidate the distinct design elements that separate one self-supervised model from another.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Subjects (LC)
Associated Schools / Departments
School of Computer Science (UK)
eprints ID
77884
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Fatehi,Kavan,20167617,corrections.pdf
Type
Full-text
Description
Examined
Size
3.47 MB
Format
Adobe PDF
Checksum (MD5)
62d74422545366771d3694396a80a71d