Learning disentangled speech representations

Abstract

A variety of informational factors are contained within the speech signal and a single short recording of speech reveals much more than the spoken words. The best method to extract and represent informational factors from the speech signal ultimately depends on which informational factors are desired and how they will be used. In addition, sometimes methods will capture more than one informational factor at the same time such as speaker identity, spoken content, and speaker prosody.

The goal of this dissertation is to explore different ways to deconstruct the speech signal into abstract representations that can be learned and later reused in various speech technology tasks. This task of deconstructing, also known as disentanglement, is a form of distributed representation learning. As a general approach to disentanglement, there are some guiding principles that elaborate what a learned representation should contain as well as how it should function. In particular, learned representations should contain all of the requisite information in a more compact manner, be interpretable, remove nuisance factors of irrelevant information, be useful in downstream tasks, and independent of the task at hand. The learned representations should also be able to answer counter-factual questions.

In some cases, learned speech representations can be re-assembled in different ways according to the requirements of downstream applications. For example, in a voice conversion task, the speech content is retained while the speaker identity is changed. And in a content-privacy task, some targeted content may be concealed without affecting how surrounding words sound. While there is no single-best method to disentangle all types of factors, some end-to-end approaches demonstrate a promising degree of generalization to diverse speech tasks.

This thesis explores a variety of use-cases for disentangled representations including phone recognition, speaker diarization, linguistic code-switching, voice conversion, and content-based privacy masking. Speech representations can also be utilised for automatically assessing the quality and authenticity of speech, such as automatic MOS ratings or detecting deep fakes. The meaning of the term "disentanglement" is not well defined in previous work, and it has acquired several meanings depending on the domain (e.g. image vs. speech). Sometimes the term "disentanglement" is used interchangeably with the term "factorization". This thesis proposes that disentanglement of speech is distinct, and offers a viewpoint of disentanglement that can be considered both theoretically and practically.