Home and International students are eligible to apply. Regardless of fees status (Home or International), all fees will be paid in addition to an enhanced stipend and a research and training support grant to cover research expenses and conference attendance.
The deadline for applications is *19 April 2026*.
For more information, please visit our website:
https://slt-cdt.sheffield.ac.uk/apply
Please feel free to circulate to those who might be interested.
*Project Description: Metamodelling of Speech Domains*
Speech is a highly variable signal that is often recorded in complex
environments and under sub-optimal conditions. The information contained in
a recorded speech signal is not limited to just the words spoken; the
signal also includes, for example, information on speaker identity or
conversation style. Depending on the task at hand, different aspects of the
speech signal are important, leading to different models being used.
However, in recent years model topologies for automatic speech recognition
and many other speech processing tasks have started to converge – driven by
research focus on generalisation. Still, the issue of domain dependence
often remains. Recently there has been an increased interest in model
combination and model editing, for example through disentanglement of
so-called task vectors.
In this project we aim to explore how different aspects of speech data are
expressed in model space, in the context of automatic speech recognition
and diarisation. The objective of this work is to explore methods to
attribute elements of model spaces to skills, or specific aspects of the
data. This can be used either as input in hypermodelling, where new models
for specific domains are generated, or for improved structuring in model
training and design.
Work on this project will require research into novel methods to represent
model variations and attribute them to specific attributes and tasks. The
value of such models should then be demonstrated by informing training and
inference processes. A range of different strategies can be explored,
including new ways to derive model distributions and model parameter
predictions. Experiments should be conducted on a range of tasks of
different complexity in the context of different data domains, for example
speech classification, speech recognition, and diarisation.