This thesis documents the application of Evolutionary Algorithms to the Acoustic Articulatory Inversion of vowels using the Praat physical model. Speech synthesis by computational means has been a field for many decades, giving rise to several differing approaches. While the physical modelling approach has been a continually active area of research, it has yet to prove a viable solution to applied TTS systems partly due to the complexity of the model parameters, hence why an efficient and robust method of Acoustic-Articulatory Inversion of parameters for these models would benefit the field. To do so, the use of Evolutionary Algorithms as a means to solve this is proposed. Firstly an introduction to the problem is given, and a taxonomy of Acoustic Articulatory Inversion is developed that draws connections and distinctions between relevant adjacent subfields. A methodology for performing experiments emphasising statistical rigour and the ASECP framework for facilitating the processing are both presented. Then experiments comparing numerous fitness function designs and algorithm configurations are executed. Additionally, multi-model experiments ascertain the generality of these algorithms across different types of physical models and their control parameters. The work concludes with identifying shortcomings of the research and several concrete suggestions for future work.
Date of Award | 2024 |
---|
Original language | English |
---|
Supervisor | Eduardo Miranda (Director of Studies (First Supervisor)) & Alexis Kirke (Other Supervisor) |
---|
An Evolutionary Algorithms Approach to Physical Modelling of Vocal Synthesis
Drayton, J. (Author). 2024
Student thesis: PhD