Lead Research Organisation: University of Edinburgh
Department Name: Sch of Informatics


Here we explore prosody modelling without explicit supervision using the FastSpeech 2 framework. A prosody encoder is jointly trained with the framework which allows for conditioning it on any reference utterance and transferring certain acoustic correlates of prosody from the reference to the output. The proposed model adds explicit control of acoustic properties in the output and reaches similar objective results when compared to the prior work. We discuss how the choice of objective metrics and definition of prosody for this task forces us to consider prosody in a narrow way and point out possible improvements for related future endeavours.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S022481/1 01/04/2019 30/09/2027
2435424 Studentship EP/S022481/1 01/09/2020 31/08/2024 Atli Sigurgeirsson