Speech Science Forum - Daniel van Niekerk (UCL)

Abstract:

High-quality articulatory speech synthesis has many potential applications in speech science and technology, however, developing appropriate mappings from linguistic specification to articulatory gestures is currently a difficult and time consuming manual process. Furthermore, manually defined gestures are likely to be incomplete or only appropriate for a specific language or dialect and need to be revised for new vocal tract models. This problem can be addressed by a procedure that learns to articulate phonologically appropriate utterances automatically –  a task resembling speech acquisition.

In this talk, I will present a personal account of progress made during the project “High-quality simulation of early vocal learning” which was conceptualised as a computational investigation of phonetic assumptions regarding coarticulation and articulatory synchronisation within the domain of the syllable, and critical mechanisms such as visual and corrective feedback. This will include a discussion of experiments with specific findings and I will motivate that our work and computational framework has made progress towards constructing a functional, albeit abstract, embodied agent engaged in vocal learning.