Modelling speech prosody based on communicative functions and articulatory dynamics
Yi Xu
(University College London)
It has become increasingly urgent in recent years to improve our
understanding of speech prosody. This is because prosody not only
conveys important communicative meanings by itself, but also is a
major source of variability in the segmental aspect of speech. One
of the most effective ways of studying prosody is through
quantitative modeling, as the modeling process forces us to make
our theoretical assumptions as explicit as possible and to subject
them to highly rigorous testing. In this talk, I will report the
findings of our recent efforts in prosodic modeling based on an
articulatory-functional approach. Our aim was to build a model
that is at once functionally driven and articulatorily based.
Functionally driven means that the categorical units of the system
are defined in terms of meaningful communicative functions such as
lexical stress, focus, sentence modality, as opposed to units that
are primarily defined in terms of phonetic properties.
Articulatorily-based means that critical articulatory dynamics are
algorithmically incorporated as properties intrinsic to the model,
as opposed to algorithms that bypass articulatory mechanisms and
model surface forms directly. The initial results of modeling two
very different languages, Mandarin and English, demonstrate that
such an approach can generate prosody that is both communicatively
effective and highly natural. These encouraging results suggest that
much more research can be conducted to explore the theoretical
implications of the articulatory-functional approach to prosody as
well as its practical applications in speech technology.
|
|
