| |
Abstract:
Surface phonetic variation observed during speech production
is typically the result of complex interaction between the
central nervous system and peripheral biomechanics. Hypotheses
about the nature of underlying control strategies and the
influence of peripheral feedback cannot be evaluated properly
until the action of the periphery has been taken into account.
Current attempts to assess the division of labour between central
and peripheral processes are limited by a number of modelling
assumptions. Models of the periphery are typically based on
simple kinematic descriptions that make unrealistic assumptions
about articulator dynamics and the role of muscle geometry.
Models of central control are typically based on over-simplistic
attempts to associate invariant phonetic targets or control
parameter trajectories with individual phonemes. Within such
frameworks, it is difficult to represent natural variations in
control strategy that may occur for different contexts or
speakers, and it is difficult to capture the physical mechanisms
that determine the influence of the periphery. It would perhaps
be more useful to attempt to characterize and assess the variety
of possible control strategies that might be responsible for
observed articulatory movement, within a frame of reference that
directly reflects the underlying physics. This suggests a
statistical approach based on an explicit biomechanical model of
the vocal tract. The aim of this paper is to propose a stochastic
framework for integrating a statistical description of possible
control hypotheses with an explicit deterministic model of
peripheral dynamics. A general mathematical model for
representing probabilistic families of control trajectories is
developed, based on the theory of hidden Markov processes, and it
is shown how the model can be used to control a physical
simulation of the periphery. The contribution of the paper is to
demonstrate that control hypotheses can always be formulated in
statistical terms as probability distributions on a function
space of control parameters, which the peripheral system
transforms into corresponding probability distributions of
observed articulatory variation. The interest of the framework
lies in the possibility of explicitly calculating and predicting
the relationship between systematic statistical variation in
control trajectories and corresponding observed patterns of
articulatory correlation. Simulation results are presented using
a biomechanical model of the jaw, hyoid, and larynx.
|
