Speech-driven head motion synthesis based on a trajectory model.
View/ Open
Date
2007Author
Hofer, Gregor
Shimodaira, Hiroshi
Yamagishi, Junichi
Metadata
Abstract
Making human-like characters more natural and life-like requires
more inventive approaches than current standard techniques such
as synthesis using text features or triggers. In this poster we present
a novel approach to automatically synthesise head motion based on
speech features. Previous work has focused on frame wise modelling
of motion [Busso et al. 2007] or has treated the speach data
and motion data streams separately [Brand 1999], although the trajectories
of the head motion and speech features are highly correlated
and dynamically change over several frames. To model
longer units of motion and speech and to reproduce their trajectories
during synthesis, we utilise a promising time series stochastic
model called ”Trajectory Hidden Markov Models” [Zen et al.
2007]. Its parameter generation algorithm can produce motion trajectories
from sequences of units of motion and speech. These
two kinds of data are simultaneously modelled by using a multistream
type of the trajectory HMMs. The models can be viewed
as a Kalman-smoother-like approach, and thereby are capable of
producing smooth trajectories.