A Window-based method for target estimation

Authors: Paul Konstantin Krug

Abstract:

The TARGET-APPROXIMATION-MODEL (TAM) describes time continuous articulatory contours through a low-pass filtered sequence of linear functions (targets). Such a representation of articulatory dynamics is well suited for computational simulations of articulatory speech production. The transfer of articulatory trajectories, e.g. measurement data, into the TAM representation is often of particular interest. Although the open source software TARGETOPTIMIZER allows to perform such a transfer successfully, its computational complexity is at least of order O(n3), whereby n is the number of targets to be estimated. This work presents a sequential fit with joint local optimization based on the TARGETOPTIMIZER backend. The proposed solution reduces the computational complexity to O(n).