Learning New Articulator Trajectories for a Speech Production Model using Artificial Neural Networks

We present a novel method for generating additional pseudo-articulator trajectories suitable for use within the framework of a stochastically trained speech production system recently developed at CUED. The system is initialised by inverting a codebook of (articulator, spectral vector) pairs, and the target positions for a set of pseudo-articulators and the mapping from these to speech spectral vectors are then jointly optimised using linearised Kalman filtering and an assembly of neural networks. A separate network is then used to hypothesise a new articulator trajectory as a function of the existing articulatorsand the output error of the system. The techniques used to initialiseand train the system are described, and preliminary results for the generation of new pseudo-articulatory inputs are presented.