A neural clustering algorithm for estimating visible articulatory trajectory

Abstract

The bimodal acoustic-visual nature of speech establishes sound correlations between its audio component and the corresponding articulatory information associated to the time-varying geometry of the vocal tract. In this paper we propose an estimation structure consisting of a simplified Time-Delay Neural Network (TDNN) working on 4-5 dimensional cepstrum trajectories provided by a preceding clusterization layer based on a Self Organizing Map (SOM). The use of this pre-processing layer has allowed an effective non-linear clusterization of cepstrum vectors thus simplifying of one order the complexity of the resulting system while maintaining unchanged the global estimation performances. The achieved results are shown in terms estimation precision and robustness with reference to previously published results.