Deep acoustic-to-articulatory inversion mapping with latent trajectory modeling

This paper presents a novel implementation of latent trajectory modeling in a deep acoustic-to-articulatory inversion mapping framework. In the conventional methods, i.e., the Gaussian mixture model (GMM)- and the deep neural network (DNN)- based inversion mappings, the frame interdependency can be considered while generating articulatory parameter trajectories with the use of an explicit constraint between static and dynamic features. However, in training these models, such a constraint is not considered, and therefore, the trained model is not optimum for the mapping procedure. In this paper, we address this problem by introducing a latent trajectory modeling into the DNN-based inversion mapping. In the latent trajectory model, the frame interdependency can be well considered, in both training and mapping, by using a soft-constraint between static and dynamic features. The experimental results demonstrate that the proposed latent trajectory DNN (LTDNN)-based inversion mapping outperforms the conventional and the state-of-the-art inversion mapping systems.