Temporal Convolution Network Based Joint Optimization of Acoustic-to-Articulatory Inversion

Articulatory features are proved to be efficient in the area of speech recognition and synthesis. However, acquiring articulatory has always been a difficult research hotspot. A lightweight accurate model is significant meaning. In this study, we propose novel temporal convolution network-based acoustic-to-articulatory inversion system. The acoustic feature converted into high-dimensional hidden space map through with frame-level correlations taken account. Meanwhile, construct two-part target function combining prediction’s Root Mean Square Error (RMSE) sequences’ Pearson Correlation Coefficient (PCC) jointly optimize performance specific from both aspects. We also further conducted an analysis on impact weight between two parts final model. Extensive experiments have shown that our, networks (TCN) outperformed Bi-derectional Long Short Term Memory by 1.18 mm RMSE 0.845 PCC 14 parameters when optimizing evenly