Autoregressive Articulatory WaveNet Flow for Speaker-Independent Acoustic-to-Articulatory Inversion

In this paper we introduce a new speaker independent method for Acoustic-to-Articulatory Inversion. The proposed architecture, Speaker Independent-Articulatory WaveNet (SI-AWN), models the relationship between acoustic and articulatory features by conditioning the articulatory trajectories on acoustic features and then utilizes the structure for unseen target speakers. We evaluate the proposed SI-AWN on the Electro Magnetic Articulography corpus of Mandarin Accented English (EMA-MAE), using the pool of acoustic-articulatory information from 35 reference speakers and testing on target speakers that include male, female, native and non-native speakers. The results suggest that SI-AWN improves the performance of the acoustic-to-articulatory inversion process compared to the baseline Maximum Likelihood Regression-Parallel Reference Speaker Weighting (MLLR-PRSW) method by 21 percent. To the best of our knowledge, this is the first application of a WaveNet-like synthesis approach to the problem of Speaker Independent Acoustic-to-Articulatory Inversion, and results are comparable to or better than the best currently published systems.