Articulatory-WaveNet: Deep Autoregressive Model for Acoustic-to-Articulatory Inversion

Acoustic-to-Articulatory Inversion, the estimation of articulatory kinematics from speech, is an important problem which has received significant attention in recent years. Estimated articulatory movements from such models can be used for many applications, including speech synthesis, automatic speech recognition, and facial kinematics for talking-head animation devices. Knowledge about the position of the articulators can also be extremely useful in speech therapy systems and Computer-Aided Language Learning (CALL) and Computer-Aided Pronunciation Training (CAPT) systems for second language learners. Acoustic-to-Articulatory Inversion is a challenging problem due to the complexity of articulation patterns and significant inter-speaker differences. This is even more challenging when applied to non-native speakers without any kinematic training data. This dissertation attempts to address these problems through the development of up-graded architectures for Articulatory Inversion. The proposed Articulatory-WaveNet architecture is based on a dilated causal convolutional layer structure that improves the Acoustic-to-Articulatory Inversion estimated results for both speaker-dependent and speaker-independent scenarios. The system has been evaluated on the ElectroMagnetic Articulography corpus of Mandarin Accented English (EMA-MAE) corpus, consisting of 39 speakers including both native English speakers and Mandarin accented English speakers. Results show that Articulatory-WaveNet improves the performance of the speaker-dependent and speaker-independent Acoustic-to-Articulatory Inversion systems significantly compared to the previously reported results