The Phoneme-Level Articulator Dynamics for Pronunciation Animation

Abstract

Speech visualization can be extended to a task of pronunciation animation for language learners. In this paper, a three dimensional English articulation database is recorded using Carstens Electro-Magnetic Articulograph (EMA AG500). An HMM-based visual synthesis method for continuous speech is implemented to recover 3D articulatory information. The synthesized articulations are then compared to the EMA recordings for objective evaluation. Using a data-driven 3D talking head, the distinctions between the confusable phonemes can be depicted through both external and internal articulatory movements. The experiments have demonstrated that the HMM-based synthesis with limited training data can achieve the minimum RMS error of less than 2mm. The synthesized articulatory movements can be used for computer assisted pronunciation training.