Statistical methods for predicting articulation from speech
Acoustic-to-Articulatory Inversion has been a special topic in speech technology research for a quite a long while. It tries to address the basic problem of connecting speech production with the acoustics produced. Given the acoustics, the goal is to predict the articulation used to produce that sound. The talk will first briefly cover the different ways in which people have tried to achieve inversion in the past and then in detail describe the challenges that make inversion difficult. I will then present my own work, focused on two statistical modelling techniques to predict articulation from sound and studies on the properties of the acoustic-articulatory relationship. These have investigated the non-uniqueness of the acoustic-to-articulatory mapping and the identification of important parts of the articulatory trajectory. Finally a brief sketch about the possible applications and future work will be discussed.