Visual speech synthesis by modelling coarticulation dynamics using a non-parametric switching state-space model