Audiovisual speech inversion by switching dynamical modeling governed by a Hidden Markov process

We propose a unified framework to recover articulation from audiovisual speech. The nonlinear audiovisual-to-articulatory mapping is modeled by means of a switching linear dynamical system. Switching is governed by a state sequence determined via a Hidden Markov Model alignment process. Mel Frequency Cepstral Coefficients are extracted from audio while visual analysis is performed using Active Appearance Models. The articulatory state is represented by the coordinates of points on important articulators, e.g., tongue and lips. To evaluate our inversion approach, instead of just using the conventional correlation coefficients and root mean squared errors, we introduce a novel evaluation scheme that is more specific to the inversion problem. Prediction errors in the positions of the articulators are weighted differently depending on their relevant importance in the production of the corresponding sound. The applied weights are determined by an articulatory classification analysis using Support Vector Machines with a radial basis function kernel. Experiments are conducted in the audiovisual-articulatory MOCHA database.