Articulatory parameter generation using unsupervised Hidden Markov Models

We present an acoustic-to-articulatory inversion method based on unsupervised Hidden Markov Models. A global HMM is first trained from the acoustic and articulatory data. This model is then split in two sub-models which represent the acoustic part and the articulatory part of the data. These two sub-models are linked through the fact that they are deduced from the same global model.