Crosslingual adaptation of semi-continuous HMMS using maximum likelihood and maximum a posteriori convex regression

In this work we present a novel adaptation design for semicontinuous HMMs (SCHMM). The method, which is developed in the scope of a crosslingual model adaptation task, consists in adjusting the states´ mixture weights associated to the prototype densities of the codebook. The mixture weights of the target language are modelled as convex combinations of prototype weights. They are defined by an acoustic regression scheme applied to the source models, followed by a refinement using probabilistic latent semantic analysis (PLSA). In order to find suitable combination weights for the convex combinations we present a maximum likelihood (ML) as well as a maximum a posteriori (MAP) estimate. Thus, we name them maximum likelihood convex regression (MLCR) and maximum a posteriori convex regression (MAPCR). Finally, a crosslingual model adaptation task transferring multilingual Spanish-English-German HMMs to Slovenian demonstrates the performance of the method.