MLLR-PRSW for Kinematic-Independent Acoustic-to-Articulatory Inversion

This paper presents an improved method for kinematic-independent acoustic-to-articulatory inversion, using acoustic adaptation to estimate weights for articulatory model creation from reference speakers. Paired acoustic and articulatory data from the Marquette Electromagnetic Articulography corpus of Mandarin Accented English (EMAMAE) are used for experimental evaluation. The new method is a modification of the Parallel Reference Speaker Weighting (PRSW) inversion algorithm, in which two separate methods are used for acoustic and articulatory adaptation. A Maximum Likelihood Linear Regression (MLLR) approach is used for acoustic adaptation model and the PRSW weighted reference speaker approach is used for articulatory model adaptation. The new MLLR-PRSW adaptation method outperforms the baseline PRSW method on inversion of new test subjects where no kinematic data is used for training, providing estimated trajectories very close to the results from speaker dependent models that do use kinematic data.