An analysis-by-synthesis approach to modeling real-time MRI articulatory data using the task dynamic application framework.
Erik Bresch, Louis Goldstein, Shrikanth S. Narayanan · The Journal of the Acoustical Society of America · 2009
We report on a method of modeling real-time MRI articulatory speech data using the Haskins task dynamic application (TaDA) framework. TaDA models speech using a set of discrete dynamical regimes that control the formation of vocal tract constrictions (gestures). An utterance can be specified by a gestural score: the pattern of activation of these regimes in time. Individual model articulator degrees of freedom are automatically coordinated according the concurrent demands of the unfolding constrictions. Our modeling procedure consists of two stages: (1) After determining the outline of the midsagittal upper airway, time series of constriction measurements are derived which allow the estimation of the subject-specific parameters relating the articulator and constriction domains. (2) Gradient descent is utilized to adjust the activation intervals of the gestural score generated by TaDA for that utterance so that the synthesized vocal tract constriction evolution matches the observed MRI time series. Additionally, information related to glottal control, which cannot be derived from the MR images, is extracted from the simultaneously recorded audio, and it is used to optimize the timing of glottal abduction gestures in the score. Adequacy of fit is evaluated at articulatory, constriction, and perceptual levels. [Work supported by NIH.]