Task dynamic and articulatory recovery of lip and velar approximations under model mismatch conditions

An algorithm for recovering task dynamics and speech articulator movements from speech acoustics was tested under various model mismatch conditions. There was evidence of articulatory compensation to recover tract-variable (constriction) trajectories in speech produced with a lip approximation under sufficiently constrained conditions. However, in more extensive studies of lip and velar approximations, the recovered tract-variable trajectories were also different from those of the data-producing utterance. This phenomenon can occur because the matching criterion in the analysis-by-synthesis procedure is an acoustic criterion and the correspondence between the tract-variable trajectories and the acoustic output is not exact. While there may be some tract-variable compensation to attain a good acoustic match, there is evidence of a correspondence between how well the tract-variable trajectories match and how well the formant frequencies match in particular instances.