The project's goal is to build methods and algorithms for the recovery of articulatory movement from acoustic speech signals. In the immediate future this will be accomplished employing analysis-by-techniques for sonorant sounds of English, but with a view to extend this to obstruents and the speech of other languages. There are four main areas of research that will be conducted to reach these goals. The first area is to study the relationship between small changes in articulation and the resulting acoustics, which is feasible now that a large amount of simultaneously recorded articulatory movement and acoustic data are available. The second area of research is in the kinematics of articulatory movement, particularly that of the tongue. Flesh point data enables a data-driven approach to the modeling of tongue kinematics. A quantitative approach to the kinematics of line segments between flesh points, secant lines, is being pursued because it is a first approximation to a kinematics of tongue shape, which is more closely related to acoustics than the flesh points themselves. Because the method for articulatory recovery is analysis-by-synthesis, it is necessary to have an articulatory synthesizer for an internal speech production model. The third area of research is the construction of an articulatory synthesizer that includes knowledge gained from the second area of research. With progress already made in articulatory synthesis, it is possible to concentrate on the kinematic control of the synthesizer. This also will be done from a data-driven approach with piecewise polynomials fit to secant line kinematic trajectories derived from flesh point data. Finally, the fourth area of research is in the recovery algorithms themselves, which include methods for normalization between different vocal tracts and the articulatory synthesizer's vocal tract. Both veridical (actual space-time articulatory trajectories are reproduced) and non-veridical (categorical segmental properties are reproduced for perceptually accurate identification) articulatory will be tested. Veridical articulatory recovery from speech acoustics is useful for the laboratory and clinic when acoustic and partial articulatory information are available and the scientist or clinician wants to know more about articulation. Non-veridical articulatory recovery is important for models of speech and language learning.
|McGowan, Rebecca W; McGowan, Richard S; Denny, Margaret et al. (2014) A longitudinal study of very young children's vowel production. J Speech Lang Hear Res 57:1-15|
|Jackson, Michel T-T; McGowan, Richard S (2008) Predicting midsagittal pharyngeal dimensions from measures of anterior tongue position in Swedish vowels: statistical considerations. J Acoust Soc Am 123:336-46|