Production modeling for articulatory recovery

Cress, Llc, Lexington MA

Investigators

Linked publications & trials

Abstract

The speech production model developed in this work will serve as an internal model in an analysis-by-synthesis algorithm intended to recover articulatory movement from speech acoustics. The speech production model will have three components: 1) an improved version of the Haskins articulatory synthesizer, ASY, 2) an improved task-dynamic model, and 3) an inverse normalizing map, which relates ASY vocal tract shapes to human vocal tract shapes. In order for ASY and the task-dynamic model to serve as components of a model of a human talker it is necessary that the task-dynamic, with end effectors in ASY, induce the articulatory kinematics observed in the talker as the image of the inverse normalizing map. ASY and the task-dynamic model are parts of a veridical model of human speech production only in conjunction with an inverse normalizing map. The approach here will be to make ASY and the task-dynamic model as realistic as possible without compromising their relative simplicity. This will enable inverse normalizing maps to be mathematically regular functions, which is an important property for articulatory recovery. There have been many empirical studies on vocal tract shape performed since the time ASY was constructed that can be included into ASY. Improvements in the transfer function calculation can also be made at this time. With improvements to ASY will necessarily come improvements in the task-dynamic model, particularly tongue control and control during obstruent production. Various methods for constructing the inverse normalizing map, along with its ability to map articulatory kinematics, will also be tested. With all three components, the task-dynamic model will be tested using X-ray microbeam data of human speech production.