Semi-automatic extraction of vocal tract movements from cineradiographic data

Since high speed X-ray films still provide the best dynamic view of the entire vocal tract, large existing databases have been preserved and are available for the speech research community. We propose a new technique for facilitating the extraction of the vocal tract shape and the movements of the articulators from complete sequences of these databases. The method was first developed for the extraction of the tongue movements in “Wioland”. It has been adapted to a sequence of the ATR database, Laval43 (provided by Rochette). The method, based on the retromarking algorithm, combines the human expertise applied for marking a small number of key images, and the automatic processing of the video data. It has been extended to other articulators (lips, velum) in order to obtain the shape and the sections of the complete vocal tract. Quantitative evaluations of the estimate error and a comparison with Thimm and Luettin (1999) are achieved.