Abstract:

Articulatory recovery in an analysis-by-synthesis procedure must account for individual differences in vocal tracts. In the procedure described here, all individual vocal tracts are referenced to the vocal tract of an articulatory synthesizer, known as the standard vocal tract. In order to normalize for any given talker, characteristics of the standard vocal tract are adjusted until its sound output closely matches the sound output of the talker over a range of speech sounds. After these standard vocal tract characteristics are fixed, the analysis-by-synthesis procedure is enabled to recover the articulation from the talker. Experiments are conducted with articulatory and acoustic data from the Wisconsin x-ray microbeam database. The midsagittal sections of the standard vocal tract are made to match those of the x-ray microbeam over a set of phones. Characteristics of the standard vocal tract are adjusted so that its acoustic output matches that of the talker as closely as possible over this range of phones. The normalization procedure is checked by choosing a novel utterance for recovery, and comparing the resulting recovered midsaggitals with and without prior normalization. [This work was supported by Grant NIDCD 01247.]