A codebook of articulatory shapes
M. M. Sondhi, Juergen Schroeter, J. N. Larar · The Journal of the Acoustical Society of America · 1987
The acoustical properties of a vocal tract depend on its shape which, at frequencies below about 4 kHz, is adequately described by the area function (i.e., cross-sectional area as a function of position from glottis to lips). Obtaining even this simplified description of the tract is a difficult problem. The solution that is proposed is to construct a linked codebook of area functions and corresponding LPC vectors. Then, a given LPC vector is transformed to an area function by finding the closest LPC vector in the codebook and accessing the linked area function. The authors have constructed just such a codebook (to our knowledge, the first). At present, it has about 1500 shapes clustered from a training set of about 10 000 shapes generated by a vocal tract model [P. Mermelstein, J. Acoust. Soc. Am. 53, 1070–1082 (1973)]. Distance between area functions is defined as the Itakura distance between the corresponding LPC vectars. Handling the unique problems arising from this definition of distance and from the large size of the training set will be described. The plan for refining and enlarging the codebook will also be discussed.