Acoustic-to-articulatory mapping codebook constraint for determining vocal-tract length for inverse speech problem and articulatory synthesis

Abstract

A new method to generate an acoustic-to-articulatory mapping codebook for constraining vocal-tract (VT) length in the inverse speech problem, for variable VT length model, is proposed. The codebook design scheme starts with generating an initial articulatory parameter candidate list (including VT length) with proper quantitative levels. Then it derives VT areas from the parameters and calculates formant parameters. After removing inappropriate candidates, a codebook with uniquely formant-to-VT mapping is finalized by clustering. During the clustering, VT length constraint versus formants is conducted by fitting a two-dimensional interpolation function for irregularly spaced data of VT-to-formant relation inventory. The proposal can facilitate inverse speech solution and articulatory synthesis with variable VT length model.