Articulatory analysis using a codebook for articulatory based low bit-rate speech coding

Abstract

Fundamental to the success of the articulatory based speech coding is the mapping from acoustics to articulatory description. As the mapping is not unique and based on articulatory continuity criteria, the non-uniqueness of the articulatory trajectories is solved using a forward dynamic network. In this paper, we present new results on forward dynamic network used to estimate articulatory trajectories when using an improved articulatory codebook for acoustic-to-articulatory mapping. The improvement on the codebook design is based on a new model that provides more details on the vocal tract area function and on more appropriate articulatory parameter samplings according to the articulatory-acoustics relation.