Component-specific temporal decomposition: application to enhanced speech coding and co-articulation analysis

摘要

Atal in 1983 proposed a speech coding algorithm called temporal decomposition (TD), which decomposes a time sequence of LPC derived log-area parameters, into a sequence of overlapping event/interpolation functions corresponding to their associated event vectors. The method provides a “global” event function corresponding to all the area parameters. In this work, we extend Atal’s methodology to obtain the component level event functions corresponding to each area parameter and empirically show that the signal reconstruction error reduces using the component level target function rather than a global target function. This work can further be used to study the coarticulation behavior corresponding to individual articulators by inferring the component level event functions when the proposed method is used for decomposing the articulatory representations.