New basis functions for sinusoidal decompositions

A set of basis functions for unvoiced frames that incorporate some knowledge about the peripheral auditory system and speech perception is presented. The basis functions are sinusoids modulated by stochastic signals and are called narrowband basis functions, since their energy remains concentrated in the vicinity of the centre frequency of each sinusoid. The use of these narrowband functions to model unvoiced fricatives makes it possible to obtain very high-quality synthetic speech, with a small number of parameters, and without tonal artifacts. Being well suited to model unvoiced sounds, the superposition of narrowband basis functions is not only a natural extension of the concept of sinusoidal decomposition, but also opens a path toward high-quality speech coding at medium-to-low bit rates.<
>