Speech resynthesis from 1/f β noise

Abstract

Speech can be resynthesized from acoustic parameters estimated from the motion of the vocal tract articulators, head, and face. This has shown that speech information is broadly distributed both in space and time across the physiological systems associated with speech production, and correspondingly, it provides a basis for exploring what properties of signal structure are relevant for specifying the relation between acoustics and physiology. In this presentation, we use the resynthesis paradigm to explore the extent to which the acoustic-articulatory relation can be specified in terms of the structure of fluctuations in the speech signal. Specifically, we demonstrate that speech acoustics closely matching the original can be obtained by replacing the vocal tract kinematic signals typically used in the re-synthesis paradigm with surrogate signals having 1/f β fractal fluctuation structure—a known (Voss & Clarke 1975) statistical property of the speech amplitude envelope that holds for low frequencies corr...