Time scale modification of noises using a spectral and statistical model

Abstract

Some natural sounds, such as speech parts can essentially be considered as noises. For instance, models suppose noisy parts of sounds as weak parts and apply basic approximations. But transformations such as time stretching do not preserve the noisy characteristics of sounds. Moreover, we show that those transformations introduce artificial intensity variations. In this paper we propose a spectral model for noise modeling which takes into account the statistical properties of such sounds. The analysis is based on the classical spectral models. The synthesis consists of randomly defining sinusoidal components. These components are then added using the adapted overlap-add method to keep statistical moments constant. Time scaling operations using this approach are described. Experiments on artificial sounds (filtered white noises) as well as natural sounds such as consonants and whispered vowels, show impressive enhancement in quality. Infinite time stretching transformations of such noises can be perfectly performed.