Restoration of voiced speech signals preserving prosodic features

Abstract

The paper deals with the restoration of voiced speech signals that contain clicks, noise, and gaps of sufficient length so that entire phonemes are lost. A particular focus is on restoring the proper prosodic elements: accent and stress. The importance of the problem is grounded by the fact that the meaning of some words (known as homographs) is solely dependent on the prosody. A new restoration method, exploiting the developed simple polynomial accent model with averaged speech signal characteristics of intensity and a fundamental frequency period as its parameters, is proposed. Feasibility of the method is confirmed by experimental investigations of the restoration of both one period and multiple periods of a voiced speech signal, examination of an instantaneous error, the total mean-square-error, and the influence of sampling frequency on the restoration quality.