A spectral model for nonstationary voiced speech

Spectral modeling is a key step in several speech processing applications. This paper presents a novel model capable of accurately handling the short-time spectrum of nonstationary voiced speech. As is well known, locally stationary, i.e., periodic, voiced speech exhibits a well-defined spectral line structure. It is shown in this paper that locally nonstationary voiced speech, i.e., speech with variations of both pitch and vocal tract, within the analysis window, still exhibits, in a generalized form, a spectral line structure, each line being formed by a series of well-defined components. The applications of this model to several speech processing problems, namely those involving speech prediction and pitch detection are presented, along with supporting experimental results.