Time-varying autoregressive modeling of audio and speech signals

Conventional linear predictive techniques for modeling of speech and audio signals are based on an assumption that a signal is stationary within each analysis frame. However, natural signals are often continuously time-varying, i.e., nonstationary. Therefore this assumption might not be well justified. In this paper, we study a time-varying autoregressive (TVAR) modeling technique in which this restriction is relaxed. A frequency-warped formulation of the Subba Rao-Liporace TVAR algorithm is introduced in the article. The applicability of the presented methodology to various speech and audio signal processing tasks is illustrated and discussed. It is also shown that the TVAR scheme yields an efficient parametrization for time-varying sounds.