On a time-varying complex speech analysis

LPC methods are successfully employed in a broad range of speech processing. However, the LPC methods can not estimate time-varying parameters since observed signals are assumed as stationary within the analysis interval. On the other hand, several complex LPC methods for analytic signals have already been proposed. These methods take some advantages over conventional LPC, i.e., more accurate spectral estimation, smaller computation, and so on. This paper presents a new non-recursive complex signal analysis for analytic speech signals by introducing a time-varying AR model in which the parameters are represented by complex basis expansion. The complex AR coefficients can be efficiently estimated by solving linear equation by means of an extended version of LDU decomposition. The experimental result with Japanese natural speech demonstrates that the proposed method offers accurate estimate of time-varying speech spectrum.