One significant problem with sinusoidal modeling of the speech signal is due to the use of standard Fourier Transform for a quasi-periodic signal. The analysis accuracy is severely limited by the lack of stationarity of the analyzed segment, since the analysis is based on the conventional Fourier Transform. An improved analysis technique based on the Generalized Fourier Transform (GFT) with quadratic phase will be discussed in this paper. Speech signal is modeled as a sum of harmonic cosines but with nonlinear phases. A technique for estimation of the time-varying model parameters from the GFT spectrum is proposed. It will be shown that the modeling gain can be improved significantly by inclusion of a single additional parameter in the analysis procedure.