In this paper, we propose a new speech probability distribution, twosided generalized gamma distribution (G_D) for an efficient parametric characterization of speech spectra. G_D forms a generalized class of parametric distributions including the Gaussian, Laplacian and Gamma probability density functions (pdf's) as special cases. All the parameters associated with the G_D are estimated by the on-line tracking procedure according to the maximum likelihood principle. Likelihoods, coefficients of variation (CV's), and Kolmogorov-Smirnov (KS) tests show that G_D can model the distribution of the real speech signal more accurately than the conventional Gaussian, Laplacian, Gamma pdf or generalized Gaussian distribution (GGD).