Voice conversion is one of useful techniques to enhance pathological speech to be perceived as normal speech, although it concerns also the modifications of normal source speaker's speech to be perceived as if a target speaker had uttered it. The parameters to be converted are obtained by matching the spectral envelope of the vocal tract for the source and the target speech. Gaussian Mixture Models (GMMs) parameters are determined for providing conversion functions. The main contribution of our study consists in the prediction of Fourier cepstrum coefficients related to the excitation signal. Such a prediction leads to a satisfactory voice conversion system. Subjective perceptual results indicate that the proposed approach yields significant improvements in quality of the converted voice.