In this paper, a new algorithm for a robust voiced-unvoiced classifi- cation of frames in telephone speech is presented. This algorithm is associated to a continuous pitch contour extraction procedure based on a discrete logarithmic Fourier transform and a dynamic programing search as introduced in [1]. While the pitch extraction procedure is forced to find a pitch value for every frame, even in unvoiced regions, the algorithm presented in this paper is capable of making a robust voiced-unvoiced decision by using multiple features and twostate ergodic hidden Markov modeling (HMM).