This thesis proposes a new speech separation algorithm for the mixture of unvoiced speech and voiced speech. After reconstruction of voiced speech using a sinusoidal separation algorithm, Kalman filter separates the residual mixture. The parameters of Kalman filter are estimated by filtering into low spectral region for voiced residual and high spectral region for unvoiced, and applying a codebook for the unvoiced speech. The mixtures of speeches are separated by the proposed algorithm together with the sinusoidal separation algorithm. According to the classification of mixtures, the mixtures of voiced speeches are separated by the sinusoidal algorithm, and the mixtures of voiced and unvoiced are separated by the proposed algorithm.