This paper presents a novel voice morphing system which reproduces high quality speech while maintaining the majority of the target characteristics. Bi-GMM is named for using GMM technique to estimate mapping functions as well as a codebook generated by GMM either. Compared with the traditional GMM technique, a maximum likelihood estimation framework combined with codebook compensation technique is proposed to overcome the overly smoothed problem caused by conventional GMM. Furthermore, in order to alleviate the discontinuities between frames, a time domain median filter is applied. The STRAIGHT algorithm is adopted for the analysis and synthesis process. The objective and subjective evaluations show that the quality of the speech converted by the proposed method is significantly improved compared with the results by the traditional GMM method.