This paper presents a CELP speech coding system based on mel-generalized cepstral analysis. In the mel-generalized cepstral analysis, we can vary the model spectrum continuously from AR to cepstral modeling by changing the value of a parameter y and we can choose an appropriate model spectrum. Furthermore, the spectrum represented by mel-generalized cepstrum has frequency resolution similar to that of human ear. Since the perceptual weighting and postfiltering are carried out through the mel-generalized cepstrum, we expect the perceptual performance of the proposed coder to be improved. The subjective performance test indicates that the quality of the proposed CELP coder is about 2 dB higher than that of the conventional one.