In this paper, we examine the robustness of an ADPCM coder based on adaptive mel-cepstral analysis. To improve the noisy channel performance, we use traditional techniques: the leakage factor and sign function. The subjective speech quality of the proposed 16kb/s coder and G.726 coder in terms of the opinion equivalent Q is measured and compared. It is shown that the proposed coder produces much higher quality speech than that of 16kb/s G.726 at BER(Bit Error Rate)=0 and BER=10~3. Although the coder scores 4dB lower than 32kb/s G.726 at BER=0, the improvement of more than 5dB is achieved by the proposed coder over G.726 at BER=10~3.