Speech transmission at 32 kbps by ADPCM coding offers a potential saving of a factor of 2 in bits transmitted per conversation. To evaluate the extent to which such codecs introduce perceptible degradation in the speech signal, two systems, one employing a third‐order fixed predictor, the second a third order adaptive predictor were simulated. Evaluation through relative preference tests shows a slight degradation in quality compared to Mu‐255 8 bit log PCM companding under high loading conditions (>20 dBm0) and a slight improvement in quality under low loading conditions. Comparison of our subjective preference results with a modified segmental measure of signal‐to‐noise ratio agrees with the findings of McDermott et al. [Bell Syst. Tech. J. 57, 1597–1618 (1978)] for simpler codecs, namely that the segmental SNR is a significantly better predictor of the subjective quality of coded speech than the SNR measured over an entire utterance.