The purpose of this paper is to build an automatic evaluation model of English pronunciation. The technique proposed in Yoon (2009) is extended to include the evaluation of segmental proficiency in the form of spectral comparison. A target English sentence is uttered by a group of “good” Korean speakers learning English, another group of “bad” speakers and the third group of model native speakers of English. Each of the utterances of the good and bad groups is compared against all of the utterances of the model group in terms of four factors; spectrum, F0, intensity and duration. Comparison of the spectral envelops, F0 contours, intensity contours and segmental durations between the two matching utterances yielded 4-dimensional coordinates, which were then plotted in a 4-dimensional model space. This creates an automatic evaluation model of English pronunciation proficiency in the form of a discriminant function. The leave-one-out crossvalidation created 24 models the predictive power of which was then tested with the use of discriminant analyses. The tests reveal that the prediction rate was 67% for the good group and 80% for the bad group.