Twelve isolated sustained English vowels were magnetically recorded and then presented to English speaking subjects in pairs. The subjects were asked to estimate the perceptual similarity of each pair by drawing a horizontal line. The length of the line corresponded to the degree of similarity. A short line represented a more similar pair; a long line, a less similar pair. The average length of all lines for all auditors was computed for each pair of vowels. These average values were summarized in a matrix and were used in subsequent analysis. Two statistical models were used in the analysis: Multidimensional Scaling (MDS: Shepard, Kruskal) and Hierarchical Clustering Schemes (HCS: Johnson). These models bring out the most prominent dimensions in the set of similarity values and are interpretable generally as phonetic features. Feature strength and feature markedness were also computed. [This work was supported by The National Institutes of Health, U. S. Department of Health, Education, and Welfare.]