Similarities among phonemes may be measured using auditory discriminations made by human observers or by comparing numerically the features obtained from the acoustic data. A variation of the latter approach is proposed which does not require one to decide beforehand which feature dimensions are relevant. A procedure based on a neural network learning model was developed which yielded a unique, multi-dimensional descriptor for each phoneme. The similarity of a phoneme to each other phoneme could then be easily observed by comparing their respective descriptor vectors. Two methods were employed to examine informally how the distance between any pair of descriptors varied with the perceptual similarity of the phonemes they represented. A dendrogram resulting from a hierarchical clustering analysis of the 49 vectors showed a reasonable grouping of phonemes. Also, a plot of the 49 dimensional arrays mapped to two dimensions via Sammon's (1969) algorithm, showed that spatially adjacent phonemes were quite likely to be perceptually similar phonemes.