Cluster analysis is a straightforward approach to classification of sounds, but the distance matrices required for input are problematical. Using a limited number of frequency time coordinates to represent each sound allows efficient handling of large data sets, as well as classification of data not used to define the clusters. Cross-correlation analysis offers the power of complete comparison of sounds, but no way to classify sounds not included in the cross- correlation. We evaluate McCowan's elegant parametric method of comparing dolphin whistles (PCA of 20 equally spaced frequency measurements) for agreement with less efficient but more powerful cross-correlation methods, on an independent data set of bottlenose dolphin whistles. Kmeans clustering of principal components produced different clusters from 3 other methods. Distance matrices generated with McCowan's 20 variables and with cross correlation analysis of time-normalized frequency contours were similar, but distance matrices based on non-normalized contours differed greatly. Our results suggest that McCowan's method might be improved by including duration as a variable in the PCA.