Published online by Cambridge University Press: 22 April 2026
Sol Saporta has presented data which are claimed to confirm his hypothesis that ‘the average frequency of a consonant cluster is a function of the difference between the phonemes in the cluster: low frequencies are expected for clusters which are either extremely similar or extremely dissimilar; high frequencies are expected for clusters which are at neither extreme.‘ The difference between any two phonemes is measured as the number of distinctive features in which they are different. A second article, in which Saporta collaborated with Kathryn C. Keller, considers the possibility that the results might have been expected by chance. The concerns of the present note are (1) to point out that Keller and Saporta did not adequately resolve the question of the chance nature of the results, (2) to present the appropriate mathematical and statistical techniques for doing so, and (3) to make new assessments of some of the data analyzed by Keller and Saporta.
1 Sol Saporta, Frequency of consonant clusters, Lg. 31.25–30 (1955).
2 Kathryn C. Keller and Sol Saporta, The frequency of consonant clusters in Chontal, IJAL 23.28–35 (1957).
3 Certain algebraic manipulations would make it possible, however, to circumvent the computation of phoneme probabilities and to predict the distribution of distinctive-feature differences between pairs of phonemes directly. For example, for a system in which there are n distinctive features in each phoneme, it can be shown that the sum of the probabilities for phoneme pairs having zero differences in distinctive features is
Similarly, the sum of the probabilities for phoneme pairs having n differences in distinctive features is
Expressions for other sums in this series are still more complicated.
4 The fact that we distinguish between pi. and p.i might be regarded as a restriction upon the completely random clustering of phonemes, which, indeed, it is. The formulation given, however, is perfectly general, for if one chooses to assume that phonemes are equally likely to occur as first and as second members of clusters, one can determine a single probability for each phoneme and assume this probability pi. = p.i. We have distinguished between pi. and p.i because we observe very large differences in the tendency of some phonemes to appear as first or as second members of clusters, and it has seemed wise to take account of this observation in constructing our probability model.
5 This is not the same as the conditional probability of j given i, a quantity arising in information theory and computed from empirical data as the proportion of instances of i which are actually followed by j. Our values of p.i(i) are theoretical expectancies against which we wish to assess the actual conditional probabilities.
6 See, for example, S. Siegel, Nonparametric statistics for the behavioral sciences 42–7 (New York, 1956).
7 John B. Carroll, Transitional probabilities of English phonemes, Unpublished Ms. (Cambridge, Mass., 1952).
7a All tables are assembled at the end of the article.
8 R. Jakobson, C. G. M. Fant, and M. Halle, Preliminaries to speech analysis (Technical Report No. 13, Acoustics Laboratory, MIT.; Cambridge, Mass., Jan. 1952).