An Experiment in Cluster Detection [Letter to the Editor]
James E. Dammann · IBM Journal of Research and Development · 1966
This study was designed as a practical test of the cluster finding and coding method described in an earlier paper.1In order to test such a method it is necessary to use data in which the clusters are known. The clusters for this experiment were generated by taking measurements on sustained spoken vowels. The utterance of each vowel was sampled at a 20-millisecond rate through 15 filters, and the outputs of the filters were quantized into 13 levels. Each time sample was taken as a pattern, thus producing a cluster of patterns from an utterance of one vowel. Eight vowels were used: |i|, |I|, |ε|, |a|, |ɔ|, |з|, |O| and |u|. These are the vowel sounds found in the words “beet,” “bit,” “bet,” “pot,” “brought,” “Bert,” “boat” and “boot.” From the eight classes a total of 250 patterns were generated. Figure 1 shows a typical sample from the data of each class in column 2 and two additional samples in columns 1 and 3 to give an indication of the divergence that occurred among the samples.