X. Speech Analysis

Most vowel identifiers constructed in the past were designed on the principle of "pattern matching"; that is, for each vowel a pattern of some set of measurable parameters, such as a particular energy density spectrum, frequency location of the formants, axis crossing density, and so on, was chosen to represent the average or "ideal" set for that vowel. These patterns were stored in the apparatus, and the results of measurements on an unknown input were compared with the patterns in order to obtain the best match, which was interpreted as the uttered vowel. Implicit in these schemes is the assumption that with each utterance of a vowel, the speaker is trying to hit a complicated target pattern of such measurable parameters. This assumption has never been proven, nor is it logically the only possible one. An alternative model, the distinctive feature model, stresses the fact that identification is possible so long as the speaker differentiates each vowel from all other vowels in his language in some consistent manner. Since with optimal coding it is possible to distinguish 2 n different vowels by means of n binary features, a distinctive feature model is evidently very economical. It also obviates the necessity of assuming the stringent requirements of accuracy in producing the vowels that are demanded in the "pattern matching" hypothesis. As long as the relevant features are properly produced, identification will be possible, regardless of other measurable features of the pattern. The description of these relevant features in terms of electrical measurements may be quite complex, but the feature "pattern" associated with each vowel is simply stated in binary terms of presence or absence of the pertinent features. Thus this approach eliminates the storage of complex parameter patterns, but may entail an involved measurement procedure to extract information in terms of distinctive features. Since all properties of the acoustic stimulus except those serving to distinguish different words in the language are ignored as carrying more information about the speaker, the linguistic and extra-linguistic context, and so forth, than about the vowel that is spoken, identification should be relatively independent of the speaker provided his dialect possesses the selected features. Table X-l(a) gives a summary of the results of a theoretical analysis of a set of American English vowels on the basis of the four pertinent distinctive features. If the indicated binary decision is made for each feature, every vowel shown will be uniquely specified. Note that in four cases specification can be made on the basis of only three