The Uses of Experiment in Language Description

Abstract

To assess the validity of a grammar, one must derive utterances from it and then determine the acceptability of these utterances to native speakers of the language. At the level of syntax it is easy enough to isolate and recombine the relevant elements (words) and thus generate appropriate utterances, however difficult it may be to devise meaningful tests. In phonology, on the other hand, there should be no particular problem about designing tests; the difficulty is, rather, in deriving testable utterances, particularly when one insists, as he must, that the phonologically relevant elements (phonemes) be freely commutable. This paper attempts to deal with the latter problem.

Until recently the linguist who wanted to move from phonology to utterance could only use the human vocal apparatus, the flexibility and controllability of which is inadequate for the purpose. Instrumental methods seemed equally inadequate until, with the advent of magnetic tape recordings, it appeared that one might make commutation tests by cutting and rearranging tape segments of phoneme length. This has not proved to be feasible, however, and for reasons that have to do with certain very fundamental properties of speech.

Newly developed techniques for synthesizing speech make it possible now to write a phonological description from which testable utterances can be recovered by precisely defined operations. On this basis the phonology becomes, in effect, a set of rules for synthesis, with explicit procedures for going from a sequence of phonemes to their actualization as sound. Such rules are now available in acoustic terms, written at a phoneme or sub-phoneme level, and therefore appropriate as a basis for commutation testing. While these acoustic rules constitute a workable and testable phonology, they are not quite so simple as one might wish. Thus, in the case of at least one phoneme it is necessary to have two rules, one for each of two classes of vowel contexts; more generally, we must apply built-in “modifiers” to the various classes of phonemes in order to accommodate them to various positions in the syllable.

We believe that these complications can be reduced, and a simpler phonology achieved, by stating rules for synthesis in articulatory rather than acoustic terms. These articulatory rules should not, however, describe the changing shape of the vocal tract, since this would produce a result that is, like the acoustic rules, less than ideally simple; if greater economy of description is to be attained, the rules must rather be written in terms of the motor commands that actuate the articulators. Preliminary studies of muscle action potentials, recorded from several articulatory organs, tend so far to support this assumption. We are, however, a long way from having sufficient knowledge about these motor command patterns; moreover, there exists at the present time no synthesizer that will accept such commands as input and thereby permit a rigorous test of a phonology in these terms.

The procedures which lead to a phonological description in terms of acoustic rules for synthesis provide other data which should be of interest to the linguist, even though they are not critically relevant to his primary concern. One learns, for example, that various phoneme classes have very different perceptual properties; we are led then, to speculate that these differences may make the phonemes differentially efficient as vehicles of information and so determine their linguistic roles.