On "high" and "low" in English.

Abstract

The subject matter of this paper is twofold. On the one hand, it is concerned with a general problem in the analysis of data from forced-choice labelling experiments. On the other hand, it is concerned with a specific set of experiments relating to the perception of English intonation patterns. The usual procedure of analysis of forced-choice labelling experiments involves drawing a graph. The value of the acoustic parameter varied is the X variable, and the percentage of subjects giving a particular response is the Y variable. In some cases, more than one choice of X is possible, and the experimenter needs a method of assessing the relative merits of different choices. In the present paper, a statistical technique (multiple linear regression analysis) is described which makes such assessment possible. In the specific set of experiments considered, English-speaking subjects were asked to evaluate synthetic rising-intonation "yes"s and "no"s as "high" or "low". In the analysis of these experiments, the chief problem was determining the "best" choice of X — subjects' responses might depend on Fo level or on extent of rise or on both at once, and other factors (formant frequency, lexical item) might also influence their judgments. Statistical analysis indicated that the best choices of X were (1) log endpoint Fo and (2) log starting Fo together with extent of rise. In both cases, lexical item had to be taken into account as well. The same statistical techniques were applied to published data from experiments in which subjects were asked to label synthetic English words varying in Fo contour as "question" or "statement". The patterning of responses in these experiments was found to be similar to the patterning in the high/low experiments. In general, with regard to such experiments, it seems possible to say that the best choices of X are (1) log endpoint Fo and (2) sets of parameters which sum to this value. Subjects adapt their perception to the stimuli at hand, evaluating the ones with highest ending Fo as "high" or "question", the ones with lowest ending Fo as "low" or "statement". The value of endpoint Fo for which 50 per cent of subjects respond "high" or "question" is at the middle (on a logarithmic scale) of the total range of Fo values spanned by the stimuli. Finally, it is suggested (though not conclusively demonstrated) that the distinction question/statement in English depends primarily on the Fo level at the end of the utterance, rather than on the shape of the Fo contour.