Impressionistic Coding for Phrasal Stress

For the purposes of this book, stress is taken to be an intuitive classification of syllables as relatively more and less prominent within a spoken utterance: impressionistic classification of phrasal stress. This is not a pattern of stress values predicted from the structure of the message, but a subjective coding of impressions of prominence. For this reason, results may be somewhat different from, e.g., Crystal and House (1990) who used lexical stress classifications. Phrasal stress has a strong effect on vowel quality; in order to see what these effects are, it was necessary to code each vowel for its degree of stress.

While intuitions about lexical stress and stress patterns in complex nominals and clitic groups are quite clear, it is well known (Lieberman 1963) that the intersubjective reliability of phrasal stress coding leaves something to be desired. My own reliability tests described below found 33#3310% disagreement using three levels of stress; four levels may increase intercoder inconsistency to 35% (Mazzie & Veatch, 1986). This kind of observer ``noise'' is probably characteristic of all impressionistic phonetic judgments per se. (This is not true, of course, for judgments of phonological or lexical identity, which are extremely reliable.) To some extent this reflects issues of definition and of training. There may well be deeper problems, a matter that we cannot resolve here.

Here, as clearly as I can define it, is the procedure I used to code phrasal stress, expressed as a set of instructions to a coder.

Before coding for stress, listen a few times to a few minutes of speech, and try to listen for the rhythm of her/his speech. You will be more consistent if you have an intuitive feeling for what that speaker's rhythm sounds like.

Listen to entire utterances (breath groups) rather than single syllables, so that you can hear the overall rhythmic contour. Try to keep the sound of the entire utterance, or many-syllable parts of it, in your mind at one time. Then locate the most prominent syllable(s) which seem to carry the primary beats of the utterance. Don't concentrate on the sound of an individual syllable, but listen to the whole utterance and pick out the prominent parts or the beats. Pay particular attention to prominences marked by pitch inflection. If beats do not stand out in your mind, then try repeating the utterance to yourself in a monotone, without melody, while overemphasizing the rhythm, as if counting meter in poetry. Put the beats of the rhythm where they sound right to you (that is, where you would put them), even if you're not sure the utterance itself had that pattern of beats. Then listen back to the utterance for confirmation or denial of your intuitive analysis. As you listen to the utterance, if what you hear conflicts with the rhythmic pattern in your mind, then change the rhythm in your mind, to one that is in accord with the actual utterance. Repeat this until you have something that fits. Try to find the best interpretation, even though there may sometimes be more than one interpretation that fits.

When you have a rhythmic pattern for the entire utterance in your mind at once, all the beats in that rhythmic contour are primary beats, coded with stress level 1. Then if there are other syllables that seem strong in addition to the primary beats, for example, when paying attention just to shorter phrases, code them as secondary stresses (level 2). The remaining syllables are ``unstressed''.

This method is most difficult to apply when pause-groups are very short, from one to four syllables. If the entire utterance is pronounced weakly, one may code the strongest beats as secondary stresses. If all the syllables seem prominent, one may try to separate out the pragmatically meaningful intonation contour from the stresses, and consider that some of the prominence of the syllables may derive from the intonation contour's requirements. For example, a high pitch associated with the beginning of a particular intonation contour will be realized on the first syllable in the utterance even if that syllable is unstressed.

There are special concerns to be remembered in coding utterance-final syllables. Pre-pausal lengthening applies to unstressed syllables, as well as stressed ones. Thus in long and short utterances, the last syllable may be lengthened despite being unstressed. Final lengthening does not seem to apply to the last syllable in a false start or a hesitation - this may be the primary phonetic ``cue'' to the ``editing signal'' of Labov (1972:203) and Hindle (1983) associated with false starts.^5.4 One may allow for some lengthening in final syllables, taking into account the effects of the editing-signal: these are not necessarily cues that the final syllable is stressed.

Another procedure that one might have followed in coding the stress contour is to decide the relative prominence of each syllable by making a forced-choice among stress levels, moving one syllable at a time from left to right in the utterance. This method, however, is subjectively very difficult, and it may give unstable results. I found it much easier and more robust to listen for the globally most prominent syllable(s) of the entire utterance.

The analysis is thus, loosely speaking, top-down rather than left-to-right. A final consequence of this method is that primary- and zero-stressed syllables are more common than secondary-stressed syllables, since secondary-stressed syllables are coded in a sort of afterthought.

Two consistency tests were conducted. One examined how consistent I was in coding the same material on two separate occasions; this is discussed in the chapter on Jamaican Creole, page . A second, small consistency test compared two coders. This test was conducted on the second Chicago speaker, Jim C. After ten minutes' training and discussion with a recent graduate of the 1st-year phonetics class,^5.5 128 syllables were classified for three degrees of stress. The results are displayed in a confusion matrix in Table .

Table: Confusion matrix of phrasal stress classifications by two coders.

		Mine
		Primary	Secondary	Unstressed
	Primary	28	2	4
His	Secondary	2	3	1
	Unstressed	2	4	82

There are six 2-level inconsistencies; the number of 1-level inconsistencies is 9, and the consistent classifications number 113/128 for an average of 88.3% cross-coder consistency. This is comparable to the 85% level of intercoder consistency found in Mazzie & Veatch (1986). They found only 65% consistency when coding for four levels of stress. If the secondary-stress level in this data is collapsed together with either primary stress or unstressed, leaving only two levels of stress, then the percentage of inconsistency is reduced from 11.7% to either 8.5%(11 cases, secondary classed with primary stress) or 7.8%(10 cases, secondary stress classed with unstressed). Neither is significantly better than the other, and in the stress data used below, I have collapsed secondary stress together with primary stress, as against unstressed, to form a binary, stressed-vs.-unstressed classification of syllables. Using these results to make an estimate, the degree of inconsistency across coders, making a forced-choice, binary classification, is in the range of 5% to 10%. The self-consistency test, discussed in the Jamaica chapter, page , was conducted on 231 syllables from one of the Jamaican speakers, with myself doing the classification on two occasions, nine months apart. The results are indistinguishable from these: 87% consistency using three levels, 92% consistency using two levels.

There are a number of sources of these inconsistencies. Some of them are simple errors, due to lack of attention, or carelessness,^5.6 while some inconsistencies come from other, more interesting sources. For example, it seems to me that I can often change the stress pattern I hear on a given sentence just by actively thinking of it in a different way. Both subjectively-perceived patterns seem consistent with the utterance itself as I listen to it. This is an unfortunate source of ambiguity.

Finally, the usual cues for stress (pitch inflection, duration, and loudness) may not always be correlated in a given utterance. Extra-long syllables (even in non-final position) may be melodically monotonous and lack a pitch inflection. If in a given sequence of two syllables, the first syllable has a sharp pitch inflection but is quite short, and the second is long and drawn out but has no definite rise or fall of pitch, the different cues lead to different classifications. If the coders are listening for different cues, they will be inconsistent to the degree that the cues to ``stress'' are inconsistent. While many cases are clear and the cues are strong and mutually consistent, many other cases are unclear, to such an extent that is is worth seriously considering whether stress is nothing more than the subjective interpretation of linguists. However, as we see in later chapters, this subjective interpretation correlates well with significant patterns of changes in vowel quality, as reflected in F1, F2 measurements.