Multiple targets of phrase-final lengthening in American English words

Abstract

Boundary-related lengthening has been shown to affect the phrase-final word in a number of languages, but its precise distribution within the final word has not been determined. Some evidence suggests that it can begin before the final syllable (e.g. in an earlier main-stress syllable), and that it may be progressive (e.g. may affect the coda of the final syllable more than the nucleus and the nucleus more than the onset). However, only a small number of word shapes have been examined in any one language, so the available facts under-determine models of the duration adjustment process. A survey of final lengthening in words with various stress patterns in American English, using acoustic measures, shows that, in the conditions of these experiments, (a) although most of the duration increase occurs in the phrase-final syllable rime, statistically significant lengthening of 7–18% also occurs in the main-stress syllable rime, when the main stress syllable is not the final syllable, (b) this pattern is seen in both pitch-accented and unaccented final words, suggesting that it is not the result of nuclear-accent-related lengthening, (c) the distribution of lengthening across the syllables of the final word is not straightforward, in the sense that regions between the main-stress rime and the final rime appear to be skipped or lengthened less than the regions before and after them. These results suggest that the mechanism of boundary-related lengthening is more complex than current models propose; in particular, its distribution cannot be explained without reference to the location of main lexical stress and appears to involve more than one stretch of speech, at least in American English.

Introduction

One of the primary correlates of constituent structure in spoken utterances is the pattern of duration adjustment near constituent boundaries, i.e. boundary-related lengthening. This study addresses one particular aspect of boundary-related lengthening, that is, the question of what stretch or domain of material at the end of an intonational phrase is affected by this process.

The literature suggests three major approaches to specifying the domain over which speakers adjust phrase-final duration, which can be termed the Structure-based, Content-based and Hybrid views. On the Structure-based view, final lengthening affects a stretch of speech defined by linguistic structure (e.g. a linguistic constituent such as the final-syllable rime, like -ine in Brookline, or the ‘word rime’ beginning with the vowel of the main lexical stress syllable and including all of the material following it up to the word boundary, such as -ookline in Brookline). Whatever the unit, the Structure-based view suggests a fixed domain, in the sense that the region affected by final lengthening is a structurally similar region for all phrases, even though its phonological content, i.e. its number of segments and syllables, can vary depending on the word (cf. the final rime -a of Cuba vs. -urst of Parkhurst). The discussions of lengthening of the phrase-final syllable in Klatt (1975) and of the final-syllable rime in Wightman, Shattuck-Hufnagel, Ostendorf, and Price (1992) are consistent with the Structure-based view, although these investigators did not evaluate possible alternative Structure-based domains.

On the Content-based view, the domain of lengthening is structurally variable, because its extent is determined by properties of the last segment or syllable of the phrase. For example, a variable domain could result from a lengthening gesture of fixed duration that overlaps with a greater or lesser portion of the final word depending on the number of final segments and the intrinsic length (Byrd & Saltzman, 2003; Byrd, Lee, Riggs, & Adams, 2005; Byrd, Krivokapic, & Lee, 2006). This version of the Content-based view might be termed the Overlap hypothesis, in the sense that the duration-governing element lengthens whatever part of the word that it overlaps with. Their proposal, formulated within the framework of Articulatory Phonology (Browman & Goldstein, 1992), invokes a duration-governing element called a Pi-gesture. The Pi-gesture is not itself a representation of the action of an articulator or the specification of a vocal-tract configuration, but instead is a phonological element of the gestural score; its shape describes the time period during which the articulations of the segmental gestures that it overlaps with will be slowed, as well as how much they will be slowed. That is, the Pi ‘gesture’ indicates the onset and offset of a time period during which the clock that controls the timing of other gestures ticks more slowly, and its height at any particular point in time specifies the degree of slowing at that time. If we assume that the Pi-gesture is of fixed underlying duration for all boundaries of a given type and aligns at its center¹ with the phrase boundary (Byrd, Lee, Riggs, & Adams (2005), Byrd, Krivokapic, & Lee (2006)), then for words with inherently short final gestures, the Pi-gesture is more likely to overlap with earlier portions in the final word and thus to slow them down. As a result, words with gesturally shorter and less complex final syllables (e.g. those containing lax vowels, high vowels or codas with only one consonant or even none) are more likely to show lengthening effects on earlier structural portions of the final word.² For example, lengthening might begin earlier than the final syllable in words like Cuba or Kenneth, but later in the final syllable in words like Brookline or London's. This prediction of the Overlap model stands in sharp contrast to predictions of lengthening over the same structurally defined region, such as the final-syllable rime (e.g. -a, -eth, -ine and -on's in these example words).

The third view of the lengthening domain is a Hybrid one, since it proposes that a fixed, structurally defined domain is normally lengthened in phrase-final position, but that phonological and phonetic properties of the final syllable determine whether additional, earlier lengthening can occur. This type of view is expressed by Camber-Langeveld (1997), who proposes that phrase-finality requires a certain amount of lengthening that occurs preferentially on the phrase-final syllable rime; whether or not final lengthening also affects earlier regions depends on the expandability of that final syllable rime. Thus, final lengthening can occur earlier than the final syllable only when the final syllable contains segments that are contrastively short and so cannot be expanded enough to accommodate an adequate amount of final lengthening. This might be termed the Expandability hypothesis: the final syllable is lengthened if it is expandable given the categories that it contrasts with in the language, but if not, then earlier syllables must be lengthened as well.

We will test these alternative views by comparing where lengthening begins in American English words that vary in their number of syllables, stress pattern and phonological composition of the final syllable. This analysis will reveal whether lengthening affects a stretch of speech defined by aspects of linguistic structure (e.g. by the location of main stress as predicted by the Word Rime hypothesis), or affects a structure-independent domain that varies depending on the phonological content of the phrase-final syllable. The two hypotheses that suggest a variable domain determined at least in part by the content of the final syllable (Overlap and Expandability) will be tested by comparing lengthening in full-vowel final syllables with lengthening in final syllables that contain schwa (intrinsically short in duration and proposed by Cambier-Langeveld to be less expandable than full vowels, at least in Dutch). In addition, we will compare the relative magnitude of lengthening on successive elements within the final word, to determine whether there is a constant degree of lengthening on all elements within a single domain, an increasing degree of lengthening as the boundary is approached, or an interrupted pattern of lengthening so that some portions of the domain are affected more than others.

Together these two kinds of analysis will provide some insight into an additional question: what is the mechanism that speakers use to compute duration? For example, a pattern of progressive lengthening across the subconstituents of the final word would be consistent with a motor-based model in which a speaker must slow down more and more in order to stop at the boundary. A finding of constant lengthening across all of the subconstituents of the domain would be more consistent with the marking of a constituent for lengthening at a more abstract level of representation. And if duration lengthening is interrupted, i.e. if certain elements within the domain are lengthened less or not at all, it will suggest the need for a more complex model of the final lengthening mechanism than current views provide. In the following sections we review literature relevant to the question of where lengthening begins (Section 1.1) and how it is distributed (Section 1.2), and we discuss implications of the answers to these two questions (Section 1.3).

Since most investigators assume that lengthening is concentrated in the final syllable of the phrase, only a limited number of studies have looked at earlier syllables of the final word. Among those that have done so, results are somewhat conflicting with respect to where final lengthening begins. Studies of languages other than English, looking at words with penultimate or antepenultimate main stress, are generally consistent with the hypothesis that lengthening begins at the main-stress syllable and affects all of the succeeding portions of the word (the Word Rime hypothesis), although their results are subject to alternative interpretations because so few word types are sampled. These studies include Kohler (1983) for German, Cambier-Langeveld (1997), Cambier-Langeveld (2000) for Dutch, Krull (1997) for Estonian, and Berkovits (1994) for Hebrew (although pre-final main stress is rare in Hebrew, so the target words with this stress pattern may not reflect the usual treatment of duration in this language).

Studies of English offer a somewhat conflicting picture. Laboratory studies of British English are also generally consistent with the view that final lengthening begins at the main-stressed syllable: Cambier-Langeveld (2000) found final lengthening on the main-stressed syllable as well as the final syllable in words like Johnny and Joseph, and White (2002) found that final lengthening could begin as early as a penultimate or even an antepenultimate main-stressed syllable coda (e.g. on [k] of spectre and spectacle). However, Byrd et al.'s (2006) articulatory study of the word dodo spoken by American English speakers found evidence of penultimate main-stress syllable lengthening for only one out of four speakers investigated, and Wightman et al.'s (1992) corpus study of radio news speech in American English suggests no hierarchically governed lengthening earlier than the final syllable. This may be in part because Wightman et al. did not separate out words that actually have main stress earlier than the final syllable (which is the only word population that provides a test of the hypothesis), or because a corpus study, which by definition has no controls for context variation, may not be ideal for investigating small duration effects that might get lost in the noise.

Finally, there are conflicting results for the possibility of lengthening even earlier than the main-stress syllable in Hebrew and German. Berkovits (1994) found no lengthening on unstressed penultimate syllables that precede a final main-stress syllable (e.g. on gi- of giTIT, where upper case signifies main stress) in Hebrew, while Silverman (1990) found lengthening on the um- of umLAgern in German, although this lengthening was of lesser magnitude than the lengthening on main-stress syllable UM- of UMlagern.

Taken together, these findings are broadly consistent with a Structure-based hypothesis in which phrase-final lengthening begins at the main-stress syllable in many languages. This might be called the Word Rime hypothesis. However, some of the findings are also consistent with the Content-based or Hybrid views of the domain of phrase-final lengthening. For example, Cambier-Langeveld's (2000) finding of lengthening for both syllables of words like Johnny and Joseph could be explained either on the basis of a structural domain that begins at the main-stress syllable and continues to the end of the word, or as early lengthening that occurs because the vowel of the final syllable is phonetically short, i.e. high (and therefore short because it requires less jaw movement) or reduced.

Moreover, there are other results from Dutch that support the Content-based and Hybrid views. For example, Cambier-Langeveld (1997) reports that lengthening begins earlier than the final syllable when that syllable contains a reduced (i.e. phonetically short) vowel, as in mode or tandem (whose final syllables contain reduced vowels in Dutch). In contrast, this earlier lengthening does not occur when the final syllable contains a low (phonetically long) full vowel, as in yucca or harmonica. These results are consistent with Cambier-Langeveld's Expandability hypothesis (because long vowels may lengthen without endangering contrasts, while short vowels may not), and with Byrd and Saltzman's (2003) Overlap hypothesis (because the Pi-gesture is more likely to overlap an earlier syllable like mo- or tan- if the final syllable contains phonetically short segmental material like -de or –dem). These results are difficult to reconcile with a Structure-based view no matter what the hypothesized structural constituent, since that view does not predict any re-definition of the lengthened constituent due to the type or number of segments at the end of the word.

In sum, the available data under-determine the choice among theories of where speakers begin phrase-final lengthening; to determine the most appropriate theory, data are needed from sets of words in which both the location of the main-stress syllable and the phonological content of the final syllable are varied systematically. The studies described in this paper attempt to determine the most appropriate account of final lengthening in American English, by examining words with a variety of numbers of syllables, stress patterns and vowel qualities in the final syllable. The behaviour of words that vary in the position of main stress will provide a direct test of the hypothesis that lengthening begins at the lexical main-stress syllable, and the behaviour of words that have short vs. long vowels in their final syllables will test the hypothesis that lengthening begins earlier if the final syllable is phonetically short (i.e. unexpandable and/or more subject to overlap with a Pi-gesture). In addition to addressing the question of the point of onset of phrase-final lengthening, the experiments also address a related question: once it has begun, how is lengthening distributed between its onset point and the end of the phrase?

Just as for the question of where final lengthening begins, there are several proposed answers to the question of how it is distributed. One possibility is that all lengthened portions of the final word lengthen equally. Several studies suggest that this is not the case, although acoustic measurement criteria, e.g. whether vowel durations included preceding VOT or not, may have differed across studies and may have affected the pattern and/or magnitude of reported results. For example, Kohler (1983) found that, for German words with penultimate main stress, the magnitude of lengthening on the final syllable was greatest (ranging from 87% to 176%), while lengthening on the immediately preceding main-stress syllable was of a lesser magnitude (ranging only between 15% and 31%). Silverman's (1990) study of German showed that, for the word umLAgern, the pre-main-stress syllable um- lengthened less than the later (but still non-final) main-stress syllable -LA- (6–7% vs. 11–17%). Both of these findings are consistent with a second hypothesis about the distribution of final lengthening: that it increases progressively over the course of the final word. Berkovits (1994) provided additional evidence for progressive lengthening from Hebrew. She showed that in disyllabic words with initial main stress, such as DUdik, more of the phrase-final lengthening occurs on the phrase-final rhyme than on the penultimate main-stress syllable. Berkovits (1993a), Berkovits (1993b), Berkovits (1994) also found progressive lengthening within the final syllable for Hebrew, i.e. codas show proportionally more lengthening than nuclei, at least for words with final stops and fricatives. In addition, Byrd et al.'s (2006) study of the American English disyllabic word dodo showed that when closing gestures leading up to the second [d] in that word were longer in phrase-final vs. phrase-medial position (for one of their subjects), they lengthened less than later pre-boundary gestures.

Another line of evidence consistent with the Progressive Lengthening hypothesis of increased lengthening as the speaker approaches the boundary is that effects of the prosodic hierarchy are more reliably detected in elements of the final syllable than in earlier portions of the final word. For Dutch, Cambier-Langeveld (1997) reports that it is the word-final segment (rather than earlier segments) that most strikingly reflects the four different levels of constituents in the prosodic hierarchy (i.e. smallest amount of lengthening for the Prosodic Word, with successive increases for the successively higher-level constituents of Phonological Phrase, Intonational Phrase and Utterance). If lengthening is greatest toward the end of the final syllable, then that is the location where it will be easiest to detect a hierarchical effect. Wightman et al. (1992) found similar evidence in English, where hierarchical lengthening was more reliable in the rhyme of the final syllable than in its onset.

Thus there are two aspects to the progressive lengthening hypothesis: within the final word, successive syllables may be lengthened more, and within the final syllable, successive syllabic subcomponents may be lengthened more. In this study, we test these two aspects of the hypothesis separately, by comparing duration changes of successive elements throughout phrase-final target words that have varying numbers of syllables and stress patterns.

Knowing the onset point of final lengthening and whether it increases toward the boundary still does not answer the question of how lengthening is distributed throughout the final word: Namely, does it affect a single continuous domain? In support of what might be called the Multiple Domain hypothesis, there is evidence that some elements of the phrase-final word are skipped or lengthened less than their surrounding elements. For example, in her study of 5 words in read laboratory speech in Dutch, Cambier-Langeveld (1997) found that, when a reduced (i.e. unexpandable) final syllable resulted in lengthening on the preceding syllable, duration measures for some intervening segments were not affected, e.g. the /n/ of tandem and the /d/ of mode were not lengthened phrase finally, although earlier elements (ta- and mo-) and later elements (-em and -e) were. These patterns are not consistent with the strong view that progressive lengthening operates over a single continuous domain, but they are consistent with the view that wherever lengthening does occur, it is greater in magnitude the closer it is to the boundary. The observation that some elements of the final word are skipped by the lengthening process (or show insignificant effects) leaves open the question of what kinds of elements can be treated in this way. In sum, like the question of where final lengthening begins, the question of how it is distributed over its domain is not adequately addressed in the literature. The present study is designed to resolve this issue for American English.

In this study, we examine the effects of lexical stress patterns on final lengthening by comparing the effects of an Intonational Phrase boundary on different types of phrase-final target words: one- and two-syllable words with final main stress (e.g. Maine, Bangkok, Tibet), two-, three- and four-syllable words with penultimate main stress (e.g. Kenneth, Brookline, Manassas, Chappaquiddick) and three-syllable words with antepenultimate stress (e.g. Mendelson, Thomasburg). We also examine the effects of varying the phonological nature and intrinsic duration of the final syllable, comparing intrinsically long full-vowel final syllables (e.g. Charlestown) with reduced-vowel final syllables (e.g. Thomas). By using a number of word types and tokens, we hope to discover systematic patterns that are not revealed by smaller word-type samples.

Answers to questions about the onset and distribution of pre-boundary lengthening in American English will be useful for several reasons. First, this information will help to distinguish among competing models of speech production planning that make contrasting predictions about lengthening patterns because they invoke different representations and mechanisms for the computation of duration adjustments (i.e. the Structure-based Word Rime hypothesis, the Content-based Overlap hypothesis, and the Hybrid Expandability hypothesis). Second, knowing what portions of the final word are lengthened and by how much will make possible improvements in the naturalness of synthesized speech. A third useful aspect is that this information can be incorporated into the design of experiments to distinguish among models of the perception of phrase boundaries. For example, one theory proposes that listeners perceive phrase boundaries when they perceive a disruption in the quasi-regular global rhythm of phrasal stresses due to duration lengthening (Lehiste, 1977; Scott, 1982). Other possibilities depend more on detecting local boundary-related cues such as lengthened duration of boundary-adjacent phonemic material, relative to the temporal characteristics of either (1) immediately adjacent material, or (2) stored non-boundary-adjacent prototypes or exemplars. It will not be possible to design critical tests of the global rhythmic disruption view vs. the local cue view of boundary perception until we know with certainty where boundary-related lengthening occurs in normal utterances. This will permit the construction of stimuli in which duration adjustments are restricted to segments within the cross-word foot that are not normally affected by boundary-related lengthening in normal speech. Such stimuli will provide a strong test of the somewhat counter-intuitive but interesting hypothesis that it is rhythm disruption per se that cues a boundary rather than lengthening of any particular constituent or at any particular location. If this view is correct, then lengthening in any portion of the target foot (and not just the portions that are normally lengthened by speakers) should encourage the perception of a boundary within that foot, because it disrupts the rhythm.