Rethinking Reduction

Interdisciplinary Perspectives on Conditions, Mechanisms, and Domains for Phonetic Variation

Francesco Cangemi, Meghan Clayards, Oliver Niebuhr, Barbara Schuppler, Margaret Zellers, Francesco Cangemi, Meghan Clayards, Oliver Niebuhr, Barbara Schuppler, Margaret Zellers

• Editore: De Gruyter Mouton
• Anno: 2018
• ISBN:
  9783110521719
• Edizione: 1
• Argomento: Lingue e linguistica
• Categoria: Fonetica e fonologia

Phonetically reduced forms are plentiful, theoretically interesting, and a key challenge for automatic speech recognition systems. Yet canonical forms are still central to models of production and perception. Drawing from different fields and diverse languages, this volume brings new insights to the debate on abstractions and canonical forms in linguistics: their psychological reality, descriptive adequacy, and technical implementability.

Margaret Zellers, Barbara Schuppler, and Meghan Clayards

1Introduction, or: why rethink reduction?

Margaret Zellers, Kiel University

Barbara Schuppler, Graz University of Technology

Meghan Clayards, McGill University

Abstract: In phonetic reduction, segments may be shorter, less clearly articulated, or absent, compared to “canonical” or dictionary forms. While traditionally considered “slurred” or deficient, recent work has shown that reduction phenomena are highly complex. This chapter takes a historical and multidisciplinary approach, describing how views of reduced speech have evolved over time in the domains of phonetics, speech perception, and automatic speech recognition. We bring these perspectives together to raise several questions: Are reduced forms really inferior to canonical forms? Is the scope of reduction best described at the level of the feature, syllable, or larger unit? Does reduction generally occur in places where the content is more predictable? The chapters of this book are then contextualized with regard to these questions.

1.1Introduction

This book is focused around the phenomenon of phonetic reduction in speech. In phonetic reduction, segments may be shorter, less clearly articulated, or absent, compared to “canonical” or dictionary forms. A classical view on reduction is given by Jakobson and Halle (1956):

Since in various circumstances the distinctive load of the phonemes is actually reduced for the listener, the speaker, in his turn, is relieved of executing all the sound distinctions in his message: the number of effaced features, omitted phonemes and simplified sequences may be considerable in a blurred and rapid style of speaking. The sound shape of speech may be no less elliptic than its syntactic composition.… But, once the necessity arises, speech that is elliptic on the semantic or feature level, is readily translated by the utterer into an explicit form which, if needed, is apprehended by the listener in all its explicitness.

The slurred fashion of pronunciation is but an abbreviated derivative from the explicit clearspeech form which carries the highest amount of information.… When analyzing the patterns of phonemes and distinctive features composing them, one must resort to the fullest, optimal code at the command of the given speakers.

Jakobson and Halle (1956: 6)

Many early studies of reduction take a similar attitude to Jakobson and Halle’s characterization of reduced forms as “slurred,” “slovenly,” or otherwise deficient. However, more recent work has shown clearly that reduction is much more complex. For example, while reduction is primarily thought of as a casual speech phenomenon, it also occurs in read speech, and indeed can make read speech easier to listen to and process; speech synthesis for the blind adopts reduction phenomena to make texts easier to listen to, especially over longer stretches of time (Jande 2003). Furthermore, an increasing amount of evidence suggests that “canonical” and “reduced” forms are not simply categorical oppositions, but may rather fall along a spectrum of pronunciation variants that are more or less clearly articulated (Nolan 1992). An acoustically “absent” segment may still leave prosodic and/ or articulatory traces (Kohler and Niebuhr 2011; Niebuhr and Kohler 2011; Torreira and Ernestus 2011); and conversely, segments may be hyperarticulated to a point that, while their pronunciation may be “super-canonical,” it does not reflect the typical production of that segment (Clayards and Knowles 2015). Schuppler et al. (2012), for instance, found that in conversational Dutch only 11.7% of the tokens show canonical realizations of word-final /t/ (i.e., a voiceless closure followed by one strong burst, produced at an alveolar place of articulation).

While a great deal of current research addresses the kinds of difficulties posed by dealing with reduced forms and spontaneous speech, cf. a special issue of the Journal of Phonetics (39[3], 2011) addressing this very topic, this book will ask the question of whether the ways we think about “reduction” are helpful, and how as researchers we could potentially shift our paradigms and methodologies, leading to greater understanding of this kind of variation in phonetic forms. Thus, this book brings together work from a variety of research and language backgrounds aimed at widening our understanding of what reduction is and how we as language users make use of it.

1.2Examples of reduction

Reduction phenomena can be highly variable, particularly across languages. A few examples are provided here to illustrate some of the possibilities.

Figure 1.1 shows the acoustic realization of the utterance “we were supposed to see yesterday, but he felt really bad.” For native speakers, it is easily understandable, even though it is realized with fewer segments, with only two instead of three syllables, and with different segments than the canonical form. The remaining segments are the initial consonant, the stressed vowels, and the fricative. As frequently shown in the literature, unstressed vowels and plosive closures are deleted completely. With a segment-based approach, this word would be hard to annotate and segment, as boundaries between the segments are overlapping. (However cf. Xu and Liu 2013, who propose that segments, like prosody, are produced as a series of approximations to dynamic targets, and that segmentation is thus preferable on the basis of gesture onsets in the context of syllable structure. Thus, in the current example, the onset of the fricative occurs before the first vowel gesture is complete, as can be seen by the continuation of voicing as well as by the clear formant structure visible within the frication.)

Figure 1.1: yesterday, realiszed as [’jeʃa͡i]; Buckeye corpus

Figure 1.2: natuurlijk, canonical, Ernestus Corpus of Spontaneous Dutch

These examples of natuurlijk (“naturally”) are taken from a Dutch corpus of spontaneous dialogues (Ernestus 2000). In Dutch, natuurlijk can be used with different functions; it can be an adjective, as in “natural languages,” or it can have the discourse-oriented meaning “of course.” If presented in isolation, listeners are not able to understand the reduced form (Van de Ven, Ernestus, and Schreuder 2012), even though the canonical form used as “of course” never occurs in the corpus (Schuppler et al. 2011). When used as an adjective, however, the canonical form is observed frequently. Thus, the same sequence of phones is reduced differently for two different functions.

Figure 1.3: natuurlijk [nt’yk], Ernestus Corpus of Spontaneous Dutch

1.3A historical and multidisciplinary perspective on reduction

1.3.1Phonetics

Like Jakobson and Halle (1956), most early views on reduction coincided in considering reduced forms as lacking in some way. Jakobson and Halle thus argue that reduced forms are not worth studying compared to canonical forms, which contain the “fullest, optimal” information. However, not all contemporary researchers held the same view, and a number of phonetic studies touched on topics relating to reduction, particularly in the context of research on lexical stress. Fry (1955, 1958), for example, reports that lexical stress in American English is marked by duration and intensity ratios between syllables in disyllabic words; that is, reduced duration and intensity are associated with the unstressed syllable. Similarly, Lieberman (1960) reports that stressed syllables in American English have higher fundamental frequency (F0), higher amplitude, and longer duration than unstressed syllables. Research on Swedish by Fant (1962) demonstrates that the formants in unstressed vowels move closer to those typical of schwa (i.e., 500, 1,500, and 2,500 Hz for the first three formants). Fónagy (1966) reports for Hungarian that increased articulatory effort (as in stressed syllables), measured by EEG, is associated with higher formant amplitude and broader bandwidth.

Lindblom (1963) addresses reduction phenomena directly by asserting the existence of “physiologically invariant” vowel targets, which are more or less closely approached based on the articulatory context. He treats reduction as a phenomenon resulting from limits on articulatory speed and claims that timing is more important in influencing reduction than lexical stress; that is, targets are undershot on the basis of decreased duration, not on the basis of lexical stress alone.

As phonetic research on reduction began to move beyond analysis of lexical stress alone, a large body of work also arose in the phonological community, providing contextual analyses that provided rules for certain types of reduction (e.g., schwa allophones of vowels or elision of consonants). Kohler (1974), for example, reports rules for the common elision of schwa from the German suffix -en. Gimson (1977), Elgin (1979), and Lass (1984) provide detailed grammars including elision rules. For Lass, deletion is phonologically defined as a segment merging with a null segment, although he indirectly addresses the phenomenon of reduction by stating that deletion is often the last stage of a lenition process (1984: 187). Since he deals with a phonological process rather than a phonetic one, however, the kinds of deletions he reports are not synchronically variable (in contrast with, for example, Dutch natuurlijk, as discussed in Section 1.2).

Reduction was also addressed from a sociolinguistic perspective in this time period. Labov (1972) reports on, among other phenomena, elision of word-final alveolar plosives in New York Black American English, and social distribution of syllable-final /r/ in department store workers. He does not study reduction as a set of individual phonetic cases, but rather as a way in which people modify their speech to demonstrate group membership. Ohala (1981) studies reduction in the context of explaining sound change over time. His work is related to Lindblom’s on timing, arguing that articulatory timing is constrained by physical characteristics of articulators.

A turning point for the analysis of reduction came with Lisker’s (1984) argument that invariance is not actually something to be expected of the speech signal. He points out that while invariance was considered as important as an explanation for why listeners are able to constantly identify sequences of a limited set of sounds from the continuous speech signal, in fact, “phonemes … are not perceptual constants” (1984: 1201). This argument grew in part out of contemporaneous phonological analyses, in which articulatorily and acoustically different sounds appeared as allophones for the same phoneme. In fact, the argument that contrastiveness, rather than invariance, is essential for speech processing is fundamental for modern work on reduction.

Following this, more phonetic studies of reduction phenomena began to appear. Dalby (1986) characterizes a number of reduction phenomena occurring in fast American English speech. Kohler (1990), moving beyond his earlier phonological analysis, argues that phonology-based characterizations of reduction are too restrictive, and that reduction rules should instead be based on phonetic features, generating a larger variety of alternatives. Kohler specifically argues that phonology, being an abstraction, cannot account for the physical and/ or phonetic processes that lead to reduction, and that therefore it is an insufficient basis for analyzing reduction. Instead, he refers back to “motor economy,” as proposed by Lindblom (1963; see discussion above) and others, as a fundamental consideration.

Lindblom (1990) himself also argues against the “problem” of invariance. His theory of hyper- and hypo-articulation (H&H theory) states that articulation is influenced by both constraints on the production system (i.e., a preference for minimal expenditure of effort) and constraints on the output (i.e., the need to make oneself understood). Since segmental production must take both of these constraints into account, reduction is not a problem but simply a response to a particular set of system settings. Again, the goal is to create sufficient contrast (and thus understandability) rather than to connect to something invariant.

Articulatory Phonology (Browman and Goldstein 1990) provides a somewhat different analysis of reduction, while staying within a similar tradition. In this theory, variation in forms occurs on the basis of varying articulation rate and thus varying degrees of gestural overlap. At extreme degrees of overlap gestures can be “hidden” so that they are not acoustically/perceptually available. For instance, Browman and Goldstein’s measurements showed that word-final /t/ in word combinations like perfect memory could be absent in the acoustic signal, even though the tongue clearly moved to the alveolar ridge. Since all form variations can be accounted for in this model by increased gestural overlap and decreased gestural magnitude (i.e., gestures are never added, changed, or deleted), this model necessarily asserts that all reduction is gradient. Johnson, Flemming, and Wright (1993) follow up on the idea of a single-gradient system for speech sound forms: “[F]ortitions are more accurately seen as descriptions of the pronunciation of phonetic targets in the absence of lenitions, and hence it is apparently the case that for every lenition there is an equal and opposite fortition” (524).

More recently, new insights about ...