2.1 The importance of method

The volume's honorand, John Ohala, provides the opening piece, ‘Methods in phonology’, arguing that a scientific discipline is characterized by ‘the questions it asks; the answers given to the questions, that is, hypotheses and theories; the methods used to marshal evidence in support of the theories’ (3). It is in this third aspect that John Ohala and colleagues depart from e.g. Halle's (Reference Halle1959: 12) view that ‘the methods by which a scientific description are discovered are not of essential concern to the science in question’, i.e. Chomsky's classic rejection of the search for ‘discovery procedures’. While it is certainly true that discoveries and the nature of the evidence for them are more important than the means or process by which they are discovered, it is also true that the questions a science asks are in turn shaped by what can be measured, and how well, which depends on the techniques of the day. John Ohala notes that developments in method ‘constitute the principal engine for refinement and productive change in a discipline, helping to moderate the pace with which one theory supplants another’ (4–5); he mentions the refinement of the kymograph by Marey in the nineteenth century, which made it possible to record relatively rapid physiological events (vocal cord vibration, for example).

To John Ohala's selection of examples let me add a few more. First, the invention of the sound spectrograph enabledFootnote ² the discovery of ‘gravity’, the low second formant frequency that is the common acoustic signature of velar and labial consonants, and back and rounded vowels (Jakobson, Fant & Halle Reference Jakobson, Fant and Halle1951). Without this knowledge, the linguistically natural and frequent association of velarity with labiality (as in double articulations such as [w], [u], [], and a host of sound changes such as [x]>[f]) lacked an explanation. Second, the subsequent invention of ‘pattern-playback’ formant synthesis from tracings of spectrograms paved the way for the discovery of the categorical perception of place of articulation of plosives (Cooper et al. Reference Cooper, Delattre, Liberman, Borst and Gerstman1952, Liberman et al. Reference Liberman1967). The diversity of acoustic cues to alveolar place of articulation led Liberman Reference Liberman1957 to propose the Motor Theory of Speech Perception, a spurFootnote ³ to the move from acoustically-based distinctive features to articulatory distinctive features made by Chomsky & Halle Reference Chomsky and Halle1968. Third, the technology for making permanent physical recordings of sound and movement allows such data to be repeatedly reexamined, measured, and accumulated into large databases, such as the Switchboard corpus (Godfrey, Holliman & McDaniel Reference Godfrey, Holliman and McDaniel1992) or the Kiel Corpus of Spontaneous Speech noted in Klaus J. Kohler's chapter.

Phonological experiments may examine speech, but equally they may examine subjects' responses to speech, or behaviour in a speech-related task. In ‘What's in CVC-like things? Ways and means to look at phonological units across languages’, Bruce L. Derwing reminds us of some of the discoveries of psycholinguistics, and presents some new data from Minnan (Southern Min) Chinese. I cannot agree with his pessimism that ‘it seems highly unlikely that any of these higher-order theoretical entities [segments, syllables, morphemes, word, lexical categories etc.] are ever going to be found even in the “wetware” of a physical brain, much less as part of the physical “stream of speech”’ (327): progress on the ‘wetware’ front – albeit still at a very early stage – is exemplified at a variety of linguistic levels by e.g. Caramazza et al. Reference Caramazza, Chialant, Capasso and Miceli2000 (double dissociation of errors in producing consonants vs. vowels), Damasio & Tranel Reference Damasio and Tranel1993 (double dissociation of errors in retrieval of nouns vs. verbs), Just et al. Reference Just, Carpenter, Keller, Eddy and Thulborn1996 (recruitment of cortical regions in processing embedded clauses), and a steady stream of results using brain imaging methods. In the ‘physical stream of speech’ work in speech and language technology and its application by some to linguistics shows that units of various linguistic categories can be found if one looks (with a theory, of course), such as distinctive features (Coleman Reference Coleman2003), segments (Stevens Reference Stevens2002), and from transcriptions of speech may be discovered morphemes (Goldsmith 2001), parts of speech (Rapp Reference Rapp, Decker and Lenz2007) and so on into other aspects of syntax and semantics. Nevertheless, there is no disagreement that knowledge of language can be and is being appropriately studied through psycholinguistic experiments; this is the one chapter that I would be most likely to set linguistics undergraduates to read. Derwing and co-workers' studies of syllable structure in the minds of Korean and Minnan speakers, including preliterate schoolchildren, clearly demonstrate the language-specific psychological reality of a body+coda (CV.C) syllable structure in those languages, in contrast with the onset+rime (C.VC) syllable structure of English. As I said above, introductory linguistics textbooks will need to be rewritten.

2.2 The place of statistics in phonology

Some textbooks present experimental inference in the scientific method as a fairly straightforward application of logic of the kind: ‘If theory A, then B should happen. B didn't happen, therefore theory A is rejected/not supported’. In reality, statistics and probability theory are indispensable features of experimental inference and argumentation in all sciences, whereas the categories and rules of phonological theory are standardly discrete, not probabilistic or gradient. Statistics is essential in experimental phonology because of endemic variation in everything human. Because an effect might have multiple interacting causes, we begin by attempting to isolate and manipulate factors individually, by holding all other variables constant. For example, when using human subjects, we may regard it as important to use a group that is homogenous with regard to age group, sex, social background, language, dialect, religion, level of education, whether or not they are trained in phonetics/linguistics/psychology, time of day (people get more tired as the evening passes), whether they smoke, use drugs, are on medication, etc. The variables studied are quantitative i.e. counts or measurements. A single set of measurements may be faulty (as a result of inadvertent error), or even if accurate, unrepresentative of other situations, and hence not a suitable basis on which to make general statements. Additionally, natural phenomena are often inherently variable, to a certain degree. We would like to understand and account for that variability, rather than wish it away. To avoid the fragility of single observations, experiments are usually repeated many times e.g. through the use of several subjects, re-runs of the procedure, reproducing the work of other experimenters to confirm or refute their conclusions. Because of multiple trials, a lot of data may be collected germane to a single question. Statistical techniques are then necessary to determine the significance of the mass of data to the question under consideration. Advances in methodology, therefore, may include advances in statistics.

Slips of the tongue are a source of data regarding phonological structure and language processing that is often mentioned anecdotally but only rarely elicited experimentally. In ‘The SLIP technique as a window on the mental preparation of speech: Some methodological considerations’, Sieb Nooteboom & Hugo Quené show us in some detail why it is difficult to obtain sufficient data for such experiments to be effective. A particular problem is that because even elicited slips are quite rare, a large number of subjects needs to be recruited; however, since previous experiments have found that subjects may behave differently, a large number of subjects will introduce between-subject variability, with attendant statistical problems. Nooteboom & Quené propose a new statistical approach (new to this experimental paradigm, that is), multi-nomial logistic regression, applying it to some earlier data, avoiding spuriously significant effects and revealing a new account of what causes the lexical bias in phonological speech errors.

Phonological theorists may be comfortable to allow statistics as a necessary tool of experiments, while maintaining that phonological knowledge is symbolic, not variable or statistical. Some recent work, however, advances the proposition that linguistic knowledge may sometimes be statistical (Saffran, Aslin & Newport 1996, Bybee Reference Bybee2001, Bonatti et al. Reference Bonatti, Peña, Nespor and Mehler2005). For example, one might regard markedness as simply a veiled surrogate for usage frequency. Debate on this question is taken up in Section 2.5 below.

2.3 Timing

Many of the papers in this volume depend on the accurate measurement of time in acoustic or physiological records; measurement vs. disregard of time is one of the hallmark differences between experimental vs. theoretical phonology.

In ‘Articulatory movements and phrase boundaries’, Patrizia Bonaventura & Osamu Fujimura seek the main factors responsible for the larger and longer articulatory movements associated with prosodic phrase edges (e.g. phrase-final lengthening). In the terms of Fujimura's C/D model of articulatory control, they examine the relative role played by the magnitude and speed of articulator excursions, and of the size of the ‘gap’ between syllables – a theory-internal concept that is inferred rather than directly observed, albeit from fleshpoint movement data obtained using X-ray microbeam They find that excursion (i.e. the amount of movement) contributes most to the prediction of speed; syllable duration and boundary magnitude have a lesser effect, for some speakers. Fujimura's earlier claim of a within-speaker, invariant ‘iceberg’ pattern of velocity vs. time for articulator movements requires amendment in the light of this new data: iceberg patterns are invariant after removing a predictable linear effect of excursion.

In ‘Controlled and mechanical properties in speech: A review of the literature’, Maria-Josep Solé surveys a body of work, to which she has made key contributions, in which variation in the fine temporal structure of speech due to changes in speech rate or stress is seen to vary between languages, providing tell-tale information about which aspects of speech can or cannot be under the control of the speaker's nervous system and conversely, which aspects of timing are fixed by physics or physiology. For example, the duration of vowel nasalisation before a nasal consonant varies with speech rate in American English, but not in Spanish (where it is roughly constant in duration), from which Solé infers that in Spanish, vowel nasalisation is a low-level coarticulatory effect due to the time required to lower the velum for the up-coming nasal consonant, whereas in American English vowel nasalisation ‘vowel nasalization has become part of the language-specific instructions’. This and the various other cases in Solé's survey are extremely informative at teasing apart ‘intentional’ from ‘mechanical’ aspects of speech; though not quite synonymous with ‘phonological’ vs. ‘phonetic’, this is perhaps a more natural, less theory-dependent distinction.

Timing is also the focus of the six experiments presented by Manjari Ohala in ‘Experimental methods in the study of Hindi geminate consonants’. In the most striking of these, Experiment 2, she finds that the consonant preceding the vowel that precedes a geminate is longer than in the corresponding non-geminate condition. That is, gemination in Hindi is a prosody, in this case a pattern of durational differences between words and without medial geminates that is evidenced throughout the initial syllable. The experiment was inspired by Local & Simpson's (Reference Local, Simpson, Ohala, Hasegawa, Ohala and Granville1999) discovery of a similar non-local effect on the initial consonants of words containing geminates in Malayalam. But whereas Local & Simpson found shorter initial consonants with Malayalam medial geminates (perhaps suggesting a mechanism of temporal trade-off between the durations of successive vowels and consonants), the effect in Hindi goes in the other direction, suggesting a planned lengthening of the word-initial syllable.

The existence of (gradient) trading relations is one of the most significant discoveries concerning the relationship between phonological categories and observed phonetic forms. In ‘Applying perceptual methods to the study of phonetic variation and sound change’, Patrice Speeter Beddor, Anthony Brasher & Chandan Narayan present a perceptual experiment in which subjects identified tokens of bet, bed, bent and bend made with different proportions of vowel nasalization and nasal consonant duration, the two cues in some cases conflicting, in others cooperating. They found that some listeners treated vowel nasality and nasal consonant duration as perceptually equivalent, others did so only before /d/, and a third group gave more weight to vowel nasality than to nasal consonant duration. The pattern of trading relations and its variability between groups of listeners offers a promising explanation for the common historical change VN>ṽ, as J. Ohala has proposed. M. Grazia Busà's chapter, ‘Coarticulatory nasalization and phonological developments: Data from Italian and English nasal–fricative sequences’, takes this topic further, presenting aerometric data of oral vs. nasal timing in nasal+obstruent sequences in Northern vs. Central Italian and American English, finding (as expected) differences between all three. Northern Italian has a greater extent of anticipatory vowel nasalization, whereas Central Italian and American English a greater tendency to end nasality early, leading to a short stop closure between the nasal and following obstruent. Although both patterns are observed in both languages, it is the pattern with higher usage frequency that determines the direction of the sound change.

2.4 Acoustic vs. articulatory categories

Juliette Blevins's ‘Interpreting misperception: Beauty is in the ear of the beholder’ defends the importance of two aspects of John Ohala's mode of explanation of synchronic sound patterns: (i) misperception, which leads a language learner to think that a sound was produced in a different way than the speaker articulated it; and (ii) the non-teleological nature of sound changes, and the unnecessity of synchronic, structural markedness constraints. Misperception is well exemplified by nasal place assimilation to following oral consonants, and by the common sound change from dorsal [k^j] or [c] to coronal [ʧ], which Guion (Reference Guion1998) explains in terms of the acoustic similarity of [c] to [ʧ]. Blevins reports that Halle (Reference Halle2004) discounted Guion's acoustic and perceptual findings, preferring to explain the sound change solely in terms of articulatory feature geometry. Blevins remarks:

In contrast, Blevins's Evolutionary Phonology replaces synchronic teleology by diachronic phonetic developments which at each step might be motivated by perceptual, articulatory and/or lexical constraints. Synchronic variation is seen as offering a range of exemplars, allowing for different and sometimes ultimately contradictory patterns of development in different languages; for example, unstressed mid vowels may historically raise or lower in some languages (e.g. Bulgarian vs. Belorussian), yet in other cases unstressed high or low vowels may become mid, an apparent paradox presenting difficulties for a putative synchronic constraint relating stress to vowel quality. Blevins observes that the non-determinism of diachronic development sits naturally with stochastic views of phonology, such as that of Pierrehumbert (Reference Pierrehumbert, Gussenhoven and Warner2002, Reference Pierrehumbert, Bod, Hay and Jannedy2003) or Keith Johnson's chapter.

Guion's acoustic explanation of the [c] >[ʧ] sound change (that they sound similar), cited with approbation by Blevins, is more convincing than the articulatory explanation (that their places of articulation are in close proximity) because the prevailing scheme of articulatory classification in phonological theory allots [c] and [ʧ] to quite distinct ‘articulators’, i.e. dorsal vs. coronal. Certainly, one might easily introduce a new articulatory category based on their common passive articulator, e.g. [palatal], but such a radical addition to current conceptions of articulatory geometry would be as conceptually ‘costly’ as simply restoring the acoustic feature [acute].

In ‘A perceptual bridge between coronal and dorsal /r/’, Olle Engstrand, Johan Frid & Björn Lindblom use acoustic measurements, articulatory modelling, acoustic waveform synthesis and perceptual testing to examine the historical development of coronal>dorsal /r/ (e.g. [r]>[R] and their approximant congeners [ɹ] [ɰ], etc.) and the resultant synchronic variation common in many European languages. Through a detailed examination of their acoustics they show that despite being articulatorily quite distinct, coronal and dorsal variants lie on an acoustic continuum. Interestingly (though the authors do not comment on this), it is the frontest (alveolar) approximants, e.g. [] not the postalveolars or retroflexes, which are acoustically the most similar of the coronals to prevelar approximants []. We have here positive evidence for restoration of the acoustic categories acute, i.e. [–grave], and [±flat]: in Jakobsonian terms both [] and [] are [–grave] and [–flat], as opposed to the [+grave] uvulars and pharyngeals, and to the [+flat] postalveolars and retroflexes.Footnote ⁴ Their categorical perception experiment shows that Swedish listeners discriminate rather poorly and non-categorically between coronal vs. dorsal /r/ (as expected, since they are not phonemically distinct in Swedish nor any known living language). The authors' sophisticated articulatory modelling yields a good match to the acoustic measurements for dorsal and retroflex approximants, but for alveolars the fit is poor, demonstrating that while articulatory models are interesting and important, they can be quite inaccurate in their acoustic output, compared to real speech. For this reason, we should be wary of over-enthusiasm about articulatory modelling.

2.5 Abstract symbols vs. detailed examples

During the last decade, exemplar-based phonology has taken root as a prominent alternative to generative phonology (including Optimality Theory). The exemplar-based approach is concerned with the cognitive grounding of phonological knowledge: although phonologists have long held that phonological forms in the mental lexicon, as well as templates, constraints and rules, are to be expressed in an ‘alphabet’ of symbols that excludes the minutiae of specific instances, exemplar-based models envisage the storage of instances as they occur, with all their specific, individual details. New experiences are recognized as being similar to previous experiences by assessing the similarity of the new instance to the ‘cloud’ of stored exemplars. In essence, it is corpus-based phonology applied to the human mind. This approach has been assessed via various computational models, which are reviewed by Keith Johnson in ‘Decisions and mechanisms in exemplar-based phonology’. It derives increasing support from experiments on human behaviour in speech perception and production, and has captured the interest of experimental phonologists partly because it offers mechanisms for explaining within-category variation in phonetic detail (thus also providing a foundation for sociophonetic variation) as well as a basis for abstraction of types and between-category differences. The key, exciting evidence comes from experiments that show: (a) the plasticity of category boundaries, subject to various external factors (especially usage frequency differences between specific words, e.g. Pierrehumbert Reference Pierrehumbert, Gussenhoven and Warner2002, Wright Reference Wright, Local, Ogden and Temple2003); Johnson shows (30–31) that vowels are reduced in different proportions in homophones with different usage frequencies; (b) listener sensitivity to within.category differences (e.g. Beddor, Brasher & Narayan's chapter); and (c) listener sensitivity to fine details – sometimes arbitrary or speaker-specific – which abstractionist approaches such as phonemics and generative phonology usually exclude as linguistically irrelevant (but cf. Terrance Nearey & Peter Assmann's chapter). For example, Goldinger Reference Goldinger1996 discovered that lexical memory is sensitive to voice identity, words uttered by an unknown speaker (one of 10) on a single previous occasion a week earlier being remembered significantly more accurately than those produced by a new speaker, a result which flies in the face of the conventional expectation of phonologists (following Trubetzkoy Reference Trubetzkoy1939 [1969: 15–16]) that the details of a speaker's voice are ‘abstracted away’ by the listener, and form no part of the phonological representation of words. Hay, Warren & Drager Reference Hay, Warren and Drager2006 discovered that in identifying diphthongs that are distinct in American English but nearly merged in New Zealand English, New Zealand listeners were biased by the experimenter's accent. In a subsequent experiment, Hay & Drager Reference Hay and Drager2010 found subjects' vowel perception to be similarly biased by the national symbolism of stuffed toys (kiwis vs. kangaroos) that were merely present in the room, the explanation being that the subject's category boundary is biased by the dialect primed by the toy, even though the subjects knew the speaker to be from New Zealand. This malleability of phonological category boundaries can be explained by reference to a pool of memories of prior experiences, but from an abstractionist view it is utterly unexpected and counterintuitive. The linkage of phonetic and non-phonetic – even non-linguistic – properties, such as visual and sociolinguistic properties of experiences, has a top–down effect whereby expectations alter perception. As in generative phonology, there is no single, complete statement of ‘Exemplar-based Phonology’, but a variety of intersecting proposals (e.g. Blevins's ‘Evolutionary Phonology’ and Pierrehumbert's ‘Stochastic Phonology’) consistent with the central idea that specific experiences are memorized and used in recall, classification, etc. It seems that laboratory/experimental phonologists have at last begun to articulate a coherent and empirically preferable competitor to generative linguistics.

The new trend has its critics: in ‘Probabilistic “sliding template” models for indirect vowel normalization’, Terrance M. Nearey & Peter F. Assmann walk us through a succession of methods of normalization of speaker-specific variation in vowel acoustics, which they then competitively evaluate against data from three multi-speaker databases. To the extent that some of the models provide reasonably good vowel identification accuracyFootnote ⁵ and alternative accounts to some aspects of Johnson's model, this chapter offers a puissant defence of the abstractionist approach.

Eurie Shin's findings in ‘How do listeners compensate for phonology?’ are also claimed to support the generative view that (in this case) Korean listeners know the phonological rules that map underlying phonological forms onto observed speech signals, for example accepting intervocalic /s.p/ as a more likely source of [p.p] than /k.p/. The fact that /p.t/ or /k.t/ were sometimes considered by listeners to be possible sources of [t.t] (even though none of the [t.t] stimuli were assimilated forms) suggests to Shin that the subjects know and use the assimilation rules in processing the stimuli in these experiments – over-apply them, in fact. But this is not an argument to convince a doubter: for one thing, it disregards the bias caused by the tendency for subjects in a multiple-choice response format to use all of the available responses some of the time. Whether this could be an explanatory factor deserves investigation.

Two chapters seek to defend traditional generative phonology, while critically engaging with experimental phonology. In ‘Morphophonemics and the lexicon: A case study from Turkish’, Anne Pycha, Sharon Inkelas & Ronald Sprouse employ a corpus, the Turkish Electronic Living Lexicon, to study alternating (e.g. kanat – kanada ‘wing’ vs. ‘wing.dat’) vs. non-alternating (e.g. sanat – sanata ‘art’ vs. ‘art.dat’) words. They show that neither usage frequency, neighbourhood density, cohort size nor etymology effectively explains which forms alternate and which forms do not. But they then precipitously conclude that ‘[t]he more insightful analysis … thus appears to be the generative one, in which the underlying representation of each root contains the information necessary to predict its grammatical behavior’ (385). This does not follow: in order to learn such underlying representations and the rule of alternation, the Turkish learner needs to know the relevant word-pairs; in fact, simply memorizing them is both a necessary and sufficient basis for ‘explaining’ which stems alternate. There is no need for a learner to decompose the words into stem+suffix, abstract a single underlying stem form, infer the rule for the alternating forms and mark the non-alternating stems as resisting the rule, as the generative analysis would require, so by Occam's razor we should prefer simple memorization: the corpus/exemplar-based account proves stronger because it explains the same data without postulating unseen rules and underlying forms.