This book presents a proposal for analysing English conversational prosody in terms of ‘prosodic constructions’ that play a systematic role in regulating the flow of the conversation and conveying non-verbal messages such as speaker attitude. The first half of the book is organised around a few specific constructions (and other ‘prosodic’ effects, in particular creaky voice). After a brief introduction, there are two substantial (twenty pages each) chapters on the ‘Bookended Narrow Pitch Construction’ and on ‘Downstep Constructions’, followed by a fifteen-page chapter on creaky voice and the way it interacts with effects of the sort conveyed by the constructions. There follows a brief review of past work (chapter 5) and then another twenty-page chapter (chapter 6) on ‘Late pitch peak and its functions’ and a brief chapter (chapter 7) on ‘Expressing positive assessments’.

There is then a change of scene as Ward introduces the more technical aspects of his work in chapters 8–10; the material is technical, but the presentation is deliberately informal and unthreatening to non-mathematicians. The heart of his methodology is Principal Component Analysis (PCA). PCA is a widely used statistical technique for exploratory analysis of data that are characterized by complex patterns of variation along many dimensions; it works by identifying underlying factors (principal components) that account for the variability in the data. Ward applies PCA to measurements of the acoustic signal (various appropriately normalised measures of intensity, pitch and so on) and then looks for functional commonalities in recurring patterns of features that are identified by PCA. Crucially, the prosodic measurements are not applied to pitch contours, conversational turns, or other stretches of speech chosen a priori, but are simply taken at fixed intervals (typically 10 or 20 milliseconds) throughout a whole conversation. This is unlike the approach taken by certain other research groups, such as Gubian et al. (Reference Gubian, Torreira and Boves2015) or Lohfink et al. (Reference Lohfink, Katsika, Arvaniti, Sasha, Paola, Marija and Paul2019), who are exploring applications of PCA to the description of pitch contours. Work of that sort first identifies a set of contour types and uses PCA to refine the phonetic description of the types, whereas Ward is using PCA to help discover a suitable set of functional types that have shared acoustic properties. A key feature of his method is that it analyses the speech signal of both participants in a dialogue simultaneously, which makes it possible to identify acoustic signatures of such events as turn-switches and back-channels.

Having presented the technical heart of his project, Ward moves on to illustrate its potential in investigating the role of prosody in the management of conversation: turn-taking (chapter 11), topic management (chapter 12) and speaker ‘stance’ (chapter 13). He then rounds things off with a four-page summary of ‘The rest of English prosody’ (chapter 14) and a similarly brief ‘Envoi’ (chapter 15).

The book is engagingly written and carefully produced by Cambridge University Press. There are two minor but notable issues with the production standards. First, even in the list of references, all but three or four foreign names have been stripped of any diacritics they might once have had (e.g. Vaissière, Martínez, Möbius, Dehé, Benuš). Second, the endnote numbering throughout chapter 2 is badly mixed up, beginning with two occurrences of ‘endnote 4’ on pp. 9 and 10, and compounded by what appears to be an erroneously duplicated endnote (‘13’ and ‘14’) on p. 199; since most of the endnotes consist of nothing more than an author's name and a date, it takes some doing to work out which number goes with which endnote.

The book is complemented by a website containing not only various more technical supplementary materials, but also sound files of the examples, which unquestionably help the reader get a clear sense of what is being discussed. Importantly, the book is inspired by Ward's work in speech technology, but it displays no trace of the negative attitude to linguistics sometimes attributed to speech technologists (Taylor Reference Taylor2009: sec. 18.1). Ward has read very widely in the literature on linguistics and allied fields (including Conversational Analysis and psycholinguistics), and there is no doubt about the book's ambitious scope. Nevertheless, there is little actual engagement with past linguistic analyses; Ward's main objective is to present his own account of how it all works.

I read most of the book before noticing Nanette Veilleux's blurb on the back cover, which describes Ward's approach as ‘novel and defiantly speculative’. This is a near-perfect summary. The novelty resides primarily in the way Ward uses PCA to identify prosodic constructions. The speculation is seen in the way the communicative functions are described and classified (more on this below). And the defiance is seen in the style of presentation. Ward writes that he intends the book to be ‘both scholarly and readable’ (p. x), but he decidedly errs in the direction of readability and anecdote. Many of the trappings of conventional scholarly writing are absent. The review of past work is brief and marred by significant errors. There are no in-text references, only copious endnotes consisting mostly of a reference or two. He completely avoids the use of any standard notation system, even the IPA, though he happily introduces a small number of informal notational devices such as asterisks for *creaky voice* and a little balloon-like icon to indicate ‘late peak’. And a theme running throughout the book is that he is merely reporting the facts and not getting carried away with flights of theoretical fancy: ‘this book focuses on the facts, not issues of theory’ (p. 3). ‘The coexistence of paralinguistic and linguistic aspects of prosody] is not a problem for us in this book, since our aim is merely to describe the facts as accurately as possible . . . ’ (p. 61). ‘. . . this method is respectful of the uncertainty and complexity of real dialog . . . [I]t works in terms of evidence . . .’ (p. 90).

Yet fundamentally, Ward is presenting a taxonomy: a classification and description of a number of prosodic phenomena that contribute to the overall communicative force of an utterance. While he may be presenting facts about the constructions he describes, the constructions themselves are his own analytic constructs. Here and there he acknowledges this:

Given this acknowledgement, it is surprising how little attention he devotes to any sort of validation or to explicit comparison with other analyses. Instead, much of the book is devoted to extended exegesis of specific examples and cases. He identifies major construction types and ‘families’ of constructions and describes their pragmatic effects. In this respect what he gives us is, ironically, reminiscent of the catalogue of intonational forms and meanings described by O'Connor & Arnold (Reference O'Connor and Arnold1973), perhaps the best-known version of the twentieth-century British traditional analysis of English intonation. Ward's catalogue is considerably more ambitious: he emphasizes correctly that work like O'Connor & Arnold's (and more recent work like Pierrehumbert & Hirschberg Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990, Ladd Reference Ladd2008, Gussenhoven Reference Gussenhoven2004, or Steedman Reference Steedman2014) focuses only on intonation, whereas he aims to account for all the non-segmental features in the speech signal. But his descriptions of the nuances of specific constructions are often just as hard to pin down (and thus to falsify) as O'Connor & Arnold's. For example:

(Compare O'Connor & Arnold Reference O'Connor and Arnold1973: 54: ‘The High Drop is probably the most common way of asking [WH] questions. It avoids the seriousness and urgency of the Low Drop, and such questions sound brisk, business-like, considerate, not unfriendly.’) This problem is pervasive in any study of intonational meaning. As I wrote in Intonational Phonology, descriptions of this sort can ‘be evaluated as reasonably plausible – or reasonably implausible, depending on who is doing the evaluating. However, there is no theoretical framework within which we can undertake a comparative evaluation that would command general agreement’ (Ladd Reference Ladd2008: 150). Insisting that he is dealing with ‘facts’ and ‘evidence’ does not exempt Ward from confronting this problem.

Of course, his general research programme does suggest an obvious riposte to this criticism – namely that, by actually building a computational system that mimics conversational behaviour, we can assess the extent to which we have understood how prosody conveys conversational meanings. This response clearly has some force. But a working model does not guarantee genuine understanding of the thing modelled. We can now build natural language translation systems based on machine learning that provide serviceable translations in many cases but sporadically produce gibberish. I have no doubt that Ward's approach to identifying conversationally meaningful properties in the speech signal could contribute significantly to the operation of a complete dialogue system, but he provides no evidence that his own set of conversationally meaningful properties would make more of a contribution than someone else's set. Veilleux's blurb expresses the right amount of scepticism when she describes Ward's approach as ‘paving the way for the discipline to use his methods experimentally’ – in other words: nice work, now let's try to validate it.

To be fair, Ward is honest about what his classification does and does not do. A clear illustration involves what he calls ‘late peak’ (chapter 6). He lists no fewer than 22 functions (along the lines of ‘suggesting’, ‘requesting’, ‘inviting’, all arguably related to a root meaning ‘not action-oriented’ (p. 92)) and he evaluates this analysis by having a ‘patient research assistant’ identify the pragmatic effect of each occurrence of late peak in a corpus of 300 cases from recorded conversations. The ‘overwhelming majority’ of the 300 cases conveyed one of the 22 functions on his list. But the ‘overwhelming majority’ turns out to be 220 cases, or not quite three-quarters of the full set. This means that a computer dialogue system equipped with the tools Ward provides would potentially be involved in a conversational misunderstanding at every fourth occurrence of a late peak. It is hard to see this result as some kind of breakthrough.

***

I omit any further discussion of Ward's PCA-based discovery procedure or its potential applications and return now to take up a few points on which his treatment of past work in linguistics is seriously inadequate. One conspicuous shortcoming is the absence of any real discussion of what ‘prosody’ is. Essentially all he says is this:

Like many others going back to antiquity (see Ladd Reference Ladd2014: chapter 3 for discussion), Ward implicitly bases his working definition on alphabetic writing: prosody is everything in speech that is not written down with consonant and vowel letters. Most obviously, this implicit definition means that he treats lexical tone as prosody (e.g. p. 22). More remarkably, he explicitly states that a symbolic representation works fine for consonants and vowels (‘phonemes’), but for ‘prosody’ (i.e. everything else) we require a more elaborate multi-dimensional representation. His justification for this view includes supposed differences between the supralaryngeal articulators and the larynx:

He does not explain how, if the larynx is so sluggish, centisecond-level accuracy of voice onset time is possible in languages all over the world. He never considers the possibility that the motivation for symbolic representations of pitch contours (from tone diacritics in Vietnamese to British ‘tonetic stress marks’ and autosegmental descriptions of pitch accents) might arise from theoretical considerations that are not based on the accidents of alphabetic writing.

More generally, as I noted at the outset, the balance Ward strikes between readability and scholarship is decidedly biased against the latter, and some of his statements about the literature must really be seen as elementary errors. His most astonishing claim of this sort is that, while most past work on intonation has involved rehashing the same old oversimplified stories (like ‘questions end with a rise in pitch’), the phenomenon that he calls late peak ‘is a recent discovery’ that was ‘first described in 1977’ (p. 77). The ‘discovery’ he refers to is Bruce's (Reference Bruce1977) demonstration that the best way to analyse the Swedish word accent distinction is to take pitch movements such as rise–fall as primitives, identified independently of syllable boundaries, and to describe the distinction between Swedish Accent I and Accent II in terms of how the rise–fall pitch movement is aligned with the accented syllable. (There are other differences as well that are irrelevant here.) This analysis – and the corresponding conceptual distinction between ‘association’ and ‘alignment’ – has been widely accepted since then, but Bruce's innovation lies in his approach to describing the phonetics of pitch, not in the discovery of a new phenomenon in English or any other language. For example, Sledd (Reference Sledd1956) invoked late peaks as a technical problem for the 1950s American analysis of intonation; one of Sledd's examples was ²won³derful¹ (i.e. with the highest pitch reached in the syllable -der-). Before Sledd, many authors in the twentieth-century British tradition of intonation analysis discussed such cases; Sledd credits the wonderful example to Palmer (Reference Palmer1922). Ward does not discuss the British analyses, but he does cite Ladd (Reference Ladd1983), the paper in which I point out the connection between the British analyses and what I called ‘delayed peak’; he explicitly equates his ‘late peak’ with my ‘delayed peak’. Late peak was not discovered in 1977.

Finally, let me comment on the endnotes that Ward uses in place of in-text references. These are inadequate in two ways: first, there are many statements that cry out for attribution but have no endnote; second, the connection between works cited in the endnotes and the statements to which they are attached is sometimes obscure or downright misleading. Examples of the first sort – conspicuously unattributed claims – include the following: ‘Prosody is processed using specific brain circuits (mostly in the right hemisphere, in contrast to words, which are processed on the left)’ (p. 4); ‘[I]t is known that pitch movements, up or down, are not something that most people can reliably detect in isolation’ (p. 122). Examples of the second type – references to works that have little obvious bearing on the statement to which the endnote is attached – are plentiful (e.g. note 34 in chapter 2 [ = endnote 36 on p. 200], note 24 in chapter 5, note 12 in chapter 14). In part this is because the endnotes almost never include page numbers, even for whole books, which makes it hard to evaluate the relevance of the reference. Sometimes, however, it is because the cited works are simply misinterpreted. To give a clear example, Ward (note 16 in chapter 3, endnote on p. 200), citing a personal communication from me to Gussenhoven (Reference Gussenhoven2004: 314), supports his own analysis of the ‘Minor Third Construction’ (often known as the calling contour) by stating that I observe that this contour requires a following silence. In fact, the point at issue in Gussenhoven's discussion is the length and structural complexity of the segmental material that follows the pitch drop in the calling contour; it has nothing to do with silence at all.

In short, this book cannot be recommended as a reliable guide to past work, but it provides an easy and lucid introduction to Ward's overall project, and illustrates one way that ‘big data’ methods might be used to investigate the prosody of spontaneous dialogue. Fundamentally, though, all it presents is one more possible approach to a difficult problem. It will still be some time before we attain a deep and practically applicable understanding of everything that contributes to making a recording of a conversation so much more than a transcript.