The Relevance of Context and Experience for the Operation of Historical Sound Change

This paper is concerned with explaining how historical sound change can emerge as a consequence of the association between continuous, dynamic speech signals and phonological categories. The relevance of this research to developing socially believable speech processing machines is that sound change is both cognitive and social and also because it provides a unique insight into how the categories of speech and language and dynamic speech signals are inter-connected. A challenge is to understand how unstable conditions that can lead to sound change are connected with the more typical stable conditions in which sound change is minimal. In many phonetic models of sound change, stability and instability come about because listeners typically parse—very occasionally misparse—overlapping articulatory movements in a way that is consistent with their production. Experience-based models give greater emphasis to how interacting individuals can bring about sound change at the population level. Stability in these models is achieved through reinforcing in speech production the centroid of a probability distribution of perceived episodes that give rise to a phonological category; instability and change can be brought about under various conditions that cause different category distributions to shift incrementally and to come into contact with each other. Beyond these issues, the natural tendency to imitation in speech communication may further incrementally contribute to sound change both over adults’ lifespan and in the blending of sounds that can arise through dialect contact. The general conclusion is that the instabilities that give rise to sound change are an inevitable consequence of the same mechanisms J. Harrington (B) · F. Kleber · U. Reubold · M. Stevens Institute of Phonetics and Speech Processing, Ludwig-Maximilians-University of Munich, Munich, Germany e-mail: [email protected] F. Kleber e-mail: [email protected] U. Reubold e-mail: [email protected] M. Stevens e-mail: [email protected] © Springer International Publishing Switzerland 2016 A. Esposito and L.C. Jain (eds.), Toward Robotic Socially Believable Behaving Systems Volume II, Intelligent Systems Reference Library 106, DOI 10.1007/978-3-319-31053-4_6 61 [email protected] 62 J. Harrington et al. that are deployed in maintaining the stability between phonological categories and their association with the speech signal. A fundamental challenge in phonetics and the speech sciences is to understand how the interleaved movements of continuous speech signals are associated with categories such as consonants and vowels that function to distinguish between word meanings. The dichotomy between these two levels of representation comes about because on the one hand any speech utterance is highly context-dependent but on the other hand languages distinguish words by means of a finite cipher of abstract phonological units that can be permuted in different ways. There is abundant evidence for the context-dependent nature of speech. The same utterance can vary dramatically depending on the speaking situation and environment—whether talking to friends or in a more formal speaking situation, whether there is background noise or quiet [92]. Speech is also context-dependent because speech sounds are synchronised in a temporally overlapping way: producing speech is a shingled movement [144], so that any particular time slice of the speech signal provides the listener with information about speech sounds that have been, and that are about to be produced and in a way that is also different depending on prosodic factors to do with syllable position and the stress or prominence with which syllables are produced [9]. These are then many of the reasons why temporally reversing a speech signal of stack does not lead to an unambiguous percept of cats. Moreover, the context-dependence is not just a function of the sluggishness of the vocal organs in relation to the speed with which speech is produced, but also communicates much about the social and regional affiliations of the speaker [60, 96]. At the same time, phonological abstraction from these details is fundamental to human speech processing: there is a sense in which the different words stack, cats, acts, and scat are permutations of the same four phonological units or phonemes. There is now extensive evidence that children learn both levels of representation in speech communication: they are on the one hand responsive to acoustic information arising from continuous movement in the speech signal. But simultaneously they acquire the ability to perform phonological abstraction which allows them to recombine abstract phonological units to produce words that they have not yet encountered (e.g. Beckman et al. [10]). The task in this paper is to make use of the existing knowledge about the connections between these two very different ways of representing speech in order to explain the operation of sound change; and in turn to use what is known about sound change to place constraints on the type of architecture that is possible and required for linking these physical and abstract levels of speech sound representation. The focus will be on what the Neogrammarians of the 19th century [115, 118] termed regular sound change which they considered to be gradual and imperceptible and to apply to all words; this type of sound change was for them distinct from analogical change which was irregular, phonetically abrupt (in the sense that the change was immediate, not gradual), lacked phonetic motivation, and often applied to only a handful of words (see e.g. Hualde [68] for some examples of change by analogy). Modelling sound change is relevant to understanding how cognitive and social aspects of human speech processing are connected. The association between these two domains has been largely neglected in the 20th century, partly because whereas [email protected] 6 The Relevance of Context and Experience ... 63 generative theories of phonology draw upon highly idealised data (typically phonetic transcriptions) to develop a grammar consisting of re-write rules based on supposed linguistic universals and operating on a lexicon containing the minimum of information to represent the distinctions between words, sociolinguistic models by definition are concerned with explaining how variation in speech is conditioned by factors such as gender, age, and social class, factors that are beyond the scope of generative models. In the last 10–15 years, there have, however, been increasing attempts to reconcile these two positions largely within so-called usage-based, episodic or exemplar models of speech (see e.g. Docherty and Foulkes [31] for a recent review) that are derived from models of perception in cognitive psychology (e.g. Hintzman [65]) and that give much greater emphasis to the role of memory in human speech processing. Exemplar theory has led to computational models of how phonological categories, the lexicon, memory and speech are inter-connected (e.g. Wedel [157]); and more generally, there has been greater emphasis in the last two decades in determining how speaker-specific attributes shape and influence cognitive aspects of human speech processing both in adults [122] and in first language acquisition (e.g. Beckman et al. [10], Munson et al. [103]). Understanding how social and cognitive aspects are related in human speech processing is in turn a pre-requisite for developing socially believable systems of machine speech and language processing. The relevance of sound change in this regard is that it is evidently both cognitive and social. The cognitive aspects are largely concerned with the mechanisms by which phonological categories and speech signals are associated and how this association can sometimes become unstable providing the conditions for sound change to take place. The social aspects are more concerned with how differences between speakers in their knowledge and use of language can cause sound change to spread throughout the community. These cognitive and social bases of sound change have in the last 40–50 years been pursued within largely separate frameworks concerned on the one hand with the conditions that can give rise to sound change (in particular Ohala [112]) and those that lead to its spread across different speakers on the other (in particular Labov [85, 87]). The challenge lies in developing an integrated model of sound change that draws upon insights from both approaches. This is turn can provide fresh insights into understanding how social and cognitive aspects must be inter-connected in both human and (therefore also) machine speech processing. 6.1 The Phonetic Basis of Sound Change Much consideration has been given to the question of whether there are factors intrinsic to the structure of the language that can bring about sound change. Such questions are typically focussed on whether there are sounds and in particular sequences of sounds that are inherently unstable, either for reasons to do with speech production or because they tend to be ineffectively communicated to the listener. Such biases in either the production or the perception of speech that predispose sounds to change should also be reflected in the typological distribution of sounds and sound sequences [email protected] 64 J. Harrington et al. in the languages of the world. Thus syllables beginning with /kn, gn/ as in German Knabe, Gnade are rarer in the languages of the world than those beginning with /kl, gl/ [67]; and they are also involved in sound changes by which /k, g/ are deleted as in the evolution of English words that have fossilized the earlier (16th century) pronunciations in the orthography (knave, knight, knife, gnome, gnat etc.) but which are now in all English varieties pronounced without the initial velar stop [97]. Research concerned with the structural conditions that give rise to sound change is founded on two further ideas. Firstly, the biases that bring about sound change are directional [113]. Thus there are many sound changes by which the front vowel /y/ has developed historically from a back vowel /o/ or /u/ (e.g. the modern German Füße, with /fys/ in the first syllable historically from Proto-Germanic /fotiz/; hence also the English alternation feet, foot) but far fewer sound changes by which /y/ has retracted with the passage of time to /u/ or /o/: that is, there is a bias towards high back vowel fronting (as opposed to high front vowel retraction) both synchronically and diachronically [55]. Similarly, Guion [48] has shown that the misidentification of /ki/ as /t i/ which forms the basis of sound changes known as velar palatalization in numerous languages (e.g. English chin, but German Kinn) is far more likely than the perhaps unattested sound change by which /t i/ evolves into /ki/. Secondly, to the extent that speakers from different languages and cultures are endowed with the same physical mechanisms for producing and perceiving speech, there should be broadly similar patterns of sound change (sometimes referred to as regular sound change) in unrelated languages. The model of sound change developed by Ohala [111–113] over a number of decades is founded upon such principles. The basis for this model is that coarticulation in speech production—that is the way in which speech sounds overlap and influence each other in time—is in almost all cases accurately transmitted between speakers and hearers. An example of coarticulation in speech production is given in Fig. 6.1 which shows the distribution of the back vowel /!/ (e.g. musste, ‘had to’) in nonfronting /p/ and fronting /t/ contexts in German. In the fronting context, the tongue dorsum for /!/ is further forward in the mouth due to the influence of /t/ that has its primary constriction further forward than that of /!/ resulting acoustically in a raised second formant (F2) frequency. How does the listener deal with the type of variation shown in Fig. 6.1? Experiments over several decades [41, 93, 99] show that adult listeners of the language interpret speech production in relation to the context in which it was produced. For the present example, this implies that listeners carry out a type of perceptual transformation such that they attribute the coarticulatory fronting not to the vowel itself but to the consonantal context in which it was produced. This can be demonstrated in perception experiments by synthesising a continuum between a front and back vowel in equal steps and embedding the continuum in coarticulatory fronting and non-fronting contexts. Figure 6.2 shows the results from just such an experiment for the fronting context /jist-just/, yeast-used (past tense) and for the non-fronting context /swip-swup/, sweep-swoop. (The former is a fronting context because of the /j/ and the latter a non-fronting context because the tongue dorsum for /w/ is retracted, [email protected] 6 The Relevance of Context and Experience ... 65 Proportional (normalised) time T on gu edo rs um ( X )