themes2

advertisement
Experimental techniques
Beckman: apparent time analysis across various studies
Beddor: ultrasound
Beckman & paidologos project
The corpus is a database of speech productions recorded in the same way from
native speakers for the target language variety in day care centers, nursery schools, or
participant homes in a half dozen different locations for as many different languages. The
portions of the corpus that are used in the current paper are the Mandarin Chinese recordings
made in 2006 in Sōngyuán City, PRC, by the second author, and the Korean recordings made
in 2007 in Seoul, South Korea, by the third author.
Conditions that give rise to sound change: intrinsic to the signal
Many papers are concerned with the conditions under which phonologisation takes place.
Kirby: tonogenesis
in Phnom Penh Khmer
colloquial speech of the capital Phnom Penh /kru/ produced with post-aspiration or breathy
voice or falling pitch
... the loss of the trill has been compensated for by the enhancement or emergence of other
acoustic cues
Cues
Beckman:
a shift in the phonetic basis of the Korean contrast between lax and aspirated stops from being
primarily a contrast in degree of aspiration in phrase-initial position to becoming primarily a
contrast in the tone pattern. the size of the difference in VOT values reported in subsequent
studies has diminished decade by decade, and both Silva (2006) and Kang and Guion (2008)
report apparent-time evidence (i.e., systematic variation across talkers of different ages
recorded in the same year) that strongly suggests a change in the contrast from being
primarily a difference in degree of aspiration to becoming a difference in voice quality and
pitch level on the vowel of the initial syllable of the Accentual Phrase.
a gradual shift in the phonetic realization of Korean stops in phrase-initial position whereby
the originally moderately long VOT values for the “mildly aspirated” lax stops have
lengthened a great deal and the originally very long VOT values in the aspirated stops have
shortened so as to considerably reduce the VOT difference. but it is not resulting in a near
merger. Instead, other correlates are taking over the role of degree of aspiration in cuing the
contrast.
In general, Kirby: The implementation
of the phonetic precursor discussed here is closely related
to the proposal of Beddor (2009), who shows how a gesture initially associated with one
segment can come to be interpreted distinctively with a different segment. The
implementation of the phonetic precursor discussed here is closely related to the proposal of
Beddor (2009), who shows how a gesture initially associated with one segment can come to
be interpreted distinctively with a different segment.
Beddor’s study focuses on the emergence of contrastive nasalization, whereby the acoustic
effects of velum lowering initially associated with a nasal come to be associated with a
following vowel, but it is conceptually very similar to the temporal exchange between
duration of /r/ and onset VOT hypothesized here.
Hypoarticulation
Beddor: its starting point and asks whether reduction is especially likely to occur in highfrequency words
The theoretical motivation for this extension was to investigate the phonetic underpinnings of
sound changes leading to post-vocalic /l/ vocalization or loss. We predicted that a post-lateral
consonantal context that does not require an alveolar constriction might contribute to alveolar
reduction. Moreover, we predicted that the highly practiced articulatory routines (and,
possibly, faster lexical access) for high-frequency words with final /lClabial/ or /lCvelar/
would further contribute to alveolar reduction. Both predictions were upheld.
New to this study is the finding that, the more frequent the word ending in / clusters, the
greater the aperture in the alveolar region for lateral productions (Figure 5). This lexical
frequency effect, while significant, was small.
Nonetheless, the frequency effect is not simply one of overall smaller magnitude of lingual constrictions in highfrequency words, as shown by the finding that, for a given high-frequency word (in relation to its low-frequency
counterpart), tongue tip and tongue dorsum apertures were not correlated. For these data, then, reduction of the
tongue tip constriction is not tied to reduction of the tongue dorsum constriction.
Non-linear relationships between acoustics and production
Beddor: For reduction of the lateral alveolar constriction in labial and velar contexts to serve as the
impetus for sound change, it follows that the reduced gesture must have acoustic consequences...The
crucial point here is that, in certain vowel contexts, small changes in F2-F1 proximity in /Vl/
sequences due to reduced tongue tip constriction may result in disproportionately large perceptual
shifts.
Solé:
outlines three types of ambiguity; and it is the third kind
Role of statistical frequency/hypoarticulation in sound change
Beddor: practiced, efficient articulatory routines for high-frequency words (Bybee 2002)
it is reasonable to speculate that, should pronunciation norms begin to shift and alveolar reduction
for /l/ become exaggerated, high-frequency words would likely be at the forefront of that shift.
Conditions that give rise to sound change: speakers
children
Beckman: Indeed, already by the late 1970s, Vihman (1980) could conclude that any account
of the role of children in phonological change that is predicated on positing such axiomatic
homologies between children’s characteristic errors and common sound changes is untenable.
(See Foulkes and Vihman 2014 for an updated presentation of more data from Vihman’s own
work that supports her original arguments, and Edwards, Beckman, and Munson 2012 for a
review of some of the large body of subsequent work that also argues against invoking
substantive universals to explain phonological development.)
n predicting that the distribution of phonetic values in young children’s productions will
match the somewhat more conservative patterns of young men [in Neogrammarian change]
rather than the more advanced patterns of young women, and that listeners will expect this
we predict that the developmental pattern for children in Songyuan will differ. Specifically,
because of the break in transmission, the distribution of phonetic values in Songyuan children
will match that of the more advanced talkers in the more abrupt change from above there.
On Songyan: how the transmission of this abrupt “change from above” is reflected in the
phonological development of children whose grandparents are in the generation before the
change (same age as the older speakers in Li 2005) but whose parents are in the generation
after the change (same age as the younger speakers in Li 2005).
hus, the evidence from the corpus suggests that the system being transmitted to
the Songyuan children is the new imported standard system with the robust contrast between
retroflex and dental fricatives, and not the older Songyuan system with only the single apical
sound in contrast with the alveolopalatal or some “mixed” system that is intermediate
between the two-sibilant system and the robust three-sibilant system.
The Songyuan children show a developmental progression toward accurate retroflex and
dental fricatives that is similar to the progression documented for children in Beijing
women
Babel:
males did not imitate to the extent that females did, although differences between men and
women were moderated by voice type so both accommodated to the same degree to the least
typical voices; Females accommodated to the most attractive voices more than to the less
attractive voices, providing support for the social preferences hypothesis.
Beckman: a female-led sound change that was well underway by the 1970s, and is now close to
completion in the Seoul dialect.
those who have exposure to multiple dialects
Clopper cognitive representation of dialect;
The preceding discussion of processing benefits associated with local and standard
varieties suggests that listeners who have substantial exposure to multiple different dialects as a
result of self-reported geographic mobility and/or exposure to a non-local standard through
education or the media, will exhibit a processing benefit for all of the varieties with which they are
familiar
the processing benefits associated with the standard may contribute to dialect leveling and
convergence towards the standard
The cross-dialect processing benefits observed for participants with exposure to multiple varieties
reflect this greater flexibility in perceptual mapping, because this flexibility provides access to both
local and standard forms in speech processing tasks.
(and inherent ambiguities between speakers because they process in different ways - bring in
research by Jannedy here).
even brief exposure to previously unfamiliar talkers within an experimental setting can
provide listeners with sufficient familiarity with those talkers to facilitate processing.
In particular, individuals with exposure to relatively little sociolinguistic variation should
have relatively less variable distributions of phonological and/or lexical exemplars, whereas
individuals with exposure to more sociolinguistic variation should have more variable
distributions.
group affiliation
Clopper: the
locus of sound change in individuals Milroy and Milroy (1993) argued that
individuals with weak social networks are more likely to be linguistic innovators than
individuals with strong networks, because individuals with weak social group connections are
more susceptible to outside influences than individuals in close-knit groups.
Jannedy:
Many recent experimental results on the categorization of speech in different language and cultural
backgrounds suggest that listeners do not categorize speech in a context-free way but by attributing
speaker characteristics which may be manipulated through explicit and even very subtle implicit
information on the origin of the stimulus (Strand & Johnson 1996; Niedzielski 1999; Hay et al. 2006;
Hay & Drager 2010; Brunelle & Jannedy 2007, 2013).
young speakers from multiethnic neighborhoods are using this alternation as a social marker
implicitly coded social information has an effect on the categorization of acoustic stimuli The third
contributing factor may be the youth-style Hood German that can be frequently heard in public
spaces like the public transportation, shopping malls or on the streets. It has become a widely
distributed youth speech style throughout Berlin, although it was formerly strongly associated with
Kreuzberg
You process speakers differently depending on who you are listening to
Beckman: we predict that in perception experiments, listeners will attend differently to the
VOT and the
F0 pattern depending on the age of the talker and also (if the talker is an adult) whether the
talker is a man or a woman.
Beckman: listeners did attend to the VOT when identifying men’s or children’s stops as lax
versus aspirated in a way that they did not when listening to women.
Jannedy: Thus, the aim of this study is to evaluate the effect of implicitly coded social information
(relating to neighborhood in our case) on the categorization of acoustic stimuli with synthesized
fricatives ranging from /ç/ to /ʃ /
how sound change spreads
imitation (i.e. link between conditions that give rise to sound change and its spread)
related to e.g. recent studies by Pardo in
a study on male roommates, Pardo and colleagues found
a correlation between phonetic convergence and roommates’ self-reported closeness (Pardo,
Gibbons, Suppes, & Krauss, 2012) and on earlier studies showing that imitation is strongly
moderated by social preference in adults
Jannedy: The resistance may be due to lack of contact with the primary speakers
cognitive novelty drives an imitative response in phonetic accommodation.
listeners-turned-speakers modify their speech in response to the characteristics of the voices
they are hearing The Least Typical Male voice elicited the most imitative behavior (Figure 9);
this was true for shadowers of both genders. Male shadowers only accommodated to the least
typical voices. it seems as though the behavior of the male shadowers provides support for the
hypothesis that cognitive novelty drives an imitative response in phonetic accommodation.
Regular sound change vs. borrowing
beckman: regular sound change, which is assumed to be incremental and so should show
continuity between phonological development and the age-related variation observed in the
speech community undergoing the change. The second type is dialect borrowing, which could
show an abrupt discontinuity between developmental patterns before and after the sociohistorical circumstances that instigate it. the Seoul Korean contrast between lax and aspirated
stops which is undergoing regular sound change, and the standard Mandarin contrast
between retroflex and dental sibilants which has been borrowed recently into the Songyuan
dialect.
lexical frequency and spread
Word frequency effects on leniting sound changes are characteristically viewed as
contributing to the propagation of these changes through the lexicon. Bybee (2000, 2002),
Phillips (1984, 2001), and Pierrehumbert (2001), for example, all cite the tendency of
processes of lenition (assimilations, reductions, deletions) to affect more frequent words
before less frequent words. We apply a similar lexical diffusion interpretation to the finding
that, in the lateral productions of the speakers in this study, high-frequency words exhibited
weaker alveolar constrictions than did
phonological change by borrowing always begins with a replacement of one word form by another ;
regular sound change by incrementation
regular sound change, which is assumed to be incremental and so should show continuity
between phonological development and the age-related variation observed in the speech
community undergoing the change. The second type is dialect borrowing, which could show
an abrupt discontinuity between developmental patterns before and after the socio-historical
circumstances that instigate it.
change through borrowing (“change from above”) is typically ascribed to a specific historical
event that brings adult speakers of two divergent systems into contact.
In one community (Seoul, South Korea), there is an on-going shift in the pronunciation of
two series of voiceless stops that began more than fifty years ago. This shift is a “change from
below”
In the other community (Sōngyuán City, PRC), there is a change to the pronunciation of
sibilant consonants that has split the apical series in two, resulting in a contrast between
retroflex and dental word classes, as in the Běijīng dialect that is the basis for the standard
language. This split is clearly a “change from above” – i.e., a pattern of pronunciations from a
different dialect of Mandarin Chinese that was borrowed into the Songyuan variety within the
last thirty years.
computational models of spread - of how bias can spread across populations
Given a reasonably accurate characterization of a synchronic start state and a hypothesized
phonetic bias factor, this type of model enables predictions about directionality in
phonologization from a point in time when the ‘seeds’ of the change still lie below the
threshold of perceptibility.
The model sketched above employs a relatively simple characterization of the relationship
between teacher and learner, whereby the learner receives independent and identical
distributed random samples from the teacher and uses these to compute maximum likelihood
estimates by making strong assumptions about the form of the distributions.
The basic premise of the model of category acquisition outlined here is that learners attempt
to parse this potentially noisy speech input in an optimal fashion. Here, I extend that
framework in order to consider an additional source of potential discontinuity between
teacher and learner: changes in attention to cue
This suggests that changes in attention to the number of cue dimensions relevant to a contrast
could potentially condition abrupt changes in classification behavior across generations of
learners. JMH: So sound change is a combination of optimising the strategy for categorisation and
introducing a bias factor
In the simulations described above, it was assumed that the devoiced/fricated proportion ρ of
/r/ is misparsed by learners as increased VOT. That is, the assumption is that learners fail to
compensate for what is in some sense an effect of coarticulation. In this set of simulations, the
duration of 20% of all tokens of /r/ were reduced by a random proportion ρ as in Series II, but
now this was not added to the VOT of those same training examples.
JMH: the technique is to develop a model that unifies both stability and change
sigmoidal change across generations
Kirby, Beckman
Beckman: The model also predicts that the generation-by-generation incrementation will have
a logistic shape as the sound change progresses from its beginning to its end.
Kirby: Several
recent computational treatments of language change argue that the sigmoidal or
discontinuous nature of change can only be modeled by characterizing the behavior of
populations, noting that models where individuals receive input from a single teacher (e.g., S.
Kirby et. al 2007) have a linear dynamics that converge to a single stable state from all initial
conditions.
Models of sound change
Beyond Ohala
Kirby: This prediction runs counter to the well-established sigmoidal progression often observed in
sound changes involving consonantal contrasts, whereby changes between generations are often
discontinuous and abrupt . This prediction runs counter to the well-established sigmoidal
progression often observed in sound changes involving consonantal contrasts, whereby
changes between generations are often discontinuous and abrupt
Exemplar models
Jannedy: Pierrehumbert’s (2001) results on modeling category neutralization. In her computational
exemplar-based model of mergers, Pierrehumbert (2001) shows that marked and less frequent
categories such as /ç/ are subject of a persistent bias towards a more frequent and unmarked category
such as /ʃ /.
Kirby: parametric vs non-parametric see Vallabha et
al. (2007)
As the notion of category is ill-defined in exemplar-based approaches – the categories are, in
a very real sense, defined by the experienced tokens themselves – it is not clear how
variability in the signal should become information for the listener in such a model.
Kirby: differences with exemplar models: supervised vs. unsupervised learning
As the notion of category is ill-defined in exemplar-based approaches – the categories are, in
a very real sense, defined by the experienced tokens themselves – it is not clear how
variability in the signal should become information for the listener in such a model.
Another approach to phonological category learning and change from a non-exemplar
perspec- tive is Boersma and Hamann (2008), who consider the evolution of phonological
categories in a constraint-based framework.
In order to avoid computing auditory distances, their model dis- cretizes the continuous
auditory cue space into an arbitrary number of cue constraints, which are then ranked using a
standard constraint-demotion algorithm. However, this approach requires that auditory
constraints be ranked relative to a universal ‘ease of articulation’ curve, a concept that has
proven particularly difficult to rigorously quantify (Pouplier 2003). In addition, since the
authors’ primary focus was on accounting for auditory dispersion, they do not explore the
question of how multiple phonetic cues would interact in their model.
In spite of the difference between these models, superficial and otherwise, it is important to
point out that the basic predictions of all the models mentioned above are consistent with the
idea that changes to phonological systems may be emergent, rather than explicitly goaloriented.
Clopper: more familiar
exemplars are more strongly represented in the exemplar space than
less frequently encountered exemplars and are therefore easier to process. In addition,
recently encountered exemplars are more strongly represented in the exemplar space than
exemplars with a longer time lag, which can account for the rapid adaptation to variation that
is observed in the laboratory.
aken together, these findings are consistent with a model in which exposure to multiple
different varieties leads to more variable representations. These variable representations are
defined by distributions with greater variance or bandwidth than the distributions of less
variable input, which provides a larger space of potential mappings between the signal and
linguistic categories. The cross-dialect processing benefits observed for participants with
exposure to multiple varieties reflect this greater flexibility in perceptual mapping, because
this flexibility provides access to both local and standard forms in speech processing tasks.
That is, a given stimulus can be mapped onto more than one linguistic representation, which
increases the likelihood that the stimulus will be accurately identified. However, this one-tomany mapping can also lead to processing costs due to competition between possible
linguistic categories.
In an exemplar model, the competing linguistic representations would be realized with greater
variance and potentially bimodal distributions of exemplars.
Crucially, an exemplar-based perspective allows us to account for both maintenance of
sociolinguistic variation through the entrenchment of local forms and for dialect leveling
through change towards the standard. In particular, maintenance can be achieved when the
exemplars associated with the local variety are more frequent and/or more heavily weighted
in processing than the standard variety.
Other modelling points
Kirby: A further assumption made in the present work is that the learner knows the number of
category labels in advance. This is a common assumption made in work applying mixture
models to the task of phonetic category learning (de Boer and Kuhl 2003; Lin 2005), in part
because of the difficulty involved with jointly inferring both the number of mixture
components and the parameters of those components.
In their computational work on English and Japanese vowel category learn- ing, Vallabha et
al. (2007) compared parametric and nonparametric versions of an online mixture estimation
algorithm. They found the parametric algorithm significantly outperformed the nonparametric one, which they attributed to the nonparametric estimator’s inherent lack of
constraints on the underlying category structure.
Topics for Future research
Beckman:
However, there are substantial gaps in our understanding both of the extent of cross-cultural
variability in language socialization and of how this might affect the mechanisms of
phonological change that must be addressed before we can fully understand the relationship
between the time courses of the two.
we need better models of the social forces driving both phonological development in an
individual learner and phonological change in a speech community before we can begin to
understand the relationship between the two.
Beckman: However, we need considerably more detailed ethnographic information on
language socialization patterns in early childhood in the two speech communities before we
can be confident of our interpretations of these two cases. And we also need more case
studies of developmental patterns in other communities where there are changes affecting
consonants contrasts like the ones studied here before we can extend the interpretations to be
a more confident general hypothesis about differing timelines for regular sound change versus
borrowing.
Download