A psycholinguistic perspective on the acquisition of phonology

A psycholinguistic perspective on the acquisition

of phonology

Franck Ramus, Sharon Peperkamp, Anne Christophe,

Charlotte Jacquemot, Sid Kouider and

Emmanuel Dupoux

This paper discusses the target articles by Fikkert, Vihman, and Goldrick and

Larson, which address diverse aspects of the acquisition of phonology. These

topics are examined using a wide range of tasks and experimental paradigms

across different ages. Various levels of processing and representation are thus

involved.The main point of the present paper is that such data can be coherently

interpreted only within a particular information-processing model that specifies

in sufficient detail the different levels of processing and representation. We first

present the basic architecture of a model of speech perception and production,

justifying it with psycholinguistic and neuropsychological data.We then use this

model to interpret data from the target articles relative to the acquisition of

phonology.

1. Introduction

One recurrent problemin linguistic and psycholinguistic research is that data are

often taken to reflect one or two particularly prominent levels of representation

or processing, whereas many more are potentially involved. Different datasets

sometimes appear to be contradictory because they are thought to reflect incompatible

properties of one particular level of representation. But a full analysis

of the tasks and of the levels of representation involved may reveal that such

datasets actually reflect properties of different levels of representation. Therefore

they are not necessarily in conflict, but the theoretical interpretation of the

data may need to be revised.

In this paper, we discuss what we call the “standard model” of phonological

theory. This model basically distinguishes two levels of mental representation,

an underlying and a surface level.The former is a level of representation inwhich

words are represented as a sequence of abstract units (phonemes, features . . . ),

312 Franck Ramus et al.

the latter is a more detailed level of representation in which complete utterances

are represented as a sequence of speech sounds or phones. The phonological

grammar mediates between these two levels, in that it maps underlying forms

onto surface forms.

In the times when spontaneous speech corpora were the main source of

data to be interpreted, the standard phonological model was perhaps sufficient

to account for most of the phenomena of interest1. But in the days of 21st

century Laboratory Phonology, where relevant data is sought from perception

aswell as production tasks, first and second language acquisition, and where the

analysis of phonetic details in speech production as well as of their influence on

perception has becomemuch more sophisticated, this model has clearly become

insufficient. Basically, there is nothing wrong with the standard model, but it

needs to be complemented with additional components in order to account for

the extended range of available data, as some have recognised before (Hume

and Johnson 2001; Boersma 1998, 2006; Pierrehumbert, Beckman, and Ladd

2000). Furthermore it is also crucial to take into account how the postulated

information-processing architecture might plausibly develop from the initial

stage at birth to the mature stage.

In this paper, we first present the basic architecture of a model of speech

perception and production, justifying it with psycholinguistic and neuropsychological

data.We then use this model to interpret data from the target articles

relative to the acquisition of phonology.

2. An information-processing model of speech perception

and production

2.1. Basic architecture and functioning

The model presented in Figure 1 is directly inspired from the classic logogen

model (Morton 1969) and subsequent updates, variants and refinements (Norris,

McQueen, and Cutler 2000; Morton 1980; Caramazza 1997; Coltheart 1978;

Levelt 1989), as well as fromideas coming fromthe linguistic literature (Chomsky

and Halle, 1968; Jackendoff, 1997; Prince and Smolensky, 1993).

The basic principles at work in Figure 1 are as follows: 1) boxes stand for

distinct levels of representations, 2) arrows stand for “processes” that perform

a mapping (or conversion, or translation) between different levels of representations,

and 3) not all conceivable boxes and arrows are shown, only those that

are necessary for the present discussion.

A psycholinguistic perspective on the acquisition of phonology 313

Figure 1. An information processing model of speech perception and production. The

boxes in grey represent the standard model of phonological theory.Arrow 1a

corresponds to output phonological processes, 2a to output phonetic implementation.

1b corresponds to phonological parsing (inverse phonology) and

2b to perceptual phonetic decoding. In the adult, these four processes are

finely tuned to the phonological and phonetic properties to the maternal language(

s).They may be mistuned during (first or second) language acquisition,

or in cases of brain lesion or learning disability.

Themodel is centred around themental lexicon,which is a long-termmemory

store divided into at least three parts that interface with different aspects of the

world: semantic representations, phonological representations (including their

segmental content and stress or tonal pattern) and orthographic representations.

(Lexical morphosyntactic properties are not represented here, and syntax is

more generally completely out of this picture).

Here only the adult state is represented. In the initial state, the overall architecture

may be in place, but the lexicon is empty, representations are in a

universal format untainted by any language-specific category, and similarly,

processes are not trained to perform any language-specific function.

All the way from the cochlea to the primary auditory cortex, speech sounds,

like all other sounds, are encoded in a non-specific manner: this is embodied

by our acoustic representation. At a later stage of processing, speech must be

314 Franck Ramus et al.

encoded in a speech-specific manner: this is a sublexical phonological representation.

The arrow between the sublexical phonological representation and

the phonological lexicon represents auditory word recognition.

Speech production includes the selection of the appropriate words (typically

at the semantic level), the retrieval of their phonological form (from the

phonological lexicon), their assembly into a whole phonological utterance (at

the sublexical phonological level), and the conversion of this latter level into

an articulatory representation that will trigger the motor commands producing

speech (Levelt 1989).

It is evident that the standard phonological model consists of two of the boxes

represented here (in grey in Figure 1), i.e. the lexical phonological representation

and the output sublexical phonological representation, with arrow1a going from

the former to the latter representing phonological grammar. Embedding this

standard model within the more comprehensive one highlights at least two other

characteristics of the model: 1) there is an input pathway, distinct from the

output pathway, but linked with it; 2) this input pathway is also subdivided

between lexical, sub-lexical and peripheral (acoustic) levels. Before discussing

the importance of these properties for the interpretation of laboratory phonology

and language acquisition data, let us review some empirical evidence in favour

of these various levels of representation.

2.2. Input versus output systems

There has been considerable debate in the speech processing community as

to whether one should distinguish separate input and output speech systems or

postulate a common amodal one.Of course, at themost peripheral level, auditory

representations are separate from articulatory ones. The former are dedicated

to the analysis of auditory patterns, and at that level, information consists of

a continuous representation of sounds that include speech sounds as well as

non-speech sounds. The latter is dedicated to the motor planning of articulatory

gestures and consists in the specification of muscle movements/trajectories,

which are adequate both for speech sounds and for other types of vocalizations.

The interesting question concerns the more abstract sublexical phonological

level: should there be a single amodal phonological systemor two separate ones?

The strongest evidence for separate systems comes from neuropsychology, and

in particular fromcases of conduction aphasia. In this type of syndrome, patients

have a relatively intact comprehension and speech production combined with

a severe impairement in the ability to repeat speech (Caramazza et al. 1981).

Jacquemot, Dupoux and Bachoud-L′evi (2007) explored the case of a patient

(FA) who could perceive and produce both real words and nonwords, but who

A psycholinguistic perspective on the acquisition of phonology 315

could not repeat nonwords. Such a deficit can be accounted for in the model

by positing that there are two distinct sublexical phonological representations,

one for perception and one for production. Specifically, the incapacity to repeat

non-words is evidence for an impaired link from the former to the latter (the

repetition of real words is not affected, since the input and the output systems

are also connected by the lexical phonological representation). A further assessment

of FA also provided evidence for the existence of two separate links

between the input and the output sublexical phonological representations, one in

each direction. Indeed, FA had no problems with tasks that require to internally

‘hear’ phonological output (without overt production), such as judging whether

two pictures’ names rhyme, or whether a picture’s name contains a previously

heard target syllable.The results showed that FAhad an intact conversion mechanism

from phonological output to phonological input. Overall, these results

strongly suggest that sublexical phonological representations in perception and

in production are separate and connected by two independent links.

2.3. Levels of representations of speech sounds

In the output pathway, as classically analysed in the linguistic literature, the

distinction between sublexical (surface) and lexical (underlying) representations

stems from the detailed examination of the phonological shape of words. Due

to a variety of phonological processes, a word can surface in a variety of ways

depending on the phonological context, speaking rate, dialectal style, etc.. Such

an architecture has been substantiated in the psycholinguistic literature: speech

production models typically acknowledge this distinction (see Levelt 1989), and

neuropsychological investigations have reported cases of specific impairments

at either level.For instance, Goldrick andRapp (2007) have recently reported two

cases of production deficits where the patients had problems producing words,

but had intact semantic and articulatory processes. In one patient, the errorswere

affected only by lexical factors such as lexical frequency and neighbourhood

density, suggesting a deficit at the level of lexical representations. In the other

patient, the errorswere affected by phonological factors such as syllable position,

place of articulation, and phoneme frequency, suggesting a deficit in sublexical

phonological representations.

There is less consensus concerning the exact format of the two levels of

representation.At the lexical level, whereas predictable variations are generally

assumed to be derived by the grammar rather than being encoded underlyingly,

the degree of abstractness of underlying representations is still amatter of debate

(see Steriade 1995 for a review). At the sublexical level, the amount of phonetic

detail argued to be present in surface representations varies. For some, there

316 Franck Ramus et al.

is a separate grammar of phonetic implementation that maps phonological surface

forms onto phonetic surface forms (Chomsky and Halle 1968; Prince and

Smolensky 1993; Keating 1990, here arrow 2a). For others, by contrast, there

is only one grammar and hence one level of surface representation (Flemming

2001). According to the model one adheres to, phonetic detail is thus either

present at the surface phonological level or introduced at a later stage. In this

paper we stay neutral with respect to this issue, and we simply assume that lexical

representations are abstract in the sense that they do not include complete

phonetic specifications of the word forms.

In the input pathway, the evidence for the distinction between acoustic and

sublexical phonological representations is probably less well known within the

linguistic literature. It rests on experiments demonstrating language-specific

effects in the processing of speech sounds. In perception, hearers show considerable

difficulties in discriminating and memorizing non-native contrasts.

For instance, Japanese listeners have persistent trouble discriminating between

English /r/ and /l/ even if these two phonemes are acoustically discriminable

(Goto 1971; Lively et al. 1994). Such language-specific effects are not limited

to segmental contrasts, but extend to suprasegmental regularities (see, for

instance, Dupoux et al. 1997; Dupoux et al. 1999; Dupoux et al. 2001). The

current interpretation of these effects is that experience with native categories

shapes sublexical phonological representations (Best and Strange 1992; Best

1995; Kuhl 2000) and that these representations are automatically activated

when processing speech. In theory, all speech stimuli could be differentiated at

the (non language-specific) acoustic level of representation. The fact that listeners

seem to find this difficult in many conditions suggests that, most of the

time, they irrepressibly activate sublexical phonological representation and fail

to attend to acoustic information that is not used contrastively in their native

language. Language specific effects with speech sounds are therefore a strong

reason to endorse the dissociation between acoustic and sublexical phonological

representations.

This interpretation is again bolstered by neuropsychological data fromaphasic

patients with no hearing impairment, who can be selectively impaired

in phonological processing (such as phoneme discrimination or identification

tasks), while they are not impaired in tasks involving non-speech sounds (Metz-

Lutz andDahl 1984; Caramazza,Berndt, and Basili 1983;Auerbach et al. 1982).

Moreover, in neuroimaging studies, it has been shown that phonological, but not

acoustic processing, involves specifically the activation of the left planum temporale

and supramarginal gyrus (Jacquemot et al. 2003; Dehaene-Lambertz et

al. 2005).

A psycholinguistic perspective on the acquisition of phonology 317

2.4. Consequences for theories of grammar and language acquisition

The usual notion of phonological grammar refers to the processes converting

lexical phonological forms into output sublexical phonological forms. It is to be

noted however that linguists and psycholinguists may differ on what counts as

a phonological process. Linguists sometimes include in the phonological grammar,

processes that are productive synchronically as well as processes that arose

diachronically but are no longer active. For instance, phonological variations

across morphologically related words (e.g., opaque/opacity, cf. Chomsky and

Halle 1968) may sometimes only reflect statistical regularities present in the

lexicon, rather than grammatical processes per se (Myers 1999). Such cases

may reflect grammatical processes that occurred in the brains of yesterday’s

speakers of the language, and that have left their mark on the shape of today’s

lexicon. For psycholinguists however, synchronic and diachronic processes are

very different: only the former require an active mental operation and need to

be acquired by the child as such, whereas the latter just reflect the content of the

lexicon.2

Whereas lexical regularities are sometimes unduly taken to reflect phonological

grammar, another kind of phonological grammar is most often overlooked:

the one that applies in the input pathway (arrow1b). Indeed, phonological variations

introduced by speakers are typically not noticed by listeners during online

speech perception, although the same listeners are typically able to hear the

differences when excised out of their context. This suggests that during on-line

speech perception, there is a mechanism (“inverse phonology”) that undoes at

least some of these variations in order to facilitate the recognition of lexical

forms. Empirical evidence for such a mechanism has been provided for assimilation

processes, showing that English- and French-speaking listeners do

mentally undo place and voice assimilations, respectively, when hearing words

in assimilatory contexts (Darcy, Peperkamp, and Dupoux 2007; Gaskell and

Marslen-Wilson 1996; Darcy et al. 2009). There are therefore two phonological

grammars, one in the output pathway (figured by arrow 1a from lexical

to sublexical representations), and one in the input pathway (figured by arrow

1b from sublexical to lexical input representations). The latter is unfortunately

much less described than the former (but see Eisner 2002; Boersma 1998, 2006).

Nevertheless it is an integral part of what the child has to learn.

In such a model, the input and output phonological grammars (1b and 1a)

are theoretically distinct entities.We assume that, as a first approximation, these

two grammars develop in a parallel fashion in children (through input-output

loops) and end up being undistinguishable in monolingual adults. However, in

cases of abnormal development, or in cases of second language acquisition, it

318 Franck Ramus et al.

is possible that input and output phonology diverge. For instance, it has been

observed that late bilinguals can sometimes show better production of a foreign

contrast than its perception (Sheldon and Strange 1982). Another example is

the Japanese vowel epenthesis process in the perception of foreign words, that

has arguably no counterpart in the synchronic production grammar (Dupoux et

al. 1999). One should note, in addition, that the arrows going from sublexical

output to articulation (2a) and from acoustics to sublexical input (2b) represent

processes that are not language independent; rather they involve categorization

or planning processes which are finely tuned to the phonological categories of

the native language (Kuhl et al. 1992).

To sum up, we argue that there are two phonological grammars, one in input

and one in output, and two more peripheral (acoustic and articulatory), yet

language-dependent, mapping processes. An alternative way to think about it

would be that phonological grammar is distributed in a partly redundant way

over several processing loci (perception, production, and at several levels). This

may seem disconcerting froma linguistic perspective, but in fact redundant and

overlapping processing systems are common in psychology and neuroscience.

Finally, the phrase “the acquisition of phonology“ may mislead one into

thinking that there is just one thing to be acquired by the child, namely phonology.

But from the previous discussion it becomes evident that there are different

components to be acquired. One component is the right format of representation

at each of the three levels of phonological representation (input and output

sublexical, and lexical).As we will see below, it is not entirely clear whether the

three levels are acquired more or less in parallel or whether lexical and output

levels seriously lag behind the input sublexical level. A second component to

learn is the lexicon itself. Two further components are both the input and the

output sublexical phonological grammars. Here again, whether there are two

distinct grammars to acquire, or whether one is the mirror image of the other

is not clear. Acquiring phonology is therefore a multi-faceted problem for the

child. Of course, all these components are not entirely independent from each

other. But neither do they completely follow from one another. Let us now look

more closely at the language acquisition data and assess what they imply for the

acquisition of each component.

3. Interpreting language acquisition data

The papers by Fikkert and by Vihman report data from both speech perception

and production by infants and toddlers. Their tacit assumption is that the entire

body of data can be taken to reflect the development of a single level of repreA

psycholinguistic perspective on the acquisition of phonology 319

sentation, that is, the lexical phonological representation. But we will see that it

may take more than one level of representation, and a careful consideration of

the tasks used to generate the data, to provide a full account.

3.1. Speech discrimination

Starting with Eimas et al. (1971), many studies on speech perception by infants

have converged on the idea that humans are born with certain universal auditory

categories, possibly shared with other species, and that these categories change

under exposure to the sounds of a particular language, reaching a relatively

stable, language-specific state around the first year of life (Werker and Tees

2005; Kuhl 2000). In terms of ourmodel, this represents a developmental tuning

of the format of the input sublexical representation3. Indeed, in these speech

perception tasks, infants typically discriminate between novel or pseudo-words,

i.e., they compare two phonological forms represented at the input sublexical

level. Furthermore,many experimental studies have shown that this tuning goes

beyond phonemic categories, and that it also includes statistical information

about the typical shape of words in the native language (stress and tonal pattern,

phonotactics, etc. Jusczyk, Luce, and Charles-Luce 1994; Jusczyk, Houston,

and Newsome 1999; Saffran, Aslin, and Newport 1996; Jusczyk, Cutler, and

Redanz 1993;Mattock andBurnham2006; Friederici, Friedrich, andChristophe

2007). Note that this phonological information about the lexicon is acquired

during the first year of life before any significant number of lexical entries

are acquired. Whereas this set of results suggest an early tuning of the input

sublexical representation, it does not inform us on the development of lexical

and output sublexical phonological representations.

3.2. Word learning

Infants acquire some of the cues helpful for word segmentation around age 9

months, and start to acquire their lexicon at the end of the first year of life

(Jusczyk 1997). Considerable controversy has arisen regarding the format of

these first lexical entries. While the default assumption would seem to be that

the lexicon is based on the same format as that acquired during the first year of

life in input sublexical representations, some researchers have posited either a

muchmore detailed representation (including acoustic details, e.g. Singh,White,

and Morgan 2008), while other researchers have posited a much less detailed

(underspecified) representation (e.g., Fikkert, this volume).This remains a complex

issue because of the experimental difficulty of unambiguously tapping the

lexical level of representation without being contaminated by methodological

confounds.

320 Franck Ramus et al.

Starting with Stager and Werker (1997), many studies have tried to teach

babies a novel word, and subsequently test how they react to a change in the

phonological shape of the word. In this “switch paradigm”, infants are familiarized

with the pairing of a novel word and the picture of an object. After

habituation, infants are presented with the same object, either together with the

same word (the same condition) or with a minimally different word (the switch

condition). The difference in looking time between same and switch conditions

is taken to indicate the degree of mismatch between the newly learnt lexical

entry and the minimally different word. This task therefore provides evidence

as to how much phonetic detail is encoded when storing a novel word for the

first time in the lexicon, and how much can immediately be retrieved from this

preliminary representation to be compared with a new item. A bottleneck at

either stage is likely to limit performance. Furthermore it is quite a demanding

task, in particular in terms of attentional resources, since it supposes a very fast

encoding (7 to 10 presentations) of a novel item.

The main result obtained by Stager and Werker (1997) and subsequently

replicated and extended (e.g., Pater, Stager, and Werker 2004), was somewhat

counterintuitive: at the age of 14 months, infants failed to notice a minimal

change in theword form, e.g. from‘bin’to ‘din’ (even though theywere perfectly

able to distinguish between ‘bin’ and ‘din’ in a speech discrimination task).

However, they performed well when the novel words were very different, such

as ‘lif ’ and ‘neem’.By the age of 17 months, infants performedwell in the wordlearning

task, even with minimal pairs of novel words. In their original paper,

Stager and Werker proposed that infants who just begin to learn words may

not pay attention to fine phonetic detail (even though they are able to perceive

and represent them). They suggested that this inattention may be beneficial

to the infant, in that it would free attentional resources for the task of wordlearning

itself. A more mundane variant of this interpretation, inspired by the

task analysis above, is simply that the word learning task, as implemented by

Stager andWerker, is much more complex and demanding than a discrimination

task, and that various difficulty factors (speed, attention, subtlety of phonetic

differences) cumulate, hence the decrease in infants’ performance, specifically

in the minimal change condition of the word learning task.

Another interpretation of these results would be that infants’ early lexical

representations are not fully phonetically detailed. This is the interpretation espoused

by Fikkert (this volume).On her account, the place of articulation feature

is unspecified for ‘coronal’. Furthermore, babies would initially (at 14 months)

represent only one place of articulation feature per word, that of the stressed

vowel. This proposal makes very specific and somewhat counter-intuitive predictions

as to the infants’behaviour in the task. For instance, Fikkert predicts that

A psycholinguistic perspective on the acquisition of phonology 321

in the switch paradigm, 14-month-old infants should not distinguish between

‘din’ and ‘don’. She also predicts that infants should distinguish ‘bin’ from‘bon’

only when ‘bon’ is presented first (and thus stored lexically). Nevertheless both

predictions are borne out by the results of Fikkert’s experiments.

But underspecification theory makes even more problematic predictions.

For instance, since ‘bin’ is encoded as ‘?’ as far as place of articulation is

concerned, it does not mismatch with anything else; thus, a child habituated

with ‘bin’ should not notice a change even to a very different word, say, ‘gom’.

Some earlier studies have contrasted words that are very different, i.e. ‘lif ’ and

‘neem’. In those experiments, infants do notice the difference (Werker et al.

1998; Stager and Werker 1997). However, these words differ by other features

than place of articulation, so they are not appropriate for this purpose.Therefore,

a more specific test of this prediction would be warranted.

Another odd consequence of the underspecification hypothesis is that ‘don’

(stored lexically as labial) mismatches with itself (because when represented

sublexically in the test phase, [d] is not labial). This predicts that ‘bon’ is recognized

as a good instance of ‘don’, whereas ‘don’ itself is not, as described

in Fikkert’s Table 8. Fikkert concludes that in the word learning task, infants

should look longer to the “switch” item (the one different from the habituation

item). However, it seems to us that this is not the correct prediction. Table 8

indeed predicts that infants, when habituated to ‘don’, should look longer to

the same item ‘don’ than to the switch item ‘bon’; it is only when habituated

to ‘bon‘ that they should look longer to the switch item ‘don’. Across the two

habituation conditions, the two opposite effects should average out to zero. But

this is not what is found.

If the interpretation of Stager and Werker’s results in terms of ‘attentional

load’ or ‘task difficulty’ is correct, then any simplification in the learning task

itself should improve the infant’s performance. Ballem and Plunkett (2005)

showed that, using a preferential-looking (rather than habituation) procedure,

14-month-olds are able to differentiate newly learnt words differing by one

phonetic feature (like ‘tuk’–‘puk’). Preferential looking is thought to be easier

because babies have the two alternatives in front of their eyes, rather than having

to appeal to memory. Thus, there is clearly an effect of task difficulty. When it

is reduced, it appears that 14-month-old infants’ may after all be able to make

fine phonetic distinctions between novel words. This predicts that if Fikkert reran

her experiments by using preferential looking rather than habituation, she

might obtain positive results for all contrasts, confirming the full specification

of 14-month-olds’ early lexical representations.

To summarise, it seems to us that the best explanation of Stager andWerker’s

results on word learning is to be couched in terms of the difficulty factors that

322 Franck Ramus et al.

bear on a given task at a given age, and therefore that it is not necessary to

postulate lexical underspecification. Nevertheless, Fikkert’s specific pattern of

results obtained at 14months with the habituation procedure cannot be explained

in terms of difficulty factors. Are there alternative ways to explain this pattern

of results?We can only point to potential methodological confounds.

For instance, take the asymmetry between ‘bin’–‘bon’ and ‘bon’–‘bin’ discrimination.

This asymmetry is manifested by the fact that in the ‘bon’–‘bin’

condition, infants look massively longer to switch item ‘bin’, while in the ‘bin’–

‘bon’condition they fail to look longer to switch item‘bon’than to ‘bin’ (Fikkert,

this volume). This could result from the addition of two independent effects:

perfect discrimination between ‘bin’ and ‘bon’, and overall preference for ‘bin’.

Why would infants prefer ‘bin’?We have no theory about that, but such preferences

are commonplace and maybe related, among other things, to the statistics

of the input . For instance, 9-month-old infants prefer to listen to more frequent

phonemes or phonotactic patterns (Friederici andWessels 1993). Regardless of

what may drive infants’ preferences, the bottom line is that the switch paradigm

is not suited to testing asymmetries in discrimination, because it does not factor

out intrinsic preferences. Baseline looking times for both words would be

necessary to interpret the looking times in the switch and same conditions.

Regarding the finding that ‘bin’–‘din’ is harder to discriminate than ‘bon’–

‘don’, this conclusion would first require a statistically significant interaction

between pair and condition before being accepted. Secondly, before interpreting

such a pattern of data, one would like to be sure that both pairs are equally

discriminable from an acoustic or phonetic point of view. This might be true in

general for these pairs of syllables, or might be an artefact of the material used

in these experiments. Indeed, any stimulus set that relies on a limited number

of tokens of each type, uttered by a single speaker, runs the risk that the items

might be discriminated on the basis of some phonetic details that might be

completely idiosyncratic to that particular speaker or even to those particular

recordings.Amore varied stimulus set, using numerous tokens uttered by several

speakers, would rule out this possibility, forcing discrimination at an abstract

phonological level of representation, therefore ensuring that task performance

actually reflects the intended representations. Regardless of their source, the

hypothesis that asymmetries might arise from the material could be assessed by

conducting acoustic measurements, or by testing the discriminability of these

two pairs under various levels of noise.

These remarks regarding the data collected by Fikkert do not show that her

conclusions are wrong, but that without additional data, simpler interpretations

are possible. Importantly, these interpretations aremotivated by the architecture

in Figure 1. Preference for certain phonological shapes arise from statistics

A psycholinguistic perspective on the acquisition of phonology 323

collected by infants at the level of the input sublexical representation. Acoustic

or phonetic effects on the discriminability of pairs of words or nonwords rest on

the fact that many experimental tasks can be performed at several processing

levels.A so-called ‘lexical’ task can involve sublexical or acoustic components;

an ‘input’task can involve output components, etc.This is because the processing

system is designed to activate automatically all the levels, not just the ones that

are of interest to the experimenter.

3.3. Word recognition

Beyond word learning tasks, a more direct way to address the phonetic specification

of lexical representations is to test the representations of those words

that are already familiar to babies. In such experiments, babies are typically presented

with two pictures of familiar objects (say a car and a ball), while hearing

sentences like ‘look at the X!’ where X either matches a lexical phonological

form (‘ball’), or introduces a minimal mispronunciation (‘gall’) (Swingley and

Aslin 2000). Results show that English-learning babies, as young as 14 monthold,

look faster to the target when it is correctly pronounced, than when it is

minimally mispronounced (Swingley and Aslin 2002), suggesting again that

their lexical representations do encode this degree of phonetic detail (see also

Fennell and Werker 2003). Note that these results are completely at odds with

Fikkert’s assumption that 14-month-olds do not represent the place of articulation

of consonants.

However, as Fikkert notes, the underspecification hypothesis makes additional

predictions for word recognition that are not directly tested in published

work (since the above authors did not compare the appropriate contrasts). Fikkert

(this volume) also ran word recognition experiments with 20- and 24-monthold

Dutch infants, but introduced one additional manipulation: she contrasted

cases in which the target word started with a coronal consonant, such as in

‘tand’ (tooth), supposedly underspecified for place of articulation, and cases in

which the target word started with a labial consonant, such as in ‘poes’ (cat),

supposedly specified for place of articulation. She showed that infants accepted

a change in the place of articulation of the first stop consonant, only when that

consonant was coronal, but not when it was labial. These results are clearly

consistent with the underspecification hypothesis, and can hardly be explained

by a different theory. This is particularly surprising since one would expect that

the lexical representations of 24 month olds would be even better specified than

those of 14- and 17-month olds.

Again, if one pushes the logic of the underspecification theory to its limits,

one inevitably makes strange predictions. For instance, let us take a word such

324 Franck Ramus et al.

as ‘did’4. According to Fikkert, its place of articulation is encoded as ‘?’, even

for older infants who can represent the place of articulation of all segments.As a

result, many other potential words or pseudowords (like ‘big’, ‘dig’, ‘bid’, ‘gig’)

should fail to mismatch with ‘did’. In other words, if 24-month-olds were put in

an experiment in which one of the words was a fully underspecified one (such

as ‘did’), they should always look longer to the picture for that underspecified

word upon hearing “look at the X”, no matter what place of articulation features

“X” carries.Wouldn’t that be very surprising?Wouldn’t that prevent them from

learning newwords? For instance, after having acquired theword ‘doll’, Englishlearning

infants would be unable to learn the word ‘ball’, since it would never

mismatch with the stored representation of ‘doll’: each new instance of ‘ball’

would be assimilated to ‘doll’. Yet there is good evidence that 14-month-olds

have correctly specified representations for both ‘ball’ and ‘doll’ (i.e. they look

longer to the correct picture when seeing the two pictures and hearing one of

the words) (Fennell andWerker 2003). Is it because all the infants tested learned

‘ball’ before ‘doll’?

Obviously the pattern of results found by Fikkert is very intriguing and does

not seem consistent with the hypothesis that infants’ lexical representations are

fully-specified.Yet this same hypothesis is well supported by independent data,

and the underspecification hypothesis makes a number of predictions that seem

hardly tenable. Overall the entire set of data reviewed here seems inconsistent,

which calls for a very close scrutiny of these data. Attention to all the methodological

details, as well as a good information-processing model, are needed to

try to understand exactly how, in each experiment, task structure might affect

performance, stimuli might introduce biases, etc. Ultimately, only replications

varying experimental procedures and stimuli will enable us to have a clearer

picture of the factors driving infants’ performance.

3.4. Word production

It has long been known that children’s performance in production lags behind

that in perception. In particular, children perceive contrasts in adult speech that

they neutralize in their own speech (see, for instance, Smith 1973). In order

to account for the multiple errors and hesitations in young children’s productions,

it has been proposed that children apply a set of phonological rules that

are not part of the adult grammar. These rules take surface adult forms as their

input and result in surface child forms, consisting of simpler phonological structures

(Smith 1973). In the framework of child phonology, the acquisition of the

adult phonological grammar thus largely consists of the gradual abolishment of

these simplifying rules (or, in Optimality Theory, of the demotion of the releA

psycholinguistic perspective on the acquisition of phonology 325

vant markedness constraints). This phonological acquisition would take several

years, beginning at around 12 months of age and lasting until around five or six

years. This is the grammatical account of child phonology. Fikkert (this volume)

makes a different proposal, which might be termed the representational account,

according towhich children’s productions are limited mainly by the features that

they can represent in their phonological lexicon. Alternatively, a quite different

hypothesis to account for children’s production data is that children rapidly converge

towards the adult phonological lexicon and grammar, and that deviations

from the adult targets merely reflect the development of the articulatory representation

and the link thereto (1b) from the output sublexical phonological

representation (Faber and Best 1994). Vihman (this volume) advocates this articulatory

account, while at the same time holding to the idea that early lexical

representations are different from adults’ (by being “holistic”).

The articulatory development hypothesis may account for several characteristics

of young children’s productions that are otherwise hard to explain. For

instance, acquisition is relatively slow and the changes in children’s productions

are gradual. Indeed, articulation is a very complex motor skill that requires the

fine coordination of some 150muscles in order to programand realizemore than

ten phonetic targets per second. Like any othermotor skill, articulation is learned

progressively. Another feature of children’s productions is that frequent words

are pronounced incorrectly for a longer period than infrequent words (Gierut

and Storkel 2002). For instance, children often preserve immature forms for

some words (say, French “trou” ‘hole’ as [kKu]) long after they have overcome

the articulatory difficulty, as shown by their correct production of other, less frequent,

words (“tronc” ’trunk’ as [tK?O]). This inverse frequency effect is difficult

to explain both on the representational and on the grammatical accounts, since

on the contrary, the more frequent a word is, the easier it should be to acquire its

correct representation, and the more evidence it provides for the modification

of phonological grammar.

Under the articulatory account, however, the effect can be explained by a

feature of the speech production system that was proposed independently for

adults: there are in fact two routes for establishing articulatory plans: a regular,

assembly route (that shown in Figure 1), and a route that retrieves stored

plans for frequent patterns (added in Figure 2). This was the “phonetic syllabary”

or “gestural score” in Levelt (1992). Here we propose that the gestural

score may include not only syllables, but also whole word forms, at least those

most frequently pronounced. This may indeed help explain some word-specific

idiosyncrasies in adult speech. Concerning child speech, the idea is that the

early words that are pronounced frequently are stored in the gestural score. Because

the child’s articulatory skills are immature, the words are stored in the

326 Franck Ramus et al.

simplified form that these skills initially allow.While articulatory skills remain

limited, the child will continue to utter this simplified form, thereby reinforcing

the stored plan. The child can hear his/her mispronunciation, thereby getting

negative feedback that will ultimately drive the modification of the stored plan.

But if the word was frequently uttered and therefore strongly reinforced, it may

take a lot of negative feedback, even after articulatory skills have improved, to

correct the stored pattern.

Figure 2. A modified information-processing model including the direct articulation

route to the gestural score containing the stored patterns of the most frequent

word forms.

Our account of children’s early production implies that it is very difficult

to interpret the source of a given deviation from the adult target: Is it a stored

pattern? Or does it reflect the standard phonological route? And, in the latter

case, does it reflect an immature phonological representation, or grammar, or

articulation? The multiplicity of factors affecting children’s surface forms, and

the traditional focus on grammar as the only factor of interest, means that we

actually know very little about the development of phonological grammar and

of output sublexical phonological representations. Child phonologists will have

to be methodologically creative if they are to tease apart the different factors.

Let us consider, for instance, children who neutralize a certain contrast in

production (e.g., Smith 1973).We have already argued that they must nevertheA

psycholinguistic perspective on the acquisition of phonology 327

less have adult-like input sublexical and lexical phonological representations.

How about their output sublexical representations?According to the articulatory

account, theymightwell be adult-like too.Away of testing this hypothesiswould

be to ask them to make judgments on the form of words. For instance, children

who merge [T] and [s] could be shown a picture of a mouth and one of a mouse

and asked which one rhymes with the word “house”. The words “mouth” and

“mouse” not having been produced by the experimenter, the children will have to

produce theminternally (in their output sublexical representation), and compare

them with “house”. This task therefore taps the output sublexical phonological

representations while bypassing articulation.The problemwith this kind of tasks

is that it requiresmetaphonological awareness, a capacity that develops relatively

late in children (4 to 6 years, depending on the unit to be judged, Duncan et al.

2006), perhaps too late to test any interesting aspect of child phonology. Again,

indirect methods such as priming might prove fruitful (in the present case, one

would be tempted to try picture-pseudoword interference, Schriefers, Meyer,

and Levelt 1990). However, this method requires averaging over many trials per

condition, so it will never provide information about the representation of any

particular item, but only of a set of items.

Consistent with the articulatory account, but also taking a broader view,

Vihman (this issue) asks more generally which factors explain variance in children’s

word forms, including variance between children’s forms and the adult

targets, variance among languages, and variance among children learning the

same language. Let us expand slightly on her answers. Children’s word forms

are shaped by:

1. Universal factors. In particular, universal factors constraining the development

of motor control of the articulators are the main source of differences

between children’s and adults’ words. Quite possibly, additional sources are

universal linguistic factors (markedness), to the extent that they can be shown

to be irreducible to motor constraints. The use of a unique place of articulation

over the whole word, as described by Fikkert (this volume) in the earliest

stage, is a good illustration of a putative linguistic constraint that can plausibly

result from a universal articulatory constraint (it is difficult to rapidly

change the place of articulation).

2. Language-specific factors.As reviewed earlier, the ambient language rapidly

shapes the child’s input sublexical and lexical representations. This must

be the main source of resemblance between the word productions of children

acquiring the same language: the targets are the same, and are correctly represented.

For the same reason, these representations are not likely to explain

much variance between children’s and adults’words.The development of the

328 Franck Ramus et al.

output phonological representations and grammar, and therefore the contributions

of these components, are largely unknown, given that they have not

been investigated independently of articulatory constraints.

3. Idiosyncratic factors: what specific words the child has heard, what s/he

wants to say. . . It is also plausible that, on top of universal articulatory constraints,

children may vary slightly in the respective control that they have

over their tongue, their lips, their larynx, and in sequential planning capacities,

etc., so that these child-specific articulatory constraints may drive the

predominant use of particular word forms. Such factors would therefore be

the main source of variance between children learning the same language.

Vihman’s (this issue) paper is largely focused on the last point. Her data lead

her to postulate a two-stage mechanism, including first the acquisition of a few

individual items (“holistically” represented), and subsequently the selection of

a subset of these forms to abstract out a template, that will thence be used to

adapt all word forms. We should caution that nothing in the data specifically

indicates the existence of two stages, or the systematic use of a real template.

The simple assumption that children speak under both universal and individual

articulatory constraints is sufficient to account for all the data presented by

Vihman. Therefore, although we largely agree with her about the main sources

of variance in children’s speech, we take the two-stage mechanism and the

template hypothesis to be metaphorical rather than explanatory.

Finally, the assumption that lexical representations are initially holistic is

neither warranted, nor necessary to account for production data (in fact it plays

no role in Vihman’s argument). As we have seen before, it is inconsistent with

perceptual and word learning data.And since children’s productions are mainly

constrained by limits on articulation, they say little about the format of phonological

representations. The very notion of a holistic representation may in fact

be fundamentally flawed: what exactly could it mean for a representation to be

holistic? What could be produced from a holistic lexical representation, if not

something totally slurred with no identifiable parts5? Children’s productions

may well be off the adult targets but they are anything but slurred.

4. Episodic memory

Episodic memory is of interest to laboratory phonologists insofar as it may

affect performance in some of the tasks they use. Such an excursion may in

particular be useful to interpret the data obtained by Goldrick and Larson (this

volume). Episodic memory refers to the memory of specific events and of all

A psycholinguistic perspective on the acquisition of phonology 329

the representations associated with these events. Whenever we hear a word, we

not only access the lexicon as indicated by the model in Figure 1, but we also

process the identity of the speaker who uttered it, his/her voice, his/her prosody,

the emotional state this prosody conveyed, the context in which the word was

uttered (time, place, situation), etc.All of this is encoded into the memory trace

for this particular episode.

In psycholinguistic research, interest in episodic memory rose steeply when

it appeared that episodic memories associated to a word could subsequently

affect its retrieval. In a typical recognition memory experiment, subjects first

undergo a study session where they hear (and for instance type to dictation) a

long list of words uttered by several speakers, then in a subsequent test session

(that can take place as much as a week later), they hear a list of both old and

new words, and must decide for each word if they had heard it in the first

session or not. It appears that their recognition performance in the test session

is increased for words uttered by the same voice as in the study session, than

for words uttered by a different voice (Goldinger 1996). Furthermore, there

is good evidence that it is not just the repetition of speaker identity that does

the priming, but the repetition of acoustic properties of the words, as there are

gradient effects of acoustic similarity (Goldinger 1996). Such results have led

some authors to propose that the phonological lexicon, rather than being made of

abstract phonological representations, ismade of episodic memories, containing

in particular all the acoustic details of words as they are heard (Goldinger 1998;

Johnson 1997; Pierrehumbert 2001).

However, it turns out that voice effects are task-dependent. They appear in

tasks involving explicit memory recall of these traces, but not in more implicit

tasks. For instance, in a similar paradigm as before, including a study and a

test session, voice effects disappear when the task in the test session is lexical

decision6 rather than lexical recognition (Luce and Lyons 1998). This is a

problem for episodic lexicon models, because in these models acoustic details

cannot be bypassed: they are the stuff that lexical entries aremade of. In another

study, Kouider and Dupoux (2005) investigated auditory subliminal priming,

in which subjects perform a lexical decision task on a target word, preceded

by a prime word that is masked so as not to be consciously heard. Even under

those totally implicit conditions, a repetition priming effect is obtained, that is,

reaction times for the lexical decision are shorter when the prime is the same

word as the target, than when it is entirely unrelated. Here, an episodic model

of the lexicon would predict a greater priming effect when prime and target are

uttered by the same voice than by different voices.Yet just the same amount of

priming was found, as predicted by an abstract model of the lexicon (Kouider

and Dupoux 2005).

330 Franck Ramus et al.

Figure 3. A modified information-processing model of the speech system including

episodic memory and executive processes. This graphical representation is

for illustration purposes and does not intend to make any claim about the

format and structure of episodic memory.

A plausible explanation for these contrasting results is that explicit memory

tasks like lexical recognition focus the subject’s attention on episodic memory

(“did I hear that word in the first session?”), while implicit memory tasks like

lexical decision don’t, and instead incite the subject to respond purely on the basis

of lexical status. Given that there is evidence for both abstract lexical effects,

and effects of episodic memories, it seems that the model that may best account

for the whole data set is simply a model that includes both an abstract lexicon,

and an episodic memory (Figure 3). Obviously the notion of an abstract lexicon

never was incompatible with episodic memory, which everybody knows must

exist for independent reasons. Interestingly, some proponents of the episodic

lexicon have recently made significant steps in this direction (Pisoni and Levi

2007; Goldinger 2007).

Another important component that the model in Figure 3 now shows explicitly

concerns the executive processes. By this we mean the cognitive function

whose role it is to receive input from various modules, from episodic and from

A psycholinguistic perspective on the acquisition of phonology 331

long-term memory, to control the execution of the task (in the context of a psychological

experiment), and to make decisions as to which behaviour to output7.

Executive processes are implicit in all cognitive models, but sometimes their absence

from the graphical representation leads one to forget that the behaviour in

any task is not directly driven by the internal representations that are the target

of the task. Now themodelmakes it obvious that responses in a given task can be

potentially influenced by many different representations, be they in the speech

system, in episodic memory or elsewhere. Responses can also be influenced by

task-specific strategies. More generally, task structure biases which representations

have the greatest influence on executive processes, hence on behavioural

responses. Thus, a lexical recognition task, which incites subjects to explicitly

search their episodic memory, allows the content of acoustic episodic memories

to influence responses. A lexical decision task, which is better performed by

searching the lexicon, and which would in fact be hindered by paying attention

to episodic memories, leaves the latter little influence on responses. Hence the

task-dependence of voice effects.

Beyond the debate between the episodic and the abstract lexicon, the broader

view afforded by the model in Figure 3 also provides an alternative way to

interpret tasks such as the one used byGoldrick and Larson (this volume).These

authors used a task drawn froma common class of paradigms, in which subjects

are exposed repeatedly to stimuli presenting a certain statistical pattern (in the

present case a biased distribution of some segments across syllabic positions),

and consequently manifest an implicit change in their behaviour in the direction

of that statistical pattern (here, they produce speech errors that tend to match

the biased distribution of the stimuli).

Goldrick and Larson use this paradigm to ask questions about what patterns

are learnable (or not) in phonological acquisition. Some caution is in order,

however,when using a ten-minute exposure in adults (who have already acquired

a phonological system) tomodel an acquisition process that extends over several

years in children. It is worth considering the possibility that part of the observed

results in the adult case could come from the episodic memory system, rather

than from the phonological representations per se.

This is not to deny that phonological learning is possible in adult speakers,

or to argue that the Goldrick and Larson’s results (this volume) are necessarily

episodic. But these results have to be compared to others that show that the

phonological system tends to resist late influences, even lasting for decades. For

instance, even after extensive training in changing the perception and production

of some non-native contrasts (like /r/ vs /l/ in Japanese learners of English)

performance remains non-native (Lively, Logan, and Pisoni 1993). Exactly under

what conditions and to what extent the phonological system may change

332 Franck Ramus et al.

remains to be established (e.g., Darcy, Peperkamp, and Dupoux 2007; Dupoux

et al. 2008; Sancier and Fowler 1997). In brief, our point is that when exposure

is very brief, and test follows immediately (in the case of Goldrick and Larson

training and test were simultaneous), it may be useful to consider the potential

influence of episodic memory. Experimental manipulations like introducing

changes in speakers, or increasing the lag between exposure and test can successfully

reduce such influences (Fowler, Napps, and Feldman 1985; Kouider

and Dupoux 2009).

To conclude, Goldrick and Larson are right that the simplicity of the phrase

“probability matching” is deceptive. Indeed we have argued that it is not always

clear what kind of learning probability matching reflects. However, the main

point of Goldrick and Larson is that their subjects seem to be influenced by

only some distribution biases. Indeed it is always interesting to observe that

some patterns are easier to learn than others, as this may reflect cognitive constraints

on the learning mechanisms.What remains to establish here, is whether

the limitations observed in the learning of statistical patterns by adult subjects

actually reflect constraints that bear on phonological acquisition (and a fortiori

on phonological acquisition by the child). An alternative possibility would be,

for instance, that certain syllables (say, those with /s/ in onset) are more easily

articulated than others for purely motoric reasons8. Such an alternative hypothesis

could be tested by assessing the baseline ease of articulation of the relevant

syllables in the absence of exposure to a biased distribution of segments.

5. Conclusion

In this paper, we have taken the papers by Fikkert, Vihman, and Goldrick and

Larson as case studies to analyse the many pitfalls that lie in the analysis of

linguistic and psycholinguistic data, and to argue for the necessity to always

analyse tasks within an information processing model in order to get clues

about the processes and levels of representations that the data may reflect.

Chomsky (1976) claimed already long ago that linguistics was a branch

of psychology. He was probably more right about it than most linguists (and

perhaps he himself) have realised. Here we argue that the methods of linguistics

should also be a branch of the methods of psychology. Indeed, all linguistic

data are behavioural data. Linguistic representations are hidden in the brain,

and can never be accessed directly by the experimenter. The experimenter can

only observe the behavioural data,which bear some (complex) relationship with

linguistic representations and grammar. Behavioural data are always collected

using a task. All tasks involve multiple levels of representation and processing.

A psycholinguistic perspective on the acquisition of phonology 333

Data interpretation therefore relies on figuring out which levels the data reflect,

hence on systematically distinguishing all the relevant levels of representation

and the associated grammars. Not surprisingly, it often takes a great deal of

methodological sophistication to design tasks that produce patterns of results

that can be unambiguously attributed to a given representation or process. To

hammer this point in, let us consider that even the simplest of all sources of

linguistic data, speaking, is a task in a non-trivial sense.We have seen for instance

in the case of child language how difficult it may be to attribute the observed

patterns to a particular level of representation or processing. This should increase

our awareness of the problems raised by the interpretation of data, and drive us

to enrich the methodological repertoire by drawing froma greater range of tasks

and by paying careful attention to the methodological conditions that may affect

task performance.

Acknowledgments

We wish to acknowledge feedback from Kie Zuraw, and financial support from the

European Commission (Neurocom project) and the Agence Nationale de la Recherche

(ANR-05-BLAN-0065-01). Correspondence to: Franck Ramus, LSCP, 29 rue d’Ulm,

75005 Paris, France. Email address: [email protected].

Notes

1. This is only true up to a point: speech corpora must be listened to and coded, which

inevitably recruits the listener’s own speech perception pathway, phonology, and

metalinguistic abilities.

2. Productivity remains a key criterion of grammatical processes. However, explicit

generation or judgement of morphological forms may not be a sufficiently stringent

test of productivity. It is likely that phonological, like syntactic grammatical processes,

are largely inaccessible to consciousness.When explicitly asked to generate

or judge a derived form, subjects may simply respond what they think is correct,

based on a rapid survey of similar examples in their mental lexicon.Therefore it may

not be surprising if such tasks yield results globally consistent with the statistics

of the lexicon, but it is not clear that this reflects real grammatical processes. One

may then wonder what kind of empirical data would constitute a sufficient proof of

productivity. The problem of experimentally tapping subjects’ unconscious mental

processes without being contaminated by their conscious beliefs and strategies is

a general one in psychology, whose solution is often to use more indirect methods

where the experimental factors being manipulated cannot be detected by subjects.A

well-known example is priming, where the relationship between a prime and a target

modulates a behavioural response (usually reaction time), unbeknownst to the subject.

In the present case, one might predict that, in a suitable experimental paradigm,

334 Franck Ramus et al.

word forms productively derived from one another would prime each other, whereas

more superficially related word forms would not (Kouider and Dupoux 2009).

3. It may be improper to talk about a phonological representation in the initial state, but

this may simply be understood as a higher-order auditory representation that adapts

to the sounds of a particular language, hence becomes phonological.

4. ‘Did’ may not be a very good example of a word very familiar to English-learning

babies, and moreover it is not imageable so as to be used in a preferential looking

experiment. But let us imagine that it were.

5. As soon as just two parts could be somewhat reliably identified, the representation

would not be holistic anymore. It might be underspecified, but not holistic (as Fikkert

points out).

6. Lexical decision involves deciding whether each item is a word or not, with typically

half the items being pseudowords.

7. To be entirely consistent, the model should show that the articulatory representation

also feeds into executive processes, which control speech just as much as other

behavioural output.

8. In addition, we find that the evidence that subjects’ production of /s/ in onset is less

affected by the exposure than that of /s/ in coda or /f/ in either onset or coda is quite

weak.Asignificant proportion × condition interaction, as well as a replicationwould

be in order before drawing conclusions on the special status of /s/ in onset.

References

Auerbach, Sanford H.,TerryAllard,MargaretNaeser,MichaelP.Alexander, andMartinL.

Albert

1982 Pure word deafness.Analysis of a casewith bilateral lesions and a defect

at the prephonemic level. Brain 105 (Pt 2):271–300.

Ballem, Kate D., and Kim Plunkett

2005 Phonological specificity in children at 1;2. Journal of Child Language

32 (1):159–73.

Best, Catherine T.

1995 A direct realist perspective on cross-language speech perception. In

Speech perception and linguistic experience: Theoretical and methodological

issues in cross-language speech research, edited byW. Strange,

167–200. Timonium MD:York Press.

Best, Catherine T., andWinifred Strange

1992 Effects of phonological and phonetic factors on cross-language perception

of approximants. Journal of Phonetics 20:305–331.

Boersma, Paul

1998 Functional phonology. University of Amsterdam: Ph.D. dissertation.

A psycholinguistic perspective on the acquisition of phonology 335

Boersma, Paul

2006 A programme for bidirectional phonology and phonetics and their acquisition

and evolution. In LOT Summerschool.Amsterdam.

Caramazza, Alfonso

1997 How many levels of processing are there in lexical access? Cognitive

Neuropsychology 14 (1):177–208.

Caramazza, Alfonso, Annamaria G. Basili, Jerry J. Koller, and Rita S. Berndt

1981 An investigation of repetition and language processing in a case of conduction

aphasia. Brain and Language 14 (2):235–71.

Caramazza, Alfonso, Rita S. Berndt, and Annamaria G. Basili

1983 The selective impairment of phonological processing: a case study. Brain

and Language 18 (1):128–74.

Chomsky, Noam

1976 Reflections on language. London: Temple Smith.

Chomsky, Noam, and Morris Halle

1968 The sound pattern of English. NewYork: Harper and Row.

Coltheart, Max

1978 Lexical access in simple reading tasks. In Strategies of Information Processing,

edited by G. Underwood, 151–216. London: Academic Press.

Darcy, Isabelle, Sharon Peperkamp, and Emmanuel Dupoux

2007 Bilinguals play by the rules: perceptual compensation for assimilation

in late L2-learners. In Laboratory Phonology 9, edited by J. Cole and J.

I. Hualde, 411–442. Berlin: Mouton de Gruyter.

Darcy, Isabelle, Franck Ramus, Anne Christophe, Katherine Kinzler, and Emmanuel

Dupoux

2009 Phonological knowledge in compensation for native and non-native assimilation.

In Variation and Gradience in Phonetics and Phonology,

edited by F. K¨ugler, C. F′ery and R. van de Vijver, 265–309. Berlin:

Mouton De Gruyter.

Dehaene-Lambertz, Ghislaine, Christophe Pallier, Willy Serniclaes, Liliane Sprenger-

Charolles,Antoinette Jobert, and Stanislas Dehaene

2005 Neural correlates of switching from auditory to speech perception. Neuroimage

24:21–33.

Duncan, Lynne G., Pascale Col′e, Philip H. K. Seymour, and Annie Magnan

2006 Differing sequences of metaphonological development in French and

English. Journal of Child Language 33 (2):369–399.

Dupoux, Emmanuel, Kazuhiko Kakehi, Yuki Hirose, Christophe Pallier, and Jacques

Mehler

1999 Epenthetic vowels in Japanese:Aperceptual illusion? Journal of Experimental

Psychology: Human Perception and Performance 25 (6):1568–

1578.

336 Franck Ramus et al.

Dupoux, Emmanuel, Christophe Pallier, Kazuhiko Kakehi, and Jacques Mehler

2001 New evidence for prelexical phonological processing in word recognition.

Language and Cognitive Processes 16 (5/6):491–505.

Dupoux, Emmanuel, Christophe Pallier, Nuria Sebastian, and Jacques Mehler

1997 A destressing “deafness” in French? Journal of Memory and Language

36:406–421.

Dupoux,Emmanuel,N′uria Sebasti′an-Gall′es,EduardoNavarrete, and Sharon Peperkamp.

2008. Persistent ’stress deafness’:The case of French learners of Spanish. Cognition

106:682–706.

Eimas, Peter D., Einar R. Siqueland, Peter W. Jusczyk, and JamesVigorito

1971 Speech perception in infants. Science 171:303–306.

Eisner, Jason

2002 Comprehension and compilation in Optimality Theory. Paper read at

40th annual meeting of the Association for Computational Linguistics,

at Philadelphia.

Faber, Alice, and Catherine T. Best

1994 The perceptual infrastructure of early phonological development. In The

Reality of Linguistic Rules, edited by R. Corrigan, S. D. Lima and G.

Iverson, 261–280. Amsterdam: John Benjamins.

Fennell, Chris T., and Janet F.Werker

2003 Early word learners’ ability to access phonetic detail in well-known

words. Language and Speech 46:245–264.

Flemming, Edward

2001 Scalar and categorical phenomena in a unified model of phonetics and

phonology. Phonology 18:7–44.

Fowler, Carol A., Shirley E. Napps, and Laurie Feldman

1985 Relations among regular and irregular morphologically related words in

the lexicon as revealed by repetition priming. Memory & Cognition 13

(3):241–255.

Friederici, Angela D., Manuela Friedrich, andAnne Christophe

2007 Brain responses in 4-month-old infants are already language specific.

Current Biology 17 (14):1208–1211.

Friederici, Angela D., and Jeanine M. I.Wessels

1993 Phonotactic knowledge of word boundaries and its use in infant speech

perception. Perception & Psychophysics 54 (3):287–295.

Gaskell, M. Gareth, andWilliam D. Marslen-Wilson

1996 Phonological variation and inference in lexical access. Journal of ExperimentalPsychology:

HumanPerception&Performance 22 (1):144–158.

Gierut, JudithA., and Holly L. Storkel

2002 Markedness and the grammar in lexical diffusion of fricatives. Clinical

Linguistics & Phonetics 16:115–134.

A psycholinguistic perspective on the acquisition of phonology 337

Goldinger, Stephen D.

1996 Words and voices: Episodic traces in spoken word identification and

recognition memory. Journal of Experimental Psychology: Learning,

Memory, and Cognition 22 (5):1166–1193.

Goldinger, Stephen D.

1998 Echoes of echoes? An episodic theory of lexical access. Psychological

Review 105 (2):251–279.

Goldinger, Stephen D.

2007 Acomplementary-systems approach to abstract and episodic speech perception.

Paper read at 16th International Congress of Phonetic Sciences,

at Saarbr¨ucken.

Goldrick, Matthew, and Brenda Rapp

2007 Lexical and post-lexical phonological representations in spoken production.

Cognition 102 (2):219–260.

Goto, H.

1971 Auditory perception by normal Japanese adults of the sounds “L” and

“R”. Neuropsychologia 9:317–323.

Hume, Elizabeth, and Keith Johnson

2001 Amodel of the interplay of speech perception and phonology. InThe role

of speech perception in phonology, edited by E. Hume and K. Johnson:

Academic Press.

Jacquemot, Charlotte, Emmanuel Dupoux, and Anne-Catherine Bachoud-L′evi

2007 Breaking the mirror: Asymmetrical disconnection between the phonological

input and output codes. Cognitive Neuropsychology 24 (1):3–22.

Jacquemot, Charlotte, Christophe Pallier, Denis LeBihan, Stanislas Dehaene, and

Emmanuel Dupoux

2003 Phonological grammar shapes the auditory cortex: a functionalmagnetic

resonance imaging study. Journal of Neuroscience 23 (29):9541–9546.

Johnson, Keith

1997 Speech perception without speaker normalization. In Talker variability

in speech processing, edited by K. Johnson and J. Mullennix, 145–166.

San Diego: Academic Press.

Jusczyk, PeterW.

1997 The discovery of spoken language. Cambridge, MA: MIT Press.

Jusczyk, PeterW., Anne Cutler, and Nancy J. Redanz

1993 Infants’ preference for the predominant stress patterns of English words.

Child Development 64:675–687.

Jusczyk, PeterW., Derek M. Houston, and Mary Newsome

1999 The beginnings of word segmentation in English-learning infants. Cognitive

Psychology 39 (3/4):159–207.

Jusczyk, PeterW., Paul A. Luce, and Jan Charles-Luce

1994 Infants’ sensitivity to phonotactic patterns in the native language. Journal

of Memory and Language 33:630–645.

338 Franck Ramus et al.

Keating, Patricia

1990 Phonetic representations in a generative grammar. Journal of Phonetics

18:321–334.

Kouider, Sid, and Emmanuel Dupoux

2005 Subliminal Speech Priming. Psychological Science 16 (8):617–625.

Kouider, Sid, and Emmanuel Dupoux.

2009. Episodic accessibility and morphological processing: Evidence from

long-term auditory priming. Acta Psychologica 130:38–47.

Kuhl, Patricia K.

2000 A new view of language acquisition. Proceedings of the National Academy

of Science USA 97 (22):11850–11857.

Kuhl, Patricia K., KarenA.Williams, Francisco Lacerda, Kenneth N. Stevens, and Bj¨orn

Lindblom

1992 Linguistic experience alters phonetic perception in infants by 6 months

of age. Science 255 (5044):606–608.

Levelt,Willem J. M.

1989 Speaking: From Intention to Articulation. Cambridge, MA: MIT Press.

Levelt,Willem J. M.

1992 Accessing words in speech production: Stages, processes and representations.

Cognition 42:1–22.

Lively, Scott E., John S. Logan, and David B. Pisoni

1993 Training Japanese listeners to identify English /r/ and /l/. II: The role

of phonetic environment and talker variability in learning new perceptual

categories. Journal Of The Acoustical Society Of America 94

(3 Pt 1):1242–1255.

Lively, ScottE.,DavidB. Pisoni,ReikoA.Yamada,Yoh’ichiTohkura, andTsuneoYamada

1994 Training Japanese listeners to identify English /r/ and /l/. III. Long-term

retention of new phonetic categories. Journal Of The Acoustical Society

Of America 96 (4):2076–2087.

Luce, Paul A., and Emily A. Lyons

1998 Specificity of memory representations for spoken words. Memory &

Cognition 26 (4):708–715.

Mattock, Karen, and Denis Burnham

2006 Chinese and English infants’ tone perception: Evidence for perceptual

reorganization. Infancy 10:241–265.

Metz-Lutz, Marie-No¨elle, and E. Dahl

1984 Analysis of word comprehension in a case of pure word deafness. Brain

and Language 23 (1):13–25.

Morton, John

1969 The interaction of information in word recognition. Psychological Review

76:165–178.

A psycholinguistic perspective on the acquisition of phonology 339

Morton, John

1980 The logogen model and orthographic structure. In Cognitive processes

in spelling, edited by U. Frith, 117–133. London: Academic Press.

Myers, James

1999 Lexical phonology and the lexicon. Rutgers Optimality Archive 330–

699.

Norris, Dennis, James M. McQueen, and Anne Cutler

2000 Merging information in speech recognition: feedback is never necessary.

Behavioral and Brain Sciences 23 (3):299–325.

Pater, Joe, Christine L. Stager, and Janet F.Werker

2004 The perceptual acquisition of phonological contrasts. Language 80 (3):

384–402.

Pierrehumbert, Janet B.

2001 Exemplar dynamics: Word frequency, lenition, and contrast. In Frequency

effects and the emergence of lexical structure, edited by J. Bybee

and P. Hopper, 137–157. Amsterdam: John Benjamins.

Pierrehumbert, Janet, Mary Beckman, and Robert Ladd

2000 Conceptual foundations of phonology as a laboratory science. In Phonological

Knowledge: Conceptual and Empirical Issues,edited byN. Burton-

Roberts, P. Carr and G. Docherty, 273–304. Oxford: Oxford University

Press.

Pisoni, David B., and SusannahV. Levi

2007 Representations and representational specificity in speech perception

and spokenword recognition. In Oxford Handbook of Psycholinguistics,

edited by M. G. Gaskell, 3–18. Oxford: Oxford University Press.

Prince, Alan, and Paul Smolensky

1993 Optimality theory: Constraint interaction in generative grammar. New

Brunswick: Rutgers University.

Saffran, Jenny R., Richard N.Aslin, and Elissa L. Newport

1996 Statistical learning by 8-month-old infants. Science 274:1926–1928.

Sancier, Michele L., and Carol A. Fowler

1997 Gestural drift in a bilingual speaker of Brazilian Portuguese and English.

Journal of Phonetics 25:421–436.

Schriefers, Herbert, Antje S. Meyer, andWillem J. M. Levelt

1990 Exploring the Time Course of Lexical Access in Language Production

- Picture-Word Interference Studies. Journal of Memory and Language

29 (1):86–102.

Sheldon, Amy, andWinifred Strange

1982 TheAcquisition of /r/ and /l/ by Japanese Learners of English: Evidence

That Speech Production Can Precede Speech Perception. Applied Psycholinguistics

3 (3):243–261.

340 Franck Ramus et al.

Singh, Leher, Katherine S.White, and James L. Morgan.

2008. Building a word-form lexicon in the face of variable input: Influences

of pitch and amplitude on early spoken word recognition. Language

Learning and Development 4 (2):157–178.

Smith, Neil

1973 The Acquisition of Phonology. A Case Study. Cambridge: Cambridge

University Press.

Stager, Christine L., and Janet F.Werker

1997 Infants listen formore phonetic detail in speech perception than in wordlearning

tasks. Nature 388 (6640):381–382.

Steriade, Donca

1995 Underspecification and markedness. In The handbook of phonological

theory, edited by J. Goldsmith, 114–174. Oxford: Blackwell.

Swingley, Daniel, and Richard N.Aslin

2000 Spoken word recognition and lexical representation in very young children.

Cognition 76 (2):147–166.

Swingley, Daniel, and Richard N.Aslin

2002 Lexical neighborhoods and the word-form representations of 14-montholds.

Psychological Science 13 (5):480–484.

Werker, Janet F., Leslie B. Cohen, Valerie L. Lloyd, Marianella Casasola, and Christine

L. Stager

1998 Acquisition of word-object associations by 14-month-old infants. Developmental

Psychology 34 (6):1289–1309.

Werker, Janet F., and Richard C. Tees

2005 Speech perception as a windowfor understanding plasticity and commitment

in language systems of the brain. Developmental Psychobiology

46 (3):233–251.