1.1 L1 perception is not always perfect

If phonological distinctiveness were the only thing that mattered to a native perceiver, we would expect speakers of a language to be equally good at discriminating all phonemes of their language, irrespective of what measurable acoustic or articulatory similarities and differences might exist between them. The centrality of phonological distinctiveness to native speech perception is exemplified by the phenomenon of categorical perception, which is characterised by great ease in discriminating between consonants that are contrastive, such as (for English speakers) /t/ and /d/, and of great difficulty in discriminating between acoustically different tokens from the same phonological category, e.g. initial /d/ realized with -30 ms voicing lead (prevoicing) versus initial /t/ realised with 0 ms voicing lead, similar to French, which are both perceived as /d/ by English listeners. If phonological distinctiveness were the only relevant parameter, English phonemes such as /p/ and /n/, which differ in terms of place and manner of articulation and voicing characteristics, should be just as easy or just as difficult for native listeners to discriminate as /p/ and /b/, which differ only in terms of voicing. The same should be true for vowels as well (although they are not perceived as categorically as consonants): Australian English vowels /ɪ/ and /i/, which differ in duration and spectral quality, should be as easy to discriminate for a native speaker as /u/ and /e/.

However, research shows that this is not the case. In particular for vowels, but even in the case of consonants, it has been shown that performance on native language speech perception tests is not always perfect: it can be very difficult to correctly identify native phonemes that are acoustically and articulatorily highly similar. In the absence of lexical information, some English fricatives (e.g., /f/ and /θ/) are notoriously difficult to distinguish, even for native listeners ([4, 5, 16], among others). [17] showed that native speakers of Canadian English often confuse native /ɛ/ for /æ/ (15%), and sometimes /i/ for /ɪ/ (5%), and /ʊ/ for /ʌ/, /u/ and /ɔ/ (between 5–2% respectively), but never confuse /ɪ/, /æ/, /ɜ/ and /a/ for any other vowel. In some cases, native phoneme discrimination may even be poorer than non-native discrimination. [18], for example, compared the discrimination of English /w j/ by native speakers of English and Danish and found that the Danish participants, for whom this is not a native contrast, much more accurately discriminated the contrast than native English speakers. They ascribed the superior non-native performance to the fact that there are many more approximant phones (phonemes and allophones, and several vowel rounding settings) in Danish than in English, providing the Danish listeners with greater opportunities to fine-tune their perceptual system to approximants.

Thus far, native speaker perceptual difficulties in consonants have only been demonstrated for continuant (e.g., fricatives) and sonorant (e.g., approximants) segments, and not for stops. The reasons for this are presumably to do with the differences between categorical and gradient (continuous) perception.

We also find suggestive (indirect) evidence that some phones might be more difficult to perceive in that some phonemes are much rarer than others cross-linguistically. One account for these inequalities is that those which occur more frequently are more differentiated from others acoustically and perceptually, which results in more linguistic redundancy and fewer misperceptions. Conversely, the rarity of certain other contrasts, such as that between dental and alveolar fricatives, may be due to their low perceptual salience and/or discriminability from adjacent consonants [19, 20].

Finally, difficulties in perception of phonological contrasts have been identified as a leading cause of sound change in languages by a number of researchers, in particular [6, 7, 8] as well as others [21, 22, 23]. Notably, in such cases, it is the native speakers of a language who are central to the scenario of perceptual difficulty leading to phonological re-analysis.

These multiple strands of independent research indicate that some native phonemic contrasts are more difficult to categorise and discriminate than others. Such difficulties have not yet been, but need to be, adequately addressed by the leading theories of speech perception. In the following, we show how PAM/PAM-L2 has the necessary conceptual framework to account for such patterns and provide testable hypotheses, if applied in conjunction with our novel proposal of innate perceptual biases in consonant perception: the Natural Referent Consonant hypothesis.

1.2 Extending PAM/PAM-L2 to difficult native phonological contrasts

Despite the above observations that some native phonemic contrasts might be more difficult than others to identify and discriminate, current theories of first (L1) and second (L2) language speech perception such as the Perceptual Assimilation Model (PAM; [1, 2]), and its extension PAM-L2 [3], the Speech Learning Model (SLM: [24]), and the Native Language Magnet model (see for instance [25]), have not yet systematically considered the theoretical implications of evidence that some contrasts might be inherently difficult to perceive, even for native speakers. This is reflected in the fact that within the cross-language literature speech perception is rarely tested using native stimuli, except to provide a control baseline for comparison with non-native listeners' performance (and here, the expectation, and perhaps requirement, is often that native perception will be near-perfect).

The Perceptual Assimilation Model and PAM-L2, however, are able to account for such a pattern of perception relatively straightforwardly. PAM/PAM-L2 assumes that native consonant perception is categorical, yet at the same time recognises a relationship between categoricity and gradiency in perception. For example, the distinction between the ‘Category Goodness’ (CG) assimilation type versus the ‘Single Category’ (SC) assimilation type (see definitions below) is, in its essence, about the perceptibility of within-category gradient variation.

The tenets of PAM/PAM-L2 allow us to examine the native phonological system of an individual by presenting him or her with phones from a different language that do not align with the phoneme boundaries of the native language of the listener. According to PAM/PAM-L2, non-native phoneme contrasts will be perceived in one of the following ways: a Two-Category (TC) assimilation where the two non-native phones are perceived as instances of two separate native phonological categories and discrimination is excellent; a Category Goodness (CG) assimilation where the two non-native phones are perceived as varying instances of the same native phoneme in which one is more native-like than the other and discrimination is moderate to good, depending on the perceptual distance between the two; Single Category (SC) where the two non-native phones are perceived as equally good (or poor) instances of a single native phoneme, and discrimination is expected to be poor; an Uncategorised-Categorised (UC) assimilation where one non-native phone is perceived as an instance of a native phoneme, but the other phone is not assimilated into any single native phoneme category or perhaps weakly assimilated to two or more categories, and discrimination is expected to be moderate to excellent, depending on the perceptual distance between the two; and finally an Uncategorised-Uncategorised (UU) assimilation where neither non-native phone is assimilated into a single native phoneme category, and discrimination success depends on the perceptual distance and extent of multiple-category overlaps between the two.

No study has hitherto applied PAM/PAM-L2 and the related experimental paradigm to adult native speech perception in scenarios where there is a high degree of acoustic and articulatory, and possibly perceptual, similarity between two contrasting L1 phonemes, resulting in less-than-perfect discrimination, even by native listeners. (A number of studies have, however, investigated infant perception of acoustically/articulatorily similar native phones; see for instance [26], and [27, 28, 29, 30]). Indeed, we propose that two distinct consonant categories, if very close together in articulatory and/or acoustic space, may overlap phonetically to such an extent that even native listeners may become unsure of whether they are hearing two distinct phonemes (a TC discrimination of highly similar phones). In the case of the Wubuy alveolar versus retroflex contrast for instance, we argue that the phonetic overlap is so great that native listeners can behave as if this were an instance of a good and a 'less good' instance of a single category (CG). Such a pattern has also been shown for fricatives in English (see references in Section 1.1). Especially if there is no lexical information to assist them, as is often the case in studies of segmental categorisation and discrimination, it is likely that discrimination accuracy will be commensurate with the discrimination accuracy expected for a non-native CG contrast, even for native listeners. As our review of Wubuy below indicates, such a scenario may indeed exist in a number of Australian languages, which have up to four coronal places of articulation for stops.

In cases of extreme perceptual difficulty, such as we are proposing, listeners may rely on universal biases in speech perception, which may be reflected in asymmetries in discrimination performance (i.e., greater ease in discriminating two phones in one presentation order than in the opposite order). Until now, evidence for perceptual asymmetries in discrimination studies, as reflections of universal biases, has been largely confined to studies of vowel perception in L1 infants and L2 adult listeners. A few studies have also found perceptual asymmetries in adult L2 perception of stop-fricative consonants /b v/ [31]. It is unclear, however, to what extent contrasts that differ in manner of articulation can inform us with respect to contrasts that differ in place of articulation. Finally, there are priming and Event Related Potential (ERP) studies that suggest that coronals may suffer an asymmetry with respect to stops with other places of articulation (labial in particular) [32; 33], though other studies using ERP [34] and eye-tracking [35] suggest that these effects do not provide evidence for a universal bias in the perception of coronals vis-a-vis phones with other places of articulation, but may instead be due to differences in the distribution of these places of articulation within the lexicon and thus to listeners' expectations about the likelihood of a particular place of articulation in a particular position within the stimulus. To our knowledge, perceptual asymmetries in discrimination have not been found, until the current study, in L1 adults, nor in the perception of consonants when manner of articulation is held constant.

Polka and colleagues ([12, 13, 14, 15]; also [36]) observed that infant L1, as well as adult L2, vowel discrimination is much better when the participants are first presented with a more central vowel and then tested for detection of change with a more peripheral vowel. More peripheral vowels appear to act as perceptual 'anchors' for detecting changes from one vowel to another. According to [13], the observed infant L1 asymmetries:

‘… could not be explained either by considering the status of the vowel in the infant's native language (i.e. whether or not the anchor vowel is in the infant's native inventory) nor by referring to universally favored vowels (i.e. predictions based on markedness). These effects were also not consistent with an acoustic bias related to a simple increase or decrease in either F1 or F2 and could not be due to differences in amplitude, duration or pitch […]. However, all of these directional asymmetries could be predicted by considering the relative position of each vowel in the vowel space defined by dimensions of vowel height […] and front/back[…]. Within this space, [..], a vowel change from a more central to a more peripheral vowel is consistently easier to discriminate compared to the same change presented in the reverse direction […]’ (p. 467).

In addition to being crucial to the understanding of potential biases in perception in general, we also believe that a focus on asymmetrical discrimination performance is crucial to understanding the effect that differences in the relative distributional frequency of individual phonemes in the input may have on discrimination of native contrasts involving those phonemes. Indeed, we propose that innate perceptual biases will only be overridden by language specific perception when there is sufficient input, i.e. relatively high frequency of occurrence, to 'reset' the perceptual system. We expand on these issues in the following section, introducing a 'Natural Referent Consonant' Hypothesis first, and then discussing the potential effects on discrimination of differences in phone frequency in the input.

1.3 A Natural Referent Consonant Hypothesis

Here, we propose a Natural Referent Consonant (NRC) hypothesis for the study of perceptual asymmetries in consonant perception. We propose that such a framework will add to the general understanding of native and non-native consonant perception spearheaded by models such as PAM/PAM-L2. The NRC hypothesis is compatible with PAM/PAM-L2, and likely also with other speech perception theories.

While the articulatory and acoustic space of consonants may not lend itself to the same notion of peripheries in an acoustic-phonetic space as is crucial to the Natural Referent Vowel hypothesis, the three major articulators (lips, tongue front, tongue back) may provide equivalent perceptual and articulatory ‘anchor points’, corresponding to the three major oral places of articulation: labial, coronal, and dorsal (c.f., [37]). From an acoustic and perceptual point of view, such articulatory anchor points are perhaps reflected in phonemes with excellent differentiation. Indeed, it might be possible to establish quantitatively that in acoustic and perceptual terms, labial, alveolar and velar are maximally differentiated from each other, although we do not pursue this here. Perhaps not surprisingly, the vast majority of the world’s languages use all—and only—these three places of articulation in consonants. Only a small percentage has multiple sub-places of articulation (manner held constant) within one of these major places. (We infer this from [19]: ‘In a very high proportion of the world's languages, segments with the same manner must be drawn from different active articulator classes’ (p. 43). However, we note that a contrast between velar and uvular voiceless stops is not uncommon, found in many Afro-Asiatic, Caucasian, North American and Turkic languages [19].) Our target language Wubuy is one of these rare languages with multiple coronal sub-places within the same manner classes.

The NRC hypothesis gives rise to a number of predictions on the basis of the general assumptions that:

1. 1). the most canonical articulation, by which we mean the one occurring most frequently for a given major place across the world's languages, will be the perceptual anchor, and that
2. 2). arguably, the canonical articulation for the coronal place is apical orientation of the tongue tip, with contact made in the alveolar region.

Indeed, the case can be made that this is an 'easy-to-achieve' coronal articulation, which provides excellent acoustic distinctiveness from the surrounding major places of articulation. [38] provides explicit discussion of ‘neutral’ vs ‘displaced’ articulations and argues that dental is the neutral passive articulation for apicals, and alveolar for laminals. Considerations of cross-linguistic frequency suggest a different interpretation, however. According to [19] 'In general, if a language has only a dental or an alveolar stop, then that stop will be laminal if it is dental and apical if it is alveolar' (p. 26; they go on to urge caution with the terms 'dental' and 'alveolar'). In addition, however, they note (op. cit.) that 'if a language has both an apical and laminal stop consonant, then the laminal consonant is likely to be more affricated'. It is unclear whether lamino-postalveolar stops also provide a canonical articulation, partly because obstruents with this articulation tend to be affricated in contrast with apical stops in the same inventory. However, according to the phonological database compiled by [39], languages with a single coronal place far outnumber those with two or more (82% for 1 vs 18% for 2 vs .04 for 3 or more). Together with the point that [19] make (above), we take this to be evidence that there are unlikely to be two canonical coronal articulations.

These biases may (indeed, must) be over-ridden by the phonological system of the native language that an infant learns when this system differs from the biases. However, we propose that they may persist in situations where the input frequency for some of the respective consonant phonemes is too low to provide robustly distinct categories for acoustically and articulatorily highly similar phones. In addition to these relatively straightforward analogues of the Natural Referent Vowel hypothesis, therefore, we suggest that the distributional frequencies of consonants overall, or in particular positions in words (syllabic, morphological), will be relevant to native listeners' discrimination. In particular we propose that:

1. 3). Consonant contrasts in which one member occurs infrequently in the speech input may continue to display the innate bias to the Natural Referent Consonant if the canonical articulation is also the more frequent member of the contrast. By this we suggest that only frequently occurring non-canonical consonants are likely to over-ride a natural initial perceptual asymmetry, while under-represented consonants are more likely to continue to display asymmetrical perception past infancy and completion of L1 acquisition and into adulthood. Below, we propose that this is the case specifically with retroflex stops in Wubuy.
2. 4). Speakers of languages that make use of only the three major oral places of articulation (labial, coronal, dorsal) and that use no distinctions within any of the major places for a given manner class will show perceptual asymmetries when presented with non-native consonants that make use of finer place contrasts within one of the major places.

The characteristics of Australian languages provide an optimal test-case for these predictions. Many have three or four-way within-coronal place contrasts in stops (and nasals and laterals). Commonly, some of these contrasts are neutralised, or are unevenly distributed, in one or more positions in words, creating the conditions for stark differences in the distributional frequency among the individual coronal consonants. We turn next to a brief description of our target langauge, Wubuy, with particular focus on its phonological and phonetic characteristics.

1.4 Wubuy

Wubuy, also known as ‘Nunggubuyu’ [11], is an indigenous Australian language spoken in south-eastern Arnhem Land, traditionally around the southern part of Blue Mud Bay in the Gulf of Carpentaria and now mostly in the settlement of Numbulwar. It is the first language for adults over the age of around 55 in the community, as well as a first or second language for many adults of the same generation in neighbouring Groote Eylandt. Wubuy is a critically endangered langauge—children are no longer acquiring it as a first language, though all children in the community are exposed to some Wubuy, both through their grandparents and through local language revitalisation efforts at the school. There are perhaps 60 first language speakers of Wubuy in Numbulwar.

The consonantal inventory of Wubuy is presented in Table 1. As can be seen from an examination of this table, the phonology of Wubuy has a four-way place distinction among the coronal stops /t ʈ ȶ/ (apico-alveolar, apico-retroflex, lamino-dental, lamino-alveopalatal). The contrast is found word-initially as well as medially, as noted above. This is unusual for languages with multiple coronal stop contrasts. A survey of the phonotactics of Australian languages presented in [10] identifies just 10 languages (from 116) where the four coronal stops are contrastive initially, as well as medially; most of these languages are now, however, extinct or moribund. Despite being endangered, Wubuy is among the 50 or so Australian languages that still have a sizeable speech community of more than 5–10 speakers [40] and, of those, one of the very few with an initial four-way contrast.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The phonemic consonant inventory of Wubuy, adapted from [11].

https://doi.org/10.1371/journal.pone.0142054.t001

Previous acoustic work on Wubuy [41, 42] shows that there are very small acoustic differences among three of the coronal stops: the apico-alveolar /t/, the apico-retroflex /ʈ/, and the lamino-dental //. We refer to this group, following [43], as the [—sharp] set of coronals. We justify this grouping further in section 4.1.1. The great degree of acoustic and articulatory overlap among the 3 [—sharp] coronals makes Wubuy ideal for testing the limitations of native segmental perception in the absence of lexical information.

Given suggestions in the literature that 'listeners are biased towards accepting a frequent pattern more often than an infrequent one' [44], we examined the distributional properties of coronal stops in Wubuy. Indeed, the distributional frequencies of the coronal stops in spoken Wubuy differ systematically across the four places of articulation in ways that provide opportunity for evaluating the influence of input frequency on overriding innate perceptual asymmetries among coronal stops, discussed in [45]. A phoneme count of 8% of the Wubuy text corpus [46] (i.e., six of the texts), shows that the lamino-alveopalatal stop is far more frequent than the dental, alveolar and retroflex stops, except in word-initial position where the alveolar is somewhat more frequent (we use 'word' here in the ordinary syntactic sense of 'terminal elements in phrase structure'). The lamino-alveopalatal is also much more frequent than the other coronal stops in initial position in verb stems (see Table 2), which typically does not correspond to word-initial position because Wubuy is a language with obligatory inflectional prefixes (see, for instance, [47, 48, 49, 50] for discussion of the cognitive saliency of stems in prefixed words). The retroflex is infrequently represented, both overall and also in stem-initial position in both verbs and nouns. This is what distinguishes the retroflex from the dental, which, although infrequent word-initially, is equal to the alveolar in the highly salient verb stem-initial position in which both are twice as frequent as the retroflex. Thus, combining the two most prominent phonotactic and morphological positions—word-initial and verb stem-initial—the order of frequency of occurrence of the coronal stops overall is: ȶ > t, > ʈ. Wubuy thus provides not only an optimal stop consonant inventory for the testing of the basic assumptions of an NRC hypothesis, it also provides differences in the distributional frequency of those native stop consonants in the spoken input, which may affect the status of these segments in the minds of speakers.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Frequencies of stop consonant phonemes in all words in 8% of the Wubuy text corpus [46], adapted from [45].

The ‘Initial’ column presents the word-initial distribution of stops in the words in the Wubuy texts, while the ‘Total’ column indicates the distribution of stop phonemes in all word positions. The ‘Verb Initial’ column indicates the distribution of stop consonants in verb stem-initial position.

https://doi.org/10.1371/journal.pone.0142054.t002

1.5 Predictions

In the following, we present two two-part studies of perception of the acoustically and articulatorily similar Wubuy coronal stops /t ʈ ȶ/. In study 1a and 1b, we tested native speakers of Wubuy and, in study 2a and 2b, we tested native speakers of Australian English as listeners who are naïve to this contrast series. Studies 1a and 2a presented the consonant targets in word-medial position (/aCa/). Studies 1b and 2b tested discrimination of the three [—sharp] stops in absolute initial position (/##Ca/).

In the case of the Wubuy listeners, we test the following competing hypotheses:

H1) All phoneme categories are distinct entities in the phonological grammar of a listener, and native language learning ensures that all native phonemes are perceived in equally categorical fashion. In the light of the literature reviewed above, however, this general hypothesis is considered unlikely to be supported.

H2) Native phoneme discrimination performance is subject to factors such as acoustic and/or articulatory similarity/overlap among native phonemes, as well as by frequency and distribution in the input. This hypothesis is compatible with PAM/PAM-L2 in conjunction with the NRC hypothesis, as argued above. For the present studies, this hypothesis would predict that:

H2a: acoustic/articulatory distinctiveness) Acoustically and articulatorily well-differentiated contrasts, i.e., /ȶ t/, /ȶ t/ and /ȶ ʈ/, will be more successfully discriminated even by non-native listeners than those involving contrasts between [—sharp] coronals, as the former three contrasts will likely result in a typical TC discrimination pattern in PAM/PAM-L2 terminology.

H2b: acoustic/articulatory overlap) In contrast, the [—sharp] stops are more likely to be confused with each other and result in poorer, though still above chance, native discrimination scores. The discrimination accuracy will depend on the specific acoustic/articulatory similarities and differences of each pair, and also by the consonants’ frequency and distribution in spoken Wubuy (because over-riding the bias towards the canonical articulations requires sufficient input: see H2c). We predict good discrimination of / ʈ/ as these are the most articulatorily and acoustically distinct of the contrasts among [—sharp] coronal stops ([43]; and c.f., [42]).

H2c: effects of frequency in the native input) Despite the claim that / ʈ/ are the most distinct pair of [—sharp] coronal stops, we nevertheless expect that discrimination of /t / will be better than both / ʈ/ and /t ʈ/, because of their differing lexical frequencies ([t, ] > ʈ). Specifically, we expect / ʈ/ and /t ʈ/ to exhibit asymmetries in discrimination, akin to those reported for vowels [12, 13, 14, 15]. Indeed, we predict that both the alveolar and the dental stops will act as perceptual anchors, due to their greater frequency in the input, and in the case of the alveolar, due to it being the 'canonical' coronal stop. Moreover, because of this we expect the relative performance levels on these contrasts to most clearly differentiate the native Wubuy listeners and Wubuy-naïve English listeners.

H2d: effect of context) Finally, we predict that Wubuy speakers will perform similarly in the two contexts examined here: /aCa/ and /##Ca/, with the exception of the /t ʈ/ contrast, which they should have more difficulty with in initial than in medial position as /ʈ/ is extremely rare in initial position (and also because of claims in the literature that /ʈ/ is difficult to distinguish from /t/ in the absence of a preceding vowel [9, 10, 22]).

In the case of the Wubuy-naïve English listeners, we test the following hypotheses:

H3) English listeners will perceive /ȶ/ as an instance of English /tʃ/ (or /dʒ/), while the [—sharp] stops will all be perceived as English /t/ (or /d/): /t/ will be perceived as a perfectly good /t/, while both // and /ʈ/ will be perceived as (somewhat) ‘odd’ /t/s. Contrasts involving the lamino-alveopalatal will thus become TC assimilations in PAM/PAM-L2 terminology, and discrimination accuracy is expected to be excellent. The [—sharp] contrasts will be perceived as CG contrasts, and discrimination will be poor to moderately good, depending on the magnitude of the acoustic and articulatory difference between the stops in each contrast. This is consistent with the report that native speakers of English struggle to discriminate English (alveolar, short-lag voiced) /d/ and French (dental, prevoiced) /d/, presumably because they perceive both the native and non-native /d/ as instances of the same native phonological category [51].

H4) On the basis of the NRC hypothesis, we further predict that order of presentation will be important for the two most acoustically/articulatorily similar coronal contrasts, / t/ and /ʈ t/ such that it will be easier to discriminate these contrasts when the 'odd' (and non-canonical) consonant is presented first.

H5) Finally, we expect English listeners to perform equally well in the two positional contexts, as they have no previous experiences with multiple coronal stops in either context.

2.1 Stimuli

We recorded three female native speakers of Wubuy (ages 51–61 years), born and raised in the Numbulwar area by native speakers of Wubuy. Two participants also reported speaking the neighbouring Aboriginal language Enindhilyakwa (Groote Eylandt) with relatives (grandparents, in-laws), and all understand and speak the community language Roper Kriol (an English-lexified creole: [52]) to some extent. All three speakers had acquired English as a second language in a classroom setting. All had basic linguistic training.

Each of the three participants produced the four target consonants /t ʈ ȶ/ in the two contexts given in (1), except for the lamino-alveopalatal /ȶ/ which was only produced in the intervocalic /aCa/ context. The target word list is shown in Table 3. Note that the primary stress is on the first syllable of each root word.

1. (1). Context 1 Phrase-medial, word-internal /aCa/ targets
   Context 2 Utterance initial, word-initial /##Ca/ targets

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Target words for each context.

Note that the orthographic representation (leftmost column for each context) represents the alveolar stop by ‘d’, the retroflex by underscoring ‘d’, and the dental by ‘dh.’ The other columns provide a phonetic transcription of the standard Wubuy pronunciation of the word in (IPA), and its English gloss.

https://doi.org/10.1371/journal.pone.0142054.t003

The target words were selected to provide a symmetrical vowel context on either side of the consonants in the /aCa/ context. The Wubuy carrier phrases were chosen so as to minimize coarticulation with the adjacent edges of the carrier phrase. The carrier phrase for the word-medial elicitations is given in (2a) in both Wubuy orthography and IPA. The carrier phrase for the utterance-initial elicitations, in (2b), is a re-ordering of the same words. Wubuy syntax is non-configurational [11], and both carrier phrases are acceptable to native speakers.

(2) a. ‘nga-yamana __________adaba’    b. ‘__________ nga-yamana adaba’

          [ˈŋa-jamana ________ˈaʈapa]            [ __________ ˈŋa-jamana    ˈaʈapa]

          1SG-say.PRS                now                                       1SG-say.PRS    now

          ‘I say __________        now’                 ‘__________, I say now’

2.2 Recording procedure and stimulus treatment

The three speakers read the target words in the carrier sentences presented in Wubuy orthography on a computer monitor in a fixed order, blocked by the type of consonant. The participants were encouraged to discuss and rehearse the words prior to the recording to ensure recognition of all target items during the recording. The participants were instructed to speak in a clear, comfortable voice as though they were speaking to a friend. Five correct utterances (as judged by the speaker herself, as well as by the other speakers who were present in the room during the recording, though they remained quiet throughout, only motioning silently with face/head and hands) were recorded for each target, resulting in a total of 45 correct utterances (5 tokens per 4 targets in the /aCa/ context, and 5 tokens per 3 targets in the /##Ca/ context). Recordings of targets containing coughs, stutters or speech or reading errors were discarded/replaced.

For the recording, we used a Shure SM10A headset cardioid microphone, an EDIROL UA-25 USB audio interface, and a laptop computer with Cool Edit 2000. All recordings had a 16-bit sampling depth with a sampling rate of 44.1 KHz. The recordings were made in a sound-attenuated professional recording studio at MARCS Institute in Sydney. The target /aCa/ and /##Ca/ sequences were excised using a Praat script and checked by the first and second author. Each excised token was given a 20 millisecond ramp-in and a 10 millisecond ramp-out using Praat.

2.3 Stimulus presentation

The excised /aCa/ and /##Ca/ targets were presented to our participants in the form of two separate randomised cross-speaker categorical XAB discrimination tasks (for both the /aCa/ and /##Ca/ contexts) programmed in Psyscope X, with the stimuli presented over headphones from a MacBook laptop computer. A cross-speaker discrimination task presents the listener with speech tokens from three different speakers in each XAB triad, and thus forces the listeners to disregard differences in voice quality and other idiosyncracies, and instead to focus exclusively on the phonological information in order to complete the task successfully.

The XAB discrimination task was explained to the participants as a task in which a ‘teacher’ (the first voice heard) was being imitated by a ‘good student’ and a ‘bad student’ (voices 2 and 3), and it was the job of the participant to indicate (with a key press on the keyboard) which of the two students (voice 2 or 3) was the ‘good student’ who copied correctly what the teacher had said. We provided this explanation to make the experimental paradigm meaningful to our participants and increase the likelihood that they would understand the task at hand and be able to complete the experiments. Similar stories and explanations are often provided in perception research, especially to children and other participants who are unfamiliar with experimental research.

For both studies, the inter-stimulus interval (ISI) between stimuli within each XAB trial was 500 ms. The response window was presented for two seconds following the playing of the third target. If the participant did not respond within the two-second window, the trial was replayed later. The inter-trial interval was one second.

In both studies, the participants were presented with six unique triads and six repetitions per contrast type, equaling a total of 36 triads per contrast for each listener. There were a total of 12 contrasts (all combinations of the four consonants /t ʈ ȶ/ as there might be differences in discriminability depending on the order of presentation) in the /aCa/ context (432 trials total per listener) and six contrasts (all combinations of the three consonants /t ʈ /, again allowing for differences in discriminability due to order of presentation) in the /##Ca/ context (216 trials total per listener). The participants in each study first completed the discrimination of the target consonants in medial position (the /aCa/ context (Study 1a, 2a), and then the discrimination of the target consonants in initial position (the /Ca/ context (Study 1b, 2b).

2.3 Participants

3.1 Native perception

The results of the /aCa/ discrimination task are presented in Fig 1. Chance performance is 50% correct discrimination. As is evident from Fig 1, the Wubuy speakers were indeed able to discriminate the four coronal stops in their native language (confirmed using one-sample Wilcoxon signed rank tests against chance performance with p values < .05 for all contrasts). The mean discriminability of the six contrasts in the medial context was 81% correct. One-sample Wilcoxon signed rank tests against chance performance also confirmed that the Wubuy speakers were able to discriminate the three [—sharp] contrasts tested in the /##Ca/ context (see Fig 1), with p values < .05 for each contrast. The mean discriminability of the three contrasts in the /##Ca/ context was 69% (range 63% correct for / ʈ/ to 72% correct discrimination for /t ). This suggests that discrimination is more difficult in initial position and a related-samples Wilcoxon signed rank test of the distribution of scores in the /aCa/ and /##Ca/ contexts confirmed this with a significant effect of context (p = .037). There was no significant effect of contrast, which is unsurprising as we did not predict an overall loss or gain in discriminability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Native discrimination scores per contrast type in the /aCa/ and /##Ca/ contexts.

% correct discrimination is presented on the y-axis, while contrast type is presented on the x-axis. Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g001

However, as laid out in H2b and H2c above, we predicted that /t ʈ/ would be less accurately discriminated than / ʈ/ and /t /, due in the case of / ʈ/ to maximal acoustic differentiation, and in the case of /t / to the fact that /t/ and // are more frequent in Wubuy input than /ʈ/. A related-samples Friedman's analysis of variance by ranks comparing the distribution of the averaged scores in the two contexts indicated that there was indeed an effect of contrast (p = .009), and Dunn-Bonferroni post-hoc comparison of the differences between the discrimination accuracy for the three averaged contrasts confirmed that the /t ʈ/ contrast was indeed discriminated less accurately than the / ʈ/ as predicted. The /t ʈ/ discrimination accuracy did not however differ significantly from that of /t /. We suspect that this is due to the poor acoustic differentiation of /t / offering little information for the listeners on which to base their discrimination [42]. In terms of an effect of context on the discrimination accuracy of individual contrasts, we expected that the discrimination accuracy of the acoustically similar /t ʈ/ would decrease from the /aCa/ context to the /##Ca/ context due to the infrequency of /ʈ/ in initial position in Wubuy input (H2d). A related-samples Wilcoxon match-pair signed-rank test did not support our hypothesis (p = .33), though the difference between the two contexts is in the predicted direction.

Interestingly, however, a related-samples Friedman's analysis of variance by ranks of each condition (/aCa/ and /##Ca/) separately showed that the discrimination performance of the participants varied with the contrast in question in the /aCa/ context, with correct discrimination ranging from 88% correct for /ʈ ȶ/, to as low as 70% for /t ʈ/ contrast (medial position p = .045; initial position p = .139, ns).

Subsequent Dunn-Bonferroni post-hoc comparison of the differences in the /aCa/ context showed that the only significant difference was in the discrimination accuracy of the apical contrast /t ʈ/ which was less accurate than discrimination of either of the apicals with the lamino-alveopalatal. This clearly suggests that the apical contrast is the most difficult to discriminate even for native speakers, and even in medial position where it is better supported by acoustic cues on both sides, i.e., during consonant closure as well as release.

3.2 Non-native perception

The results for the non-native English listeners’ /aCa/ discrimination are presented in Fig 2. As before, chance performance is 50% correct discrimination. As the figure indicates, the English listeners were able to discriminate most of the coronal stops in Wubuy (% correct discrimination ranged from non-significant discrimination of 49% for / t/ to 93% for /t ȶ/, with all three contrasts involving /ȶ/ being discriminated correctly in more than 90% of cases). One-sample Wilcoxon signed rank tests against chance performance confirmed that the English listeners were indeed able to discriminate the Wubuy stops (p = < 0.05 for /t ʈ/, /ʈ /, /t ȶ/, / ȶ/, and /ʈ ȶ/), with the exception of / t/ which was not significant in the /aCa/ condition. One-sample Wilcoxon signed rank tests against chance performance also confirmed that the English listeners performed above chance (p < 0.05 for all three contrasts) in the /##Ca/ condition. Note that the discrimination performance for / t/, which had not been discriminated in /aCa/ context, was above chance in /##Ca/ context.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Non-native discrimination scores per contrast type in the /aCa/ and /##Ca/ contexts.

% correct discrimination is presented on the y-axis, while contrast type is presented on the x-axis. Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g002

As in the case of the Wubuy listeners, a related-samples Wilcoxon signed rank test of the distribution of scores in the /aCa/ and /##Ca/ contexts showed a significant main effect of context (p = .013), which we did not predict. However, as opposed to the Wubuy listeners, the English listeners found /##Ca/ discrimination to be easier than the /aCa/ (confirmed by an independent samples Mann-Whitney U test, p < .001). This likely reflects an increased familiarity with the Wubuy stops (and perhaps the task) as the /##Ca/ study was always presented second. As predicted, there was no significant effect of contrast.

A related-samples Friedman's analysis of variance by ranks showed that there was a significant effect of contrast type in both conditions separately (/aCa/ p = .000; /##Ca/ p = .006). Subsequent Dunn-Bonferroni post-hoc comparisons showed that, in the /aCa/ context the /t / contrast and the apical contrast /t ʈ/ were less accurately discriminated than all contrasts involving a lamino-alveopalatal (all significant at or below p = .005). Post-hoc comparisons of the results in the /##Ca/ context showed that the /t / contrast was significantly more poorly discriminated than / ʈ/ (p = .012) and /t ʈ/ (p = .032). There was no significant difference in the discriminability of the apical contrast /t ʈ/ and the dental-retroflex contrast / ʈ/.

3.3 Native versus non-native perception

In order to assess what similarities and/or differences exist in the performance of the two groups with regard to each of the contrasts in each presentation environment (see Fig 3 for the /aCa/ context, and Fig 4 for the /##Ca/ context), we conducted six independent-samples Mann-Whitney U tests. Table 4 shows that the two listener groups differed in their discrimination of medial /t / (where the Wubuy listeners out-performed the English listeners), and in their discrimination of initial /t ʈ/ (where the English listeners out-performed the Wubuy listeners).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Comparison of native and non-native discrimination performance in the /aCa/ context.

% correct discrimination is presented on the y-axis, while contrast type is presented on the x-axis. Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Comparison of native and non-native discrimination performance in the /##Ca/ context.

% correct discrimination is presented on the y-axis, while contrast type is presented on the x-axis. Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Pairwise comparison, Wubuy versus English discrimination in /aCa/ and /##Ca/, excluding contrasts involving the lamino-alveopalatal.

Significant results indicated by * and Sig. in bold.

https://doi.org/10.1371/journal.pone.0142054.t004

3.4 Discrimination asymmetries in native and non-native listeners

In order to test the Natural Referent Consonant hypothesis outlined in the introduction, and its predictions in H2c and H4 that discrimination accuracy may differ in ways specific to each listener group depending on the order of presentation, we conducted one further set of analyses of our discrimination results for the contrasts /t ʈ/, / ʈ/ and /t /. As our participants were presented with a well-balanced stimulus set, where all consonants appeared first and in combination with all other consonants, we were able to split the discrimination results from each contrast into two separate sub-analyses for each contrast: one in which each member of the contrast was presented first, and one where it was presented second. The sub-divided results are presented in Table 5 (Wubuy) and Table 6 (English) below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Wubuy discrimination accuracy depending on the first presented consonant for /t ʈ/, / ʈ/ and /t / in both the /aCa/ (first mean % listed) and /##Ca/ (second mean %) contexts.

https://doi.org/10.1371/journal.pone.0142054.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. English discrimination accuracy depending on the first presented consonant for /t ʈ/, / ʈ/ and /t / in both the /aCa/ and /##Ca/ contexts.

https://doi.org/10.1371/journal.pone.0142054.t006

If our NRC hypothesis predictions are correct, we expect the following patterns:

1. Wubuy listeners will show an asymmetry in discriminating the infrequent /ʈ/ from both /t/ and //, such that discrimination will be more accurate when the infrequent consonant is presented first (and thus twice) in each triad (H2c).
2. Wubuy listeners will display no asymmetry when they are discriminating the two frequently-occurring coronals // and /t/ (H2c).
3. English listeners will show an asymmetry between /t / and /t ʈ/ such that discrimination is easier when the ‘odd’ consonant relative to English (// or /ʈ/) rather than the native (and more canonical) /t/ is presented first. They are not, unlike Wubuy listeners, expected to show any asymmetry for the contrast / ʈ/, because both phones are 'odd' in this case (H4). We qualify this with the fact that we are not basing our predictions on a categorisation task, as we have not carried out such a study. It is possible that the English participants’ perception of the Wubuy stops differs from what we are proposing.

The discrimination asymmetries for the three contrasts (and an additional measure of /t/ versus a collapsed category 'x' which is an average of the // and /ʈ/). are presented in Figs 5 and 6, below. Asymmetry values were computed by subtracting the paired percentage correct values for the two presentation orders from each other.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Discrimination asymmetries for Wubuy listeners.

Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Discrimination accuracy for English listeners.

Error bars indicate standard error of the mean. Note, however, that as the statistics applied are non-parametric, they are not necessarily related to the dispersion indicated in the error bars.

https://doi.org/10.1371/journal.pone.0142054.g006

We tested for an overall effect of presentation order for native and non-native listeners separately and found a significant effect in both listener groups (asymmetry values averaged over contrast and context and submitted to Wilcoxon signed-rank test against the hypothesised median value of 0; Wubuy listeners: p = .005; English listeners: p = .018). We then subjected the individual results for each context-contrast cell to a series of One-sample Wilcoxon signed-rank tests to determine whether there were any differences, for each group, between the performance on each contrast pair according to direction of presentation. We also compared the two groups for the difference in their performance on contrast-by-presentation order. Neither of these results differed significantly from chance. As the descriptive statistics make clear, this is probably due to the unavoidably low number of participants (approximately 25% of the entire Wubuy speech community) resulting in a loss of statistical power, and the great degree of variance in both groups, for all contrasts. Indeed, the only result which reached statistical significance (and this on a one-sample t-test against 0) was that of the contrast / ʈ/ between English and Wubuy listeners, which is the largest difference in means in the entire sample. We discuss this issue further in Section 4.

Nevertheless, the differences in the means alone (shown in Tables 5 and 6) are consistent with a difference between contrasts according to presentation order, and according to group, as predicted. We therefore subjected the results to further one-sample Wilcoxon signed rank tests against the hypothetical value of 0 (no difference between presentation orders). P-values are presented in Tables 7 and 8. As is clear, the Wubuy listeners show the expected direction of perceptual asymmetry for the contrasts /t ʈ/ and / ʈ/, consistent with the low frequency of occurrence of /ʈ/ in the speech input.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Results of one-sample Wilcoxon signed rank tests of the perceptual asymmetries of Wubuy listeners.

Significant results in bold.

https://doi.org/10.1371/journal.pone.0142054.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Results of one-sample Wilcoxon signed rank tests of the perceptual asymmetries of English listeners.

Significant results in bold.

https://doi.org/10.1371/journal.pone.0142054.t008

As a visual inspection of Fig 4B suggests, the English listeners also showed patterns of perceptual asymmetry. Indeed, statistical analyses are reasonably consistent with our predictions for two of the three contrasts. The English speakers' discrimination is asymmetrical for /t / in the /aCa/ condition, consistent with the fact that [] is a rare allophone of /t/ in English (produced in anticipatory coarticulation of a following /θ/ as in width), and does not contrast phonologically with [t] in the English input. However, the asymmetry found for this contrast in /##Ca/ did not reach significance, although the trend is the same. We also find support for the predicted null asymmetry for / ʈ/. However, in neither context did the asymmetries in the perception of /t ʈ/ reach significance, although again the trend is in the predicted direction (with greater ease of discrimination when the ‘odd’ member of the pair is presented first).

4.1 Native language attunement and phoneme frequency shapes perception of coronal stops

4.2 The Natural Referent Consonant Hypothesis

Our analysis of the asymmetries in the discrimination patterns of the native and non-native speakers is consistent with the predictions of the NRC hypothesis that the three major articulators (lips, tongue front, tongue back) may provide perceptual and articulatory ‘anchor points’, corresponding to the three major oral places of articulation: labial, coronal, and dorsal. Our findings are also consistent with the NRC hypothesis' fundamental assumptions that 1) the most canonical articulation will be the perceptual anchor in consonant discrimination; 2) the canonical articulation for the coronal place is apical orientation of the tongue tip, with contact made in the alveolar region; and 3) consonant contrasts may continue to display an innate bias towards the Natural Referent Consonant if the canonical articulation is also the more frequent member of the contrast.

We hope our framework and the studies presented here invite more research to test the NRC and its predictions. Indeed, the perceptual asymmetries observed here must be accounted for, and the NRC provides the only current testable framework. We recognise, however, that while our results are consistent with the NRC hypothesis, we cannot conclusively determine whether the discrimination behaviour of the two listener groups in our studies is due to language experience alone, or if the Natural Referent Consonant (or, perhaps more precisely, Natural Referent Place of Articulation) principles play a role in the discrimination patterns observed. In the case of the English listeners, the NRC hypothesis predicts a pattern of behaviour identical to that predicted by PAM on the basis of language specific learning of coronal consonants: /t/ would be expected to be canonical in this group, resulting in the observed asymmetries, both on the basis of language specific learning of English, and on the basis of the alveolar place of articulation being the canonical or default setting. This pattern is exactly what we observed: canonical /t/ behaves as a perceptual anchor for English speakers in the discrimination of /t/ versus // and /ʈ/.

In the case of the Wubuy listeners, we observed that the discrimination of frequently-occurring native phonological categories is symmetrical, even for coronal stops with a POA other than alveolar, e.g. the lamino-dental stop. This finding is compatible with the findings reported in [12, 13] that native language acquisition can override the initial Natural Referent Vowel effects of better discrimination when a less peripheral vowel is presented prior to presenting a more peripheral vowel, observed in infant speech perception and in cross-language/L2 perception. The fact that /ȶ/ does not act like an anchor, despite its frequency of occurrence, is likely due to the fact that it is not /t/-like, as opposed to // and /ʈ/, and easily discriminated (despite the asymmetry) by native and non-native listeners alike (see also [9]). Our results, however, also showed that perceptual asymmetries do exist even for native speakers in the special situation when one of two phonemes in a contrast is infrequent in the input. This has not been explicitly considered for native vowel perception, nor been examined experimentally, but may well apply to vowels also. In any case, our findings support the notion of a NRC bias that will only disappear if native language experience provides sufficient input for the relevant phonological contrasts to be established and reinforced. And this is arguably not the case for Wubuy /ʈ/ which is exceedingly rare in natural Wubuy input. To conclusively tease apart language specific attunement factors and a potential NRC in the context of the present study, however, we would have to present our two tasks to speakers of a language with a dental and a retroflex stop, but no alveolar stop. Here, we would expect native language attunement to override any NRC effects so that /ʈ/ followed by /t/ and // followed by /t/ would always be discriminated at the same level of success as /t/ followed by /ʈ/ and /t/ followed by //. Hindi, as well as other Indo-Iranian and Dravidian languages of South Asia, have these characteristics and we look forward to future studies of these L1 listeners’ perception of Wubuy coronal stops.