1.1. Marginal contrast

Contrast is a central tenet of phonology. Traditionally, two sounds are said to be in a relation of contrast in a given language if they can be used to distinguish minimal pairs in the lexicon. Sounds that cannot be used to distinguish minimal pairs and whose presence is predictable from the phonological context in which they appear are said to be allophonic. But it is not always straightforward whether two sounds in a given language are in a relation of contrast or allophony. Recently, there has been a surge in interest in articulating contrast in a more fine-grained, gradient way (i.a., Ernestus, 2011; D. C. Hall & Hall, 2016; K. C. Hall, 2013; Renwick, 2014; Renwick & Ladd, 2016; Scobbie & Stuart-Smith, 2008; Stevenson & Zamuner, 2017). In particular, the concept of marginal contrast¹ has garnered a substantial amount of attention. Phonological contrasts are considered marginal when their distributions lie somewhere between completely separate and totally overlapping (full allophony and full contrast respectively), a scale that can be operationalized probabilistically (K. C. Hall, 2009). Such contrasts are quite common. For example, Scobbie and Stuart-Smith (2008) discuss the phonemic status of /ʍ/ and /x/ in Scottish English. These sounds form minimal pairs (with /w/ and /k/ respectively, e.g., /wɪt͡ʃ/ – witch ~ /ʍɪt͡ʃ/ – which or /lɔx/ – loch ~ /lɔk/ – lock) for many, but not all speakers, and in some, but not all phonological contexts (e.g., only at the beginning of words for the former, but only at the end for the latter). Such minimal pairs, in addition to being subject to inter-individual variation and contextual constraints, are also relatively infrequent. This variability is measurably different from other phonological contrasts in the system (e.g., /p/~/k/) which serve to make many meaningful lexical distinctions in all contexts for all speakers (e.g., pill versus kill or lip versus lick). Similarly, contrasts that neutralize in certain contexts may be viewed as marginal. For example, in German (and many other languages), laryngeal contrasts on obstruents are neutralized in word-final position. While such contrasts are clearly part of the phonological system (they make many distinctions at the beginnings of words for all speakers, e.g., /dɔrf/ – Dorf – “village” ~ /tɔrf/ – Torf – “peat”), their lack of contrast in word-final position means that their distribution is, like the Scottish cases, partly predictable. This semi-predictability of their distribution qualifies these contrasts as ‘marginal,’ and the longstanding debate on the incomplete phonetic realization of this neutralization and its implications for representation only add to the complexity of this case (see Winter & Roettger, 2011, and references therein).

K. C. Hall (2013) laid out the first typology of such marginal relations, identifying the primary ways that can lead the distribution of contrasts to be in between full contrast and allophony. One primary way of defining contrasts as marginal is based on their distribution, such that frequency differences, phonotactic restrictions, and neutralizing phonological rule application (among others) can render contrasts marginal. But marginal contrasts can also come about through mergers. For example, in Parisian French, there is an ongoing merger of mid-vowels, rendering variable the difference between words like /epe/ – épée – “sabre” and /epɛ/ – épais – “thick” (Fagyal, Hassa, & Ngom, 2002). Over time, such mergers may make previously full contrasts somewhat or entirely allophonic. In the case of the French mid-vowels, close-mid and open-mid variants are entirely predictable based on syllable structure in varieties where the merger is complete.²

While contrasts may become (more) marginal through mergers, marginal contrasts may also come about in cases of intensive language contact. Such cases are only touched upon briefly in K. C. Hall’s (2013) typology, but represent an important piece of the marginal contrast puzzle. When languages are in contact, lexical borrowing is one of the earliest consequences (Thomason & Kaufman, 1988). Such borrowing leads to an interesting phonological situation when words that are brought in contain sounds that do not form contrasts in the borrowing language. These sounds are usually adapted to the phonology of the borrowing language (Winford, 2005). For example, in the English loanword backpacker, Dutch speakers pronounce the English /æ/, absent from the Dutch inventory, with the nearest native Dutch category, /ɛ/. Although theories of loanword adaptation differ in whether they consider this process to take place at the phonetic or phonological level (see Kang, 2011; Uffmann, 2015, for reviews), they do cover most of what happens in lexical borrowing: Foreign sounds are adapted to native ones through some mapping function.

But words that are borrowed are not always (fully) adapted to the phonology of the borrowing language (Blevins, 2009; Boretzky, 1991; Cohen, 2019; Eckman & Iverson, 2015; Haugen, 1950; Itô & Mester, 1999, 2001; Kennard & Lahiri, 2020; Lee, 2013; Ussishkin & Wedel, 2003; Zuraw, O’Flynn, & Ward, 2019, among others). This can lead to the emergence of a marginal contrast in the phonology of the borrowing language. Lee (2013), for example, described how the originally context-dependent allophone [ʃ] in Korean now forms minimal pairs with native /s/. Before heavy borrowing, /s/ was produced as [s] in the context of non-high vowels and as [ʃ] in the context of high vowels (a case of full allophony). Through the introduction of words like show from English (which contains [ʃ] followed by a non-high vowel), the two sounds became contrastive for some word pairs. This is a situation that Hall terms “mostly predictable, but with a few contrasts” (K. C. Hall, 2013, p. 231).

One approach to understanding this type of phenomenon considers the lexicon as a hierarchical set of strata. For example, Itô and Mester’s (1999, 2001) approach organizes the lexicon into strata that may be closer or further away from the native core, based on the grammatical constraints that operate on different kinds of lexical items. As the strata are hierarchically organized, certain grammatical constraints may apply to a certain subset of the lexicon while not affecting the lexicon as a whole. Cases of non-adaptation, like the emergence of /ʃ/ in words like show in Korean, are then treated as more peripheral than “assimilated foreign items”³ since they may contain segments or structures that are illicit in more core strata of the lexicon. The alternation between [s] and [ʃ] does not become completely inactive for Korean speakers when words like show are borrowed; rather, the alternation applies only to a subset of the lexicon, namely core native words, while the new contrast applies only for loanwords (for more discussion on the intermediate steps, see Simonović, 2015).

The remainder of this paper will address the psychological reality of marginal contrast. How is a contrast that is defined as marginal according to the aforementioned distributional criteria represented in speakers’ minds and what effects does that have in speech production?

1.2. Cognitive constraints on contrast emergence

A stratal analysis as described above can account for observed distributional differences. This, however, is only one part of the story. What can perception tell us about the representation of phonological contrast? Does perception strictly track lexical organization, with more marginal contrasts being perceived less well than more core ones? Loanword non-adaptation is an ideal topic to explore such questions. Here, we will consider just such a case, where a new contrast is emerging after intensive lexical borrowing, causing two sounds which were previously completely allophonic to form contrasts between native and borrowed words. We propose that the evolution of such marginal contrasts may be dependent on the link between speakers’ ability to produce and perceive the borrowed sound relative to native categories.

Evidence from perceptual studies clearly shows that listeners are sensitive to whether a pair of sounds is contrastive in their language or not (Babel & Johnson, 2010; Boomershine, Hall, Hume, & Johnson, 2008; Hume & Johnson, 2001; Whalen, Best, & Irwin, 1997, among many others). Boomershine et al. (2008), for example, demonstrated how sounds that are in a relationship of allophony are perceived as more similar to each other than sounds that are contrastive. They tested the perceived similarity of the sounds [d], [ð], and [ɾ] in populations whose native language uses these sounds differently. Spanish speakers, for whom [d] and [ð] are allophonic, perceived these two sounds to be highly similar, but found [d] and [ɾ], which correspond to different phonemes of Spanish, to be relatively dissimilar. English speakers, on the other hand, for whom [d] and [ð] correspond to two different English phonemes, perceived these two sounds to be relatively dissimilar, while [d] and [ɾ], which are both allophones of English /d/, were perceived as much more similar. Thus, the status of a contrast has clear effects on perception.

Beyond the allophonic versus contrastive dichotomy, listeners have been shown to be sensitive to the level of distinctiveness of different contrasts, even if those contrasts would all be considered part of the core native phonology (i.e., contrastive, not allophonic). For instance, French listeners have been shown to be better at recognizing nouns whose initial consonant is mispronounced in voicing than if the mispronunciation involves manner or place of articulation (e.g., listeners can recognize the non-word pagnole as the real word bagnole – “car” [voicing difference], but have a much harder time with the non-words dagnole [place difference] or vagnole [manner difference]), indicating that manner and place of articulation are somehow more important for lexical access in French (Martin & Peperkamp, 2015). A follow-up study showed that the functional load of these contrasts could explain part of this pattern (Martin & Peperkamp, 2017). In a similar vein, K.C. Hall and Hume (2013) focussed on how listeners’ impressions of similarity vary according to the functional load of the contrast in question. They found that contrasts with a lower functional load (i.e., those that are used to distinguish fewer pairs of words in the lexicon) were perceived as more similar than contrasts with higher functional loads. An entropy-based translation of predictability of distribution in different phonological environments has also been shown to be predictive of listeners’ perceptions (K. C. Hall, Letawsky, Turner, Allen, & McMullin, 2015). That study tested the discriminability of the English fricatives [f], [s], [ʃ], and [h] in onset, intervocalic, and coda positions and in the context of different vowels. They found that sound pairs with a less predictable distribution in English were better discriminated than those with a more predictable distribution. Thus, clearly perception is reflective of the gradient nature of phonological contrast.

The studies described above focus on native contrasts, but perception may tell us even more about marginal contrasts that arise through borrowing. We will refer to this process of borrowing that results in a change to the phonology of the borrowing language as non-adaptation. Perception has been proposed as a key factor in determining the way words are adapted into the borrowing language (i.a., Boersma & Hamann, 2009; Davidson, 2007; de Jong & Cho, 2012; Kang, 2003; Peperkamp & Dupoux, 2003; Peperkamp, Vendelin, & Nakamura, 2008). Peperkamp and Dupoux (2003), for example, argue that ‘phonological deafness’ drives certain patterns of loanword adaptation. According to this account, French listeners cannot perceive lexical stress contrasts (since they are not present in French) and thus adapt loanwords according to how they do perceive them (namely, without lexical stress). Thus, even if French were to borrow very extensively from a language like English, it is unlikely that minimal pairs based on a stress contrast would come to exist in the language, simply because French listeners cannot perceive the stress contrasts to begin with.

But what of phonetic contrasts that exist in a borrowing language, but do not distinguish word pairs (i.e., allophones)? It has been shown that allophones may actually be easier for listeners to perceive and produce if they are presented outside of their licensing environments (Peperkamp, Pettinato, & Dupoux, 2003; Whalen et al., 1997). For instance, the French phoneme /r/ is realized as [𝜒] when it is adjacent to a voiceless consonant and otherwise as [ʁ]. Thus, French speakers have experience with the production and perception of both sounds. Interestingly, Peperkamp et al. (2003) found that while French listeners have difficulty perceiving the allophonic [ʁ]~[𝜒] contrast when the sounds precede another consonant (precisely where they alternate in the language), this difficulty disappears when they occur in final position. This could be an important factor in cases of non-adaptation: Sounds that are present allophonically in a borrowing language may be more likely to be borrowed and lead to a marginal contrast because listeners are better able to perceive them than sounds that are completely absent from their language (though perhaps only outside of their traditional licensing environment). It might be possible, then, for a contrast between [ʁ] and [𝜒] to emerge through borrowing in French, but this might be restricted (at least initially) to contexts where the two sounds do not alternate allophonically.

Another factor that may play a role in loanword (non-)adaptation is the abstract representation of languages’ sound inventories. A non-adaptation that leads to a contrast that recombines features already present in the system (i.e., filling a gap in the inventory) adheres to the principle of feature economy (Clements, 2003). And indeed, if inventories tend to use featural distinctions maximally, non-adaptation might be facilitated in cases where a foreign sound recombines features (or gestures) already present in the borrowing language (Cohen, 2019; Ussishkin & Wedel, 2003). In this paper, we will consider a case study of a marginal contrast that is emerging in Dutch through heavy lexical borrowing and non-adaptation, involving two previously allophonic sounds whose representation recombines phonological features already active in the Dutch inventory. We return to the role of allophonic perception and feature economy in contrast emergence below.

1.3. Emergence of a marginal contrast in Dutch

Our case study concerns the emergence of a marginal contrast in Dutch, a West Germanic language spoken by some 22 million people mainly in the Netherlands and the Flanders region of Belgium (Lewis, Simons, & Fenning, 2016). Dutch, like languages all around the world, has borrowed extensively from other languages, including English, French, Italian, Indonesian, and many others. For example, in the 20th century alone, over 1,300 words were borrowed into Dutch from English (van der Sijs, 2002). A quick analysis of the complete loanword dictionary compiled by van der Sijs shows that these words come mostly from the domains of communication (e.g., information and communications technology, media), social life, sports, consumption, and science, and much more rarely concern domains like religion, geography, or time. Simply put, words borrowed in the last century pertain to our modern lifestyles. Many of these words contain sounds that do not exist natively in Dutch. Our specific interest here will be the sound /ɡ/ such as occurs in words like gate (borrowed from English) or spaghetti (borrowed from Italian).

The native Dutch stop inventory includes a voicing contrast between prevoiced and voiceless stops at both the labial and coronal places of articulation, creating minimal pairs like paard – “horse” ~ baard – “beard” and tak – “branch” ~ dak – “roof.” The voicing contrast at the velar place of articulation, however, was lost, with West Germanic */ɡ/ spirantizing to a fricative in Old Dutch (Goossens, 1974), leaving voiceless /k/ without a voiced counterpart. The Germanic past tense prefix ge- is therefore produced as [𝜒ə-]⁴ in modern Dutch words like gedaan – “done.” The phone [ɡ] is not completely absent from Dutch, though. In regressive voicing assimilation, underlying /k/ may be rendered as [ɡ] before a voiced stop, such as in i[ɡ] ben – “I am” (Booij 1999). This assimilation also occurs across stem boundaries in compound words like plakband – “adhesive tape” formed from plak and band.⁵ Dutch speakers are therefore used to producing the sound in such contexts, and also to being exposed to it auditorily, with the mapping [ɡ] → /k/. Thus, although the phone [ɡ] is present, it does not contrast in the native lexicon with [k]. Rather, it represents one end of a continuum that maps onto the category /k/. Given that [ɡ] is both an allophone (thus potentially perceptible to Dutch speakers outside of its licensing environment) and that its representation recombines features already present in the inventory, it is a prime candidate for non-adaptation. What then happens when words containing this sound are borrowed into Dutch? Previous research has revealed several outcomes.

On the one hand, foreign /ɡ/ can be adapted and pronounced with a native Dutch sound. Since Dutch shares an orthographic system with most of the languages it borrows from (namely the Latin alphabet), one such possibility is for Dutch speakers to simply produce the sound represented by the grapheme <g>, namely [𝜒] (or the corresponding regional fricative). Indeed, orthographic strategies of adaptation are not uncommon, even in the presence of other strategies (Daland, Oh, & Kim, 2015; van Coetsem, 1988; Vendelin & Peperkamp, 2006). Alternatively, the sound can be adapted while still approximating the way it is pronounced in the source language (thus produced with the phonetically closest Dutch sound, [k]⁶). On the other hand, these loanwords may be produced with a new sound [ɡ], which contrasts in voicing with native [k], leading to a structural change in the phonology of Dutch. And indeed, through borrowing and non-adaptation, [ɡ] has come to form minimal pairs with its native voiceless counterpart [k]: kool – “cabbage” ~ goal – “goal (sports)”; manco – “deficiency” ~ mango – “mango”; keet – “shack” ~ gate – “gate (at an airport)”; lekken – “to leak” ~ laggen⁷ – “(for a computer) to lag.”

It is noteworthy that these minimal pairs follow the same phonetic pattern as the voicing contrast that distinguishes paard from baard and tak from dak: namely prevoicing versus short-lag VOT. Dutch voiced stops have been reported to have a mean VOT of –85 ms for labials and –80 ms for coronals, while the corresponding voiceless stops have VOTs of 10 ms and 15 ms, respectively (Lisker & Abramson, 1964). When Dutch speakers contrast the native word kool (mean VOT of 25 ms) with the loanword goal (producing prevoicing on the /ɡ/), they are implementing this new voicing contrast within the Dutch phonetic space. As we will see, this implementation is not sensitive to the phonetics of the voicing contrast of the source language. Indeed, in English, the initial /ɡ/ is often produced with short-lag VOT, thus very close to Dutch /k/. Minimal pairs like kool~goal highlight how this contrast has become meaningful in the Dutch phonological system, beyond reference to a phonetic contrast in a source language.

The production of this new phoneme /ɡ/ in Dutch has received some attention in three previous studies (Hamann & de Jonge, 2015; van Bezooijen & Gerritsen, 1994; Van de Velde & van Hout, 2002). In two of these, /ɡ/ was considered as part of greater surveys concerning the pronunciation of loanwords in Dutch, and specifically the realization of various sounds that are not distinctive natively in the language. First, van Bezooijen and Gerritsen (1994) examined the pronunciation of loanwords by asking 75 women from various regions of the Netherlands and Belgium to read them aloud. Two of the words contained /ɡ/: goal and drugs. The two differ though, as the latter is subject to final devoicing; the production of a voiced stop in such a case would thus involve an extra degree of non-nativeness. They found evidence for all three phonetic realizations of /ɡ/ for these words: [ɡ], [k], and [𝜒], though [ɡ] occurred most often, at least in goal (49% of tokens of this word). Van de Velde and van Hout (2002) surveyed a wider range of regional populations in both the Netherlands and Belgium. They asked school teachers from four regions in the Netherlands and four regions in Belgium to read a list of loanwords aloud according to standard Dutch pronunciation. Their list included a few more words with /ɡ/ (goal, guillotine, goulash, buggy, mango), and they found similar results to van Bezooijen and Gerritsen (1994), with unadapted [ɡ] being rather common, but with the other two variants also being used. The third study used an adaptation of the map task (Brown, Anderson, Shillcock, & Yule, 1985), which the authors called the ‘menu task,’ to elicit spontaneous productions of three food-related loanwords containing /ɡ/ (gorgonzola, spaghetti, and mango) by an age-stratified sample of dyads (Hamann & de Jonge, 2015). Again, there was clear variation in the pronunciation of the words. In addition, they noticed slight differences in the way the younger and older groups pronounced spaghetti. Younger people in their sample realized this word with a fricative (following the orthographic pronunciation) more often than older individuals. The authors suggest word frequency and word age as possible explanatory factors for their findings.

While clearly establishing that there is great variation in the (non-)adaptation of loanwords containing /ɡ/ and showing strong evidence for the use of the sound [ɡ] outside of its usual allophonic context, none of these studies considered listeners’ ability to perceive the emerging contrast. As discussed above, perception plays a key role in loanword (non-)adaptation. In the present case, it is therefore important to consider individuals’ ability to perceive the new sound [ɡ] as different from their native [k], and to test if this is related to the way they produce the critical loanwords. Are the speakers that produce the most [ɡ] tokens also the ones who can perceive the contrast between [k] and [ɡ] best? Previous work has shown that the relation between perception and production at an individual level may follow complex trajectories (e.g., Pinget, Kager, & Van de Velde, 2020), and consideration of perception alongside production is thus a central concern.

To address this issue, this paper details an experimental investigation of Dutch /ɡ/ usage effects in production and in perception, and explores the relation between the two. Focussing on university students, we assess the production and perception of /ɡ/ by means of an elicitation and a discrimination task respectively; this allows us to consider the relation between individual speakers’ productions and their ability to perceive this new contrast. To preview, we find that while there is a range of variability in production, participants’ perception of the emerging contrast is very good, though not quite as good as of a contrast that is established in Dutch. We also explore the idea that speakers who produce more [ɡ] tokens also tend to be better at perceiving the emerging [k]~[ɡ] contrast, but our data do not provide strong support for this.

3.1. Production

Participants’ productions were segmented and coded manually offline by the first author. Each target segment was identified by examination of the waveform and spectrogram in Praat (Boersma & Weenink, 2020). Tokens were coded as stops if they had a visible release burst, otherwise they were coded as fricatives. Stops were coded as voiced if they had a visible voicing bar during closure, and voiceless if they did not. Voiced stops were segmented from onset of voicing to release burst; voiceless stops were segmented from release burst to onset of voicing in the following vowel. VOT durations were extracted from this segmentation using a Python script and measured relative to the release burst, such that prevoiced tokens (those with a visible voicing bar during closure) had negative VOTs and voiceless tokens (those where voicing began after the release burst) had positive VOTs. Examples are shown in Figure 2.

Figure 2

Spectrograms of two tokens of the word goal produced in isolation, one produced with [k] (left) and one produced with [ɡ] (right). The token on the left shows that the VOT for the phone [k] was measured from the release burst to the onset of voicing in the following vowel. The token on the right shows that the VOT for the phone [ɡ] was measured from the beginning of visible prevoicing up to the release burst.

Occasionally, target segments were produced in an ambiguous way. Per our coding scheme, segments with no identifiable release burst were categorized as fricatives, though a certain number of these did not impressionistically correspond to any variant of the fricative representing the grapheme <g>. Lenition of underlyingly voiceless obstruents has been documented in spontaneous Dutch speech (Schuppler, Ernestus, Scharenborg, & Boves, 2011), and these tokens could, in principle, correspond to underlying /ɡ/ or /k/. Since it was not possible to objectively determine which category these ambiguous tokens belonged to, they were not included in our analysis; this concerned around 11% of the data points (N = 148). A total of 1,264 tokens were thus retained for analysis, 617 from the sentences and 647 from the word list, with a total of 700 [ɡ] tokens, 436 [k] tokens, and 128 [𝜒] tokens.

We observed a wide range of inter-individual variability in the proportion of [ɡ] produced (see Figure 3), with some participants nearly always producing [ɡ], and others rarely producing it. We analyzed these data using logistic mixed-effects models,¹² with fixed effects being deviation coded. We first created a full model that included the fixed effect of Context (word list versus sentence reading) with random intercepts for Participant and Item, each including a random slope for Context (this was the maximal model according to our design). We then created a simpler model that excluded the factor Context and compared the simpler model to the full model using a likelihood ratio test. The full model was found to explain significantly more variance than the reduced model (β = –0.81, 95% CI = [–1.53 : –0.12]; 𝜒²(1) = 5.13, p < 0.05). This indicates that participants were more likely to produce [ɡ] if the token was produced within a sentence than if it was produced in isolation.

Figure 3

Average percentage [ɡ] production by context. Each dot represents a participant and error bars show a 95% confidence interval calculated over participant means.

In addition to whether or not each token was produced in isolation or embedded in a sentence, some target segments appeared in an intervocalic position (potentially easing the articulation of a voiced compared to a voiceless stop). Some of these always appeared in intervocalic position (e.g., when they were word internal like in the word spaghetti), while others were in intervocalic position only in the sentence reading (e.g., when they were preceded by a vowel-final word like in the phrase redelijke goulashsoep). A summary of the frequency of each token type by phonological context is shown in Table 1. To test the influence of phonological context on [ɡ] versus other production, we designed a logistic mixed-effects model with the fixed factor Intervocalic (deviation coded) and random intercepts for Participant and Item (including a random slope for Intervocalic under Participant, the maximal structure), and compared it to a simpler model excluding the factor Intervocalic using a likelihood ratio test. We combined the sentence and word list data in order to increase statistical power¹³ and found that the full model was a significantly better fit than the reduced model (β = 0.89, 95% CI = [0.39 : 1.40], 𝜒²(1) = 12.12, p < 0.001), indicating that intervocalic contexts did appear to facilitate the production of [ɡ].

Next, we considered the VOTs of stop tokens. VOT was measured as the beginning of a visible voicing bar relative to the stop release burst. VOTs for all stop tokens (i.e., both [ɡ] and [k] tokens) are visualized in Figure 4. Tokens with a positive VOT correspond to those tokens coded as [k], with a peak around the average VOT for native Dutch /k/ (25 ms, per Lisker & Abramson, 1964). Our [k] versus [ɡ] coding thus seems to properly reflect the bimodal distribution of the stop tokens’ VOTs in our dataset, and further follows the pattern reported in Lisker and Abramson (1964) with the voiced category showing a wider range than the voiceless category. To test whether there is a relation between the amount of [ɡ] usage (a representational question) and the phonetic realization of the sound (an implementational question), we measured VOTs of the [ɡ] stop tokens in individual participants and compared this with their average [ɡ] usage. A total of 700 tokens (again, combining the sentence and word list data) were coded as containing [ɡ] and were therefore eligible for VOT analysis. We examined whether speakers who produce longer prevoicing on [ɡ] tokens (thus a heightened contrast between native [k] and emerging [ɡ]) are more likely to use [ɡ] in general. We did this by performing a regression on the proportion of average [ɡ] production per participant as a function of their average [ɡ] VOT. This regression is shown in Figure 5. We indeed found that participants with longer average prevoicing on [ɡ] tokens also produced more [ɡ] tokens (adjusted R² = 0.12, t = –2.82, p < 0.01).

Figure 4

Distribution of VOT for all velar stop tokens in our production task.

Figure 5

Participants’ average [ɡ] production as a function of their average VOT for [ɡ] tokens with a fitted regression line and the 95% CI around that line.

3.2. Perception

Having considered our results in production where we observed varying rates of [ɡ] usage by participant, we turn to the results of our perception task to see if we observe similar variability. Despite the seemingly difficult design of the experiment (three different voices pronouncing long stimuli, thus a wide range of acoustics; crucial contrast in an unstressed syllable containing schwa), performance was overall very high and variability relatively low. We set an a priori inclusion criterion such that participants who performed below 80% accuracy on the established contrast were excluded from data analysis (N = 5). This concerned three participants who responded with only one of the buttons, one participant with low overall accuracy (70.1%), and one participant whose accuracy indicated they may have inversed the response buttons (27.2%). Average accuracy for the remaining 46 participants on the established contrast was 94.5% (SD = 22.8%), and 91.3% (SD = 28.2%) on the emerging contrast. Participants’ performance in the two conditions is plotted in Figure 6.

Figure 6

Participants’ average accuracy in the ABX task by contrast, plotted from chance level. Each dot represents a participant mean and error bars represent 95% confidence intervals calculated on those means.

We analyzed these data using logistic mixed-effects models, with fixed effects being deviation coded. We first created a full model that included the fixed effect of Contrast (established versus emerging) with random intercepts for Participant and Frame, each including a random slope for Contrast (this was the maximal model according to our design). We then created a simpler model that excluded the factor Contrast and compared it to the full model using a likelihood ratio test. This model was found to predict significantly less variance than the full model that included the factor Contrast (β = –0.41, 95% CI = [–0.81 : –0.01], 𝜒²(1) = 3.99, p < 0.05), indicating that the established contrast was perceived better than the emerging contrast.

We additionally conducted a follow-up analysis based on Best, McRoberts, and Goodell’s (2001) consideration of recency.¹⁴ The ABX task includes two types of trials: 1) those where X matches the initial stimulus (i.e., ABA trials) and 2) those where X matches the most recent stimulus (i.e., ABB trials). Previous work has suggested that participants in the ABX task tend to have a small bias to answer X = B (Schouten, Gerrits, & van Hessen, 2003). One potential explanation for this bias is that if participants have difficulty discriminating the critical sounds, then all three stimuli sound similar. This might lead participants to think that the last two stimuli in particular sound the same, if they have lost the auditory trace of the A stimulus. This would cause them to answer X = B more often than X = A, leading to ceiling performance on ABB trials, and lower accuracy on ABA trials. We therefore conducted a post hoc analysis splitting our data by trial type. This split is visualized in Figure 7, showing that the difference between the established and emerging contrast is greater on ABA trials than it is on ABB trials, though performance is overall still very high and variability low.

Figure 7

Participants’ average accuracy in the ABX task by contrast, separated by trial type, plotted from chance level.

To explore the possible difference between ABA and ABB trials, we re-analyzed the data including an additional deviation coded factor in our logistic regression model. Thus, we first created a full model that included the fixed effect of Contrast (established versus emerging) and a fixed effect of Trial Type (ABA versus ABB) and the interaction between these factors, along with random intercepts for Participant and Frame, and including a random slope for Contrast within Participant. This was the maximal random effects structure that converged. We then created simpler models that excluded one of the factors or the interaction and compared them to the full model using likelihood ratio tests. A model excluding Contrast was again found to predict significantly less variance than the full model (β = 0.40, 95% CI = [–0.65 : –0.16], 𝜒²(1) = 9.88, p < 0.01), indicating that the established contrast was perceived better than the emerging contrast. A model excluding Trial Type was also found to predict significantly less variance than the full model (β = 0.68, 95% CI = [0.52 : 0.84], 𝜒²(1) = 75.38, p < 0.001), indicating that performance was better on ABB trials than ABA trials, in line with a response bias for B. We also found that a model excluding the interaction predicted significantly less variance than the full model (β = 0.33, 95% CI = [0.02 : 0.64], 𝜒²(1) = 4.25, p < 0.05), indicating that the difference between the established and emerging contrasts was greater on ABA trials than on ABB trials.

3.3. Production-perception link

In order to investigate the association between participants’ production and their perception, we first established a by-participant score for each of the tasks. For the production task, we took the average proportion of [ɡ] use (unadapted compared to adapted variants), collapsing across production contexts (word list and sentence reading).

For the ABX task, we require a measure of how well each participant can discriminate the emerging contrast. Taking the raw accuracy on the emerging contrast fails to take into account task-related issues (e.g., attention). It is therefore necessary to consider performance on the emerging contrast relative to performance on the established contrast, which represents baseline performance in the task. A relative measure also makes comparison amongst participants more meaningful, since some listeners will be better at contrast perception generally. This is often done using a difference score, where participants’ average performances in one condition are subtracted from their average performance in the other. However, recent work has demonstrated that a method using the residuals extracted from a linear regression between the two conditions is a more robust measure (DeGutis, Wilmer, Mercado, & Cohan, 2013). We followed the method described in that study by performing a linear regression on the ABX accuracy data, using performance on the established contrast as a predictor for performance on the emerging contrast. From this regression, we extracted the residuals (a given point’s deviation from the fitted regression line), with lower values representing worse discrimination of the emerging [k]~[ɡ] contrast relative to the established [p]~[b] contrast. We then used these residuals in a regression predicting the average proportion of [ɡ] production by participant (that is, we used the by-participant perceptual score to predict the proportion of by-participant [ɡ] production); see Figure 8.

Figure 8

Average [ɡ] production by participant (both contexts combined) as a function of the residuals extracted from the ABX task regression with a fitted regression line and the 95% CI around that line.

The residuals from the ABX task regression were found to be a significant predictor of the proportion of [ɡ] production (adjusted R² = 0.12, t = 2.72, p < 0.01), indicating that participants who perceive the emerging contrast less robustly are less likely to produce [ɡ] in loanwords, and instead tend to adapt it as one of the native sounds [k] and [𝜒], while participants who perceive the emerging contrast more robustly are, on average, more likely to produce [ɡ]. We thus observed some evidence of a link between perception and production. Note though that the observed effect appears to be driven by a couple of participants, with most participants showing no discernible link between their production and perception scores, despite the relatively small amount of variability we observed in our perceptual results compared to our production results.

Given the results from the follow-up analysis of our ABX task, we additionally performed an analysis that used residuals from a regression that looked only at performance on ABA trials (the more difficult trials, where we observed a significantly stronger difference between the established and emerging contrasts than on ABB trials). The relation between ABA residuals and average [ɡ] production is visualized in Figure 9. We similarly found that the ABA residuals were a significant predictor of [ɡ] production (adjusted R² = 0.12, t = 2.66, p < 0.05) though again, this appears to be driven mainly by the same couple of participants.

Figure 9

Average [ɡ] production by participant (both contexts combined) as a function of the residuals extracted of the ABX task regression for ABA trials only, with a fitted regression line and the 95% CI around that line.

4.1. Production

Let us first consider our production data. Participants produced three types of tokens in loanwords which contain /ɡ/ in their source language. Some tokens were produced with the fricative [𝜒], which corresponds to the orthography: the Dutch letter <g> is pronounced as a fricative. Others were produced with stops, either the established voiceless [k] or the emerging voiced [ɡ]. We posited [ɡ] productions to represent non-adaptation while [k] tokens were posited to be adapted to the native phonology. But, before we settle down on this notion, we have to consider an alternative interpretation. Could these [k] tokens simply be faithful, English-like productions, at least for the tokens occurring in English borrowings? After all, the <g> in English goal is produced without prevoicing by native English speakers.¹⁶ However, this alternative interpretation is unlikely to fully account for our production data. While the speakers we tested were all proficient in English, considering [k] as a faithful production (thus not as an adaptation) requires speakers to have not only in-depth knowledge of English phonetics, but also in-depth knowledge of the phonetics of other source languages (e.g., how Italian native speakers produce spaghetti). Speakers would have to know to produce [k] for loanwords from English or German and [ɡ] for loanwords from French or Italian. Furthermore, if the target words were being produced as phonetically-faithful code-switches, we should expect no phonological integration at all (Poplack, Robillard, Dion, & Paolillo, 2020); clear and consistent final devoicing on items like Go Ahead Eagles in our data discounts this possibility. Finally, if such loanwords were to be produced entirely faithfully, then we should expect (some) Dutch speakers to produce words like computer with aspiration on the initial /k/, in line with the English pronunciation. But lack of aspiration on initial stops in stressed syllables is a well known feature of Dutch L2 speakers of English (Collins & Mees, 2003) and we are unaware of such productions on loanwords in Dutch. For these reasons, we considered [k] tokens to be productions adapted to the native phonology.

It seems that the phonetic realization of the crucial sound in the source language has a limited effect compared to its position in the phonological system of that language, and specifically how it contrasts with other sounds. This is in line with accounts of structural borrowing in cases of intensive language contact (like borrowing of a phoneme) that propose that such borrowing is initially restricted to a small group of highly bilingual individuals before spreading through the community (e.g., Thomason & Kaufman, 1988; Weinreich, 1953). The abundance of [ɡ] tokens we observed suggests that speakers know that /ɡ/ contrasts with their native /k/ in a specific way (i.e., voicing), or at the very least that this was true of the bilingual speakers who initially borrowed these words. The prevoiced [ɡ] is different from the way /ɡ/ is usually produced in English (though is closer to the way it is produced in other languages Dutch has borrowed from, such as French or Italian), and the contrast that it forms with [k] for our participants is thus specifically relevant to Dutch. Since Dutch speakers use a native phonetic contrast (i.e., prevoicing versus short lag VOT, in line with other stop voicing contrasts in their language), they rather implement the source language phonological contrast in a way that is meaningful within the Dutch phonetic space, even though it might not strictly correspond to the phonetic realization in the source language. We may consider how this Dutch phonetic contrast came about through borrowing from languages like English (where the contrast is usually realized as a difference in the presence versus absence of aspiration). One possibility is that the speakers who brought in the loanwords originally were aware of the abstract contrast between /ɡ/ and /k/ (in English or other source languages) and implemented the contrast in a way that is meaningful in Dutch (in analogy with other stop voicing contrasts). For example, a Dutch speaker of English might know that the English word coast – /koʊst/ – [k^hoʊst] contrasts with the English word ghost – /goʊst/ – [koʊst] and thus that there is a contrast between /k/ and /ɡ/. If they also know that words like post and boast are similarly contrasted by the presence or absence of aspiration, and that the aspirated categories map onto their native voiceless categories in Dutch while the unaspirated categories map onto their voiced categories, the jump to a voiced velar phone is reasonably straightforward. Additionally, the presence of [k]~[ɡ] allophony in Dutch might allow other listeners to more accurately perceive this new contrast when the borrowing speaker uses it (we return to this point below when discussing our perception results). Naturally, this whole process may be aided by orthography, since Dutch speakers of English are invariably literate in both languages, and attention is drawn to this contrast in English language teaching in the Netherlands.

One aspect of lexical borrowing that we have yet to address is lexical specificities that might affect the way individual words are adapted (e.g., some words may be systematically produced with the unadapted variant, while other words may display inter- or even intra-speaker variability). For example, Lev-Ari and Peperkamp (2014) mention the Arabic sound [w] in loanwords borrowed into Hebrew (with Hebrew lacking [w] natively); some words are systematically borrowed with an adapted sound [v], while others maintain the Arabic pronunciation. Any number of word-specific properties could drive such differences. Crawford (2007) suggests that the type frequency of emerging phonological structures may affect their (non-)adaptation, taking as an example the previously illicit and now often unadapted coronal+[i] sequences in Japanese (e.g., in English team). He argues that if a structure occurs in many words, listeners are more likely to encounter it, and in turn be better at perceiving it without perceptual adaptation to a native structure. Nagy (2010), however, points out that certain previously foreign sounds which recombine features already present in the borrowing language may be reliably borrowed even if they appear in only a small amount of (infrequent) words. For example, he notes the emergence of contrast between long [yː] (borrowed) and short [y] (native) in Dutch (which recombines vowel length with a native category), despite the long phone only occurring in a handful of borrowed words (e.g., centrifuge). In the statistical analysis we performed on our experimental results, we simply considered word-specific factors as a type of noise by including word as a random effect in our models. Indeed, with only 12 individual items in our production task it is difficult to disentangle the various lexical attributes of individual words (e.g., frequency, age of introduction, source language). Additionally, many (though not all) of the words have domain-specific uses; goal for example is only used in reference to scoring points during a sports match and does not replace native doel globally (cf. examples like lunch which has supplanted middageten almost entirely for most speakers). Given the restricted semantic domains in which many of these words appear and their relative rarity, it may prove rather difficult to systematically tease apart the word-level effects that might influence adaptation strategy. Future work might thus seek to include a wider range of target words in order to better ascertain which lexical characteristics influence (non-)adaptation.

One might also ponder the social aspects that could affect the emergence of the [k]~[ɡ] contrast in Dutch. Previous explorations of Dutch /ɡ/ pointed to regional variability in the use of the phone [ɡ], with speakers from more peripheral regions (those located further away from the Randstad, the main population centre of the Netherlands) generally producing fewer [ɡ] tokens and using rather more adapted tokens (van Bezooijen & Gerritsen, 1994; Van de Velde & van Hout, 2002). Specifically, van Bezooijen and Gerritsen (1994) found that the participants who produced an adapted sound such as the fricative [𝜒]—which corresponds to the orthographic pronunciation of the word—were nearly always from the far south of the Dutch language area. Van de Velde and van Hout (2002) similarly found higher proportions of an adapted variant in the south and higher proportions of [ɡ] in the north. This could potentially reflect the urban environment of the Randstad compared to the relatively rural environment in much (though not all) of the peripheral areas. While we did not set out to test this specific hypothesis, we explored this in our data post hoc, by log-transforming the population density of the home region¹⁷ of our speakers and using it as a predictor of [ɡ] production in a mixed effects logistic regression, including a random intercept for Item (which itself included a random slope for Population Density).¹⁸ We found that this model was a significantly better fit than a simpler model without the predictor Population Density, suggesting that speakers from areas with a higher population density tended to produce more [ɡ] (β = 0.32, 95% CI = [0.08 : 0.60], 𝜒²(1) = 10.86 p < 0.001). We may speculate what could drive such differences in production. Language attitudes, and specifically views of foreign languages like English being a threat to Dutch could differ between more urban versus more rural areas. However, a more straightforward interpretation might be that, in more urban areas, there is simply more contact with non-native speakers of Dutch (and potentially use of English as lingua franca, especially amongst the more highly educated in university towns), leading urban speakers to show different production patterns from rural speakers. This is a clear avenue of interest for future research.

4.2. Perception

Next, we turn to our consideration of perception alongside production. Following the variability observed in production, we might have expected to observe a high range of variability in perception. Specifically, individuals who often produce [ɡ] in loanwords could be expected to also perform well in a perceptual task requiring them to distinguish between emerging [ɡ] and native [k], with individuals who produce less [ɡ] in turn being less good at distinguishing the new sound from the native one. While we found some evidence for this, the observed effect is not overwhelming. We did find that participants discriminated the emerging contrast less well than an established contrast in Dutch, but overall accuracy was very high. This means that variability amongst participants was low, and this limited our ability to observe a strong link between production and perception. Any conclusions concerning the link between the two are therefore tentative. We will start by presenting a limitation of our perception results and then discuss two plausible lines of explanation for the lack of variability we observed, one concerning the specifics of our sample and the other concerning the specifics of [ɡ].

First, we should point out a caveat to the interpretation of our perception results. We compared perception of the emerging contrast relative to an established baseline, but we did not compare perception of the emerging contrast relative to a pair of sounds that are completely absent from the Dutch inventory. Our perceptual results indicate that the contrast between [k] and [ɡ] is perceived slightly less well than the contrast between [p] and [b], and while we take the position that this is due to the emerging nature of the phoneme /ɡ/, it is also compatible with a view that says that /ɡ/ does not form a contrast with /k/ in the Dutch inventory. Such an account would predict that the ABX results are reflective of our participants’ ability to perceive any similar ‘non-contrast.’ Inclusion of a contrast like [c]~[Ɉ] (also a stop voicing contrast) in our experiment would help disentangle these views: If our listeners have a harder time discriminating the non-contrast than the emerging [k]~[ɡ] contrast, that would be strong evidence for an abstract representation of /ɡ/ (and in line with a gradient view of contrast since [k]~[ɡ] was still perceived as less distinct than [p]~[b]). In fact, Peperkamp and Bouchon (2011), who used an ABX discrimination task with an identical procedure, including a 20-trial training phase with feedback, found that accuracy on a non-native, non-emerging contrast only reached about 70%. This suggests that performance in these kinds of tasks can in fact be much lower than what we observed for our emerging contrast. This also suggests that the feedback during training was unlikely to have provided much training on our target contrasts (beyond training on the task itself).

Additionally, certain aspects of the present study make the gradient contrastiveness account more plausible. First, our ABX task required participants to abstractly represent the different stimuli; a simple acoustic comparison (which might allow participants to distinguish even a contrast they were not familiar with) was not available due to the nature of the task. This was likely bolstered by the fact that we used two different female voices for the A and B stimuli and a male voice for the X stimulus, with the crucial contrast appearing in a non-prominent position within three-syllable non-words. Second, we observed an individual-level link between phonetic implementation and frequency of use of the new sound [ɡ] (participants who produced longer prevoicing also produced more prevoiced tokens). This is exactly the kind of relation predicted by a gradient view of contrast. Thus, although worse discrimination of a contrast like [c]~[Ɉ] compared to the emerging contrast would be strong evidence for gradient representation, evidence from similar studies in addition to features of our own design already speak to this issue.

One aspect that may have limited our ability to observe large differences in perception has to do with our experimental sample. Our testing site, Utrecht, is located in the aforementioned urban Randstad area, where [ɡ] usage is reportedly high. This means that all of our participants, who were all young university students, have been exposed to productions of [ɡ] in loanwords and may therefore be well trained in the perceptual realm, without necessarily producing many [ɡ] tokens themselves. Therefore, we may have observed less variability in the perception of the contrast than we saw in production due to the social specifics of the sample we tested. Previous work has shown that perceptual adaptation to contrasts not present in one’s native dialect does not completely follow shifts in production, though an individual level link may still be present (Evans & Iverson, 2007). This suggests that phonological perception and production may evolve separately (or at least semi-independently), and is a clear avenue for future research around social aspects affecting loanword (non-)adaptation.

A second aspect that could explain some of the lack of variability we observed in perception has to do with what makes the contrast between [k] and [ɡ] an ideal candidate to emerge in the first place: the place of [ɡ] in the native Dutch phonological system. The phone [ɡ] is not entirely absent from native Dutch phonology, as it exists as a contextual allophone of /k/ (specifically, in regressive voicing assimilation). As noted in the introduction, previous work has found that allophone perception is considerably better when the target contrast is presented outside of the native licensing environment (Peperkamp & Dupoux, 2003; Whalen et al., 1997). In our case, target [ɡ] in our ABX discrimination task was not presented in a context that licenses regressive voicing assimilation. Thus, perhaps participants’ performance in our ABX task was particularly strong since the target contrast is a well established case of allophony in Dutch. This high perceptability could in fact bolster the chances that a foreign contrast will come to distinguish words in the borrowing language. Indeed, if all speakers are capable of perceiving the contrast relatively well, it has a higher likelihood of taking hold at the population level.

Allophonic status has been pointed to in explaining some cases of phonological contrast emergence through borrowing. For example, Blevins (2009) notes that while Mekeo lacks contrastive coronals in its native lexicon, coronals do surface as allophones of velars in some phonological contexts. This, then, could explain why certain recent borrowings from English may be pronounced with coronals, like tauli for “towel.” This in time could lead to the emergence of a contrast between velars and coronals in the language. Kennard and Lahiri (2020) go so far as to propose that phonological emergence through lexical borrowing is necessarily dependent on the new contrast already existing as an allophonic alternation in the borrowing language. Our Dutch case would then be a prototypical example of this process: Dutch speakers can distinguish between [k] and [ɡ] if they occur in onset position (i.e., outside of the context where they natively alternate, namely preceding obstruents), making a non-adaptation more likely to take hold in a word like buggy.

Furthermore, it has been claimed that foreign sounds which recombine features or gestures already present in the borrowing language may be more likely to be borrowed without being adapted (Cohen, 2019; Ussishkin & Wedel, 2003). This could also favour the emergence of a phonological contrast between /k/ and /ɡ/ in Dutch, since the feature [voice] is already used to make contrasts in stop consonants. As noted above, ‘filling the gap’ in the native inventory adheres to the principle of feature economy (Clements, 2003), since [voice] becomes maximally contrastive within the stop series, in turn making the system on the whole more economical. While phonological inventories are by no means universally maximally economical, they do seem to be more economical than might be expected by chance (Coupé, Marsico, & Philippson, 2011; Dunbar & Dupoux, 2016; Mackie & Mielke, 2011). This typological evidence suggests that there may be a cognitive bias constraining learning that would lead to expectations of economical systems, and indeed, higher-order organizational pressures, such as economy, are often suggested to play a role during language acquisition (e.g., Thompson & de Boer, 2017). For our case, this could mean that Dutch-learning children might expect [k] to contrast with [ɡ], since other pairs of stops are contrasted along the same dimension. Relatively little input might therefore be required for the contrast to be acquired. It might suffice for a single word in a young child’s lexicon to contain /ɡ/ (e.g., buggy pronounced [bʏɡi], a quite common word in most Dutch-speaking children’s lives) for them to assume a voicing contrast for velar stops.

There is some evidence from Dutch-speaking children that this account is reasonable. Stoehr, Benders, van Hell, and Fikkert (2022) found that children between 3 and 6 years old produced varying rates of [ɡ] after being exposed to novel names produced by an experimenter with prevoicing (e.g., Gabi produced as [gabi]).¹⁹ Overall, 27% of target segments were produced with prevoicing (other tokens were produced like [kabi]), but at least one child produced consistent prevoicing on all tokens. This inter-speaker variability is in line with the results of our production task and with previous investigations of Dutch /ɡ/. The authors of that study argue that their results are evidence of generalization of the feature [voice] from labial and coronal contrasts (contrasts that are ubiquitous in the children’s input) to the velar place of articulation (an emerging contrast that some of the children tested may not yet have heard). This is precisely the type of generalization predicted by a feature economy account.

4.3. Putting it all together

Overall, our participants may all have been capable of perceiving the emerging contrast well, but their use of the contrast in production might be subject to further constraints (which could explain the relatively larger variability we observed in our production results). Two explanations seem likely to have a role to play. Firstly, prevoicing is more effortful to produce during velar constrictions than during coronal or labial constrictions simply due to the fact that the space behind the constriction is smaller (and pressure thus builds up more quickly). Indeed, this can explain why cross-linguistically, missing /ɡ/ from a stop inventory is more likely than missing /d/ or /b/, and also why prevoicing at more posterior places of articulation tends to be shorter (e.g., Ohala, 1983). This argument is bolstered by the fact that we observed higher [ɡ] rates in our production data in intervocalic contexts (which favour voicing) than in non-intervocalic contexts (see Table 1). Secondly, as mentioned above, we did not explore the social associations that speakers have with the different variants. There may be specific (potentially stigmatized) social meanings associated with any of the variants, and this might differ by speaker. For example, some speakers might find it pompous to use the unadapted variant in casual speech, while others might find it uneducated not to. Future research might specifically explore these potential differences.

Let us turn finally to what this case of loanword phonology can tell us about the representation of phonological contrasts. Taking the distribution of contrasts as a starting point, we can imagine that a contrast like /t/~/p/, which is ubiquitous in Dutch, and is not subject to neutralization, should be perceived on the far end of a continuum of distinctiveness. A contrast like /p/~/b/, which is also ubiquitous, but is subject to neutralization (i.e., final devoicing means these sounds are not contrastive at the ends of words) is likely to be perceived as slightly less distinctive than /t/~/p/, and should therefore be placed less far along a continuum of distinctiveness. This is in line with previous work showing that listeners’ perceptions of different contrasts reflect properties of those contrasts’ lexical distribution (e.g., K. C. Hall & Hume, 2013; K. C. Hall et al., 2015; Martin & Peperkamp, 2017). Indeed, this continuum seems to closely track K. C. Hall’s (2013) predictability of distribution continuum, but evidence from perception shows that the story is even more complex. Martin and Peperkamp (2017), for example, show that French listeners are indeed biased by their knowledge of contrast distribution in the French lexicon during word recognition, but also that those same listeners are more attentive to contrasts with a greater acoustic difference. The latter type of knowledge is posited to be part of the grammar in substance-based theories of phonology (i.a., Archangeli & Pulleyblank, 1994; Bybee, 2001; Hayes & Steriade, 2004), and can be used to explain why certain phonological rules are more common cross-linguistically than others (namely, that certain alternations are perceptually more distinct than others, see, e.g, Steriade, 2001). While proposing how language-specific and language-independent knowledge integrate to influence contrast perception is beyond the scope of this paper, the data presented here show that fine-grained knowledge should be part of our understanding of contrast. It does not do justice to the status of /ɡ/ to simply say that /ɡ/ only appears in a handful of words in Dutch (and therefore plays only a marginal role, if any, in the phonology). The contrast between /k/ and /ɡ/ is already used by many speakers, and some of the minimal pairs distinguished by this contrast may be present in Dutch children’s input (e.g., Stoehr et al., 2022, mention that some of the children in their study knew the word goal which forms a minimal pair with the native word kool, “cabbage”). Clearly we must consider contrasts to be dynamic, evolving, and intertwined with perception. The predictability of the distribution of a contrast is an important starting point, but data like we have presented here can begin to shed light on the representation of a contrast on the one hand, and also on the intricacies of a contrast’s implementation in production (including complex patterns of variability) on the other.