1.1 Unreliable beginnings are problematic

Despite differences in what is assumed to be the nature of the input (e.g., acoustic-phonetic material, articulatory gestures or constellations, abstract phonemes, or distributional information from sequences of phones), most models of word recognition, like the Cohort Model (Marslen-Wilson & Welsh, 1978), TRACE (McClelland & Elman, 1986), and Shortlist B (Norris & McQueen, 2008) assume that word recognition begins with the word-initial segment. Under most models, this first segment and the following segments feed two concurrent processes: On one hand, they activate possible lexical competitors in the listener’s vocabulary—a top-down process—and on the other, they progressively eliminate competitors that do not match the presented material—a bottom-up input-driven process.

Models like the Cohort Model, TRACE, and Shortlist B invoke connectionist, associative and/or probabilistic mechanisms and assume that the process of word recognition is linear and progressive. For instance, in Shortlist B (Norris & McQueen, 2008; see Figure 1 below), processes of word recognition begin immediately—and the first segment or two may even provide a match to a word in the lexicon—and then continue to use incoming segmental information to activate or eliminate lexical candidates. Eventually, the competition will be won by the highly activated sequence of lexical candidates that successfully accounts for the entire string without any leftover elements (Cutler, 2008; 1996; Norris & McQueen, 2008; Norris, McQueen, & Cutler, 2000).

Figure 1

Adapted from Norris and McQueen (2008). Schematic representations of word recognition in the British English phrase [ðəkætəlɒɡɪnəlaɪbrɪ] ‘the catalogue in a library’ in Shortlist B. The dashed and dotted arrows in the figure are examples of connections between non-contiguous words.

Models like Shortlist B (Norris & McQueen, 2008) attempt to also account for observations from the predictive parsing literature that many well-established factors play a role in word recognition, including sub-phonemic transitional information (Marslen-Wilson & Warren, 1994), phoneme and lexeme frequency, lexical neighbourhood density (Marslen-Wilson & Welsh, 1978; Metsala, 1997), and adherence to or violation of semantic and syntactic rules (Spoehr, 1980; Miller & Isard, 1963), such that a particular construction may favour an activity verb, a countable noun, or human referent (Smirnova et al., 2019). Research also shows that intonation, as well as word stress and tone, likewise provide cues about upcoming syntactic and lexical structures (Kjelgaard & Speer, 1999; Roll et al., 2010; 2011; Hirose & Mazuka, 2015; Söderström et al., 2017). And it is well-established that not only are whole words activated or eliminated in the lexicon, though Figure 1 might appear to suggest that this is the case, but words that are near-matches to a presented word are also activated (Davis & Taft, 2005; Taft & Forster, 1976). Additionally, word-internal morphological complexity—and speaker morphological knowledge—is also important to word recognition/processing (Bundgaard-Nielsen & Baker, 2020) and morphological competitors have been shown to play a role in the identification of complex words (see e.g., Balling & Baayen, 2012) as listeners may make decisions about the identity of some morphemes before the end of a complex word is reached.

Importantly, though often implicitly, models of continuous parsing and research into human word recognition mechanisms also assume that the input, whatever it is assumed to be in each model, is faithful to the lexical specifications of the entries in the lexicon. To quote Cutler (2008, p. 1602): “Psycholinguists largely take the ‘front end’—the initial processing applied to raw acoustic input—for granted, assuming that it will deliver a representation of the input that is in a form suitable for accessing stored lexical entities” [emphasis added]. And often, this might seem a safe and reasonable assumption to make. Indeed, in the Shortlist B example in Figure 1 (Norris & McQueen, 2008)—the British English phrase [ðəkætəlɒɡɪnəlaɪbrɪ] ‘the catalogue in a library’—few would argue with the fact that [k^hæt] ‘cat’ is stored in the minds of English-speakers as /kæt/ and will not necessarily be well-recognised if produced as [ɡæt] (a change in the voicing specifications of the initial segment that crosses the phonetic VOT boundary), or [pæt] (a change in place of articulation), or [ɣæt] or [ɰæt] (a change in manner of articulation from a stop to a fricative [ɣ] or approximant [ɰ] at the velar place of articulation) by speakers. Changes of this sort potentially lead to activation of the wrong set of lexical competitors, where the phonetic realisation results in a phonological substitution (/kæt/ ‘cat’ to /pæt/ ‘pat’ or /ɡæt/ ‘gat’, or no activation at all). These ‘wrong paths’ would require reinterpretation or repair of the input at some later point, in a scenario that is not possible under Shortlist B assumptions of linear progression from the root node (R) to a terminal node (T).

This example is extremely important because it demonstrates that psycholinguistic models of word recognition—even ones that Shortlist B, for example, rely upon include on a combination of both path probabilities (i.e., predictions about ‘what comes next’) and information about phonetic confusability between phonemes—tend to be poorly equipped to deal with variability in the input. We expect this to be particularly problematic in cases of word-initial mutations as found in Welsh, and in terms of the complex (even multidirectional) phone-to-phoneme correspondences explored in Mawng and Iwaidja in the present study.

The (insular) Celtic language family, to which Welsh belongs, is well-known for its richness in word-initial consonant mutations. For instance, the word /pont/ ‘bridge’ may appear also as /bont/ or /font/ in Welsh (Ball & Müller, 1992; Fife & King, 1998) depending on grammatical environment. Initial mutations of the kind found in Welsh are often considered morphological rather than phonological (Ussishkin et al., 2017) and raise questions about what is stored in the lexicon (is it one form, or many? And is the lexicon multi-tiered?). Welsh-style alternations in surface forms also raise questions of whether selection of correct forms is the result of rule-application (syntactic, morphological, and/or phonological), as is assumed under generativist/Chomskyan models of language processing (e.g., Chomsky & Halle, 1968; Gaskell & Marslen-Wilson, 1996) or associative patterns that can be accounted for by connectionist, relational or associative morphological processes/schemas and extensive listings in the lexicon (e.g., Rumelhart & McClelland, 1986; Walsh et al., 2010; Jackendoff & Audring, 2020), as well as in TRACE (McClelland & Elman, 1986) and Shortlist B (Norris & McQueen, 2008).

Changes to the surface form of words, and even to the initial segments of words, have also been described for many other languages, including Spanish. In Spanish, voiced stop consonants /b d ɡ/ have two regular and predictable allophones both word-initially and within words, such that stops /b ɡ/ are realised as approximants [β̞ ɣ̞] everywhere but after nasals (where they are realised as stops), and /d/ is realised as approximant [ð̞] everywhere but after nasals and laterals. The Spanish allophonic realisations, however, are unlikely to induce competition in lexical activation, as the lenited forms are not canonical realisations of other Spanish phonemes, and thus do not cause perceptual confusion or neutralisation, or ‘pernicious homophony’ (Campbell, 1996; Blevins & Wedel, 2009). That is, the changes involved are allophonic and ‘non-structure-preserving’, in Kiparsky’s (1982) sense, because the lenited allophones are not segments in the inventory, unlike the case of Welsh and other languages with initial mutation, which do involve substitutions between phonemes, i.e., neutralisation.

Within phonological theory, it has been observed that lenition and neutralisation are far less common in initial position than in later positions in the word (Houlihan, 1975; Beckman, 1998). An explanation for this fact is offered by Smith (2004, p. 1456), who notes that “Nooteboom (1981) and Taft (1984) have argued that speech processing is more efficient when positions that are especially important in (early-stage) word recognition are given as large a number as possible of phonological contrasts to draw from,” and thus word-initial and root-initial syllables are particular prone to maximise phonological contrasts and resist neutralisation (see also Wedel et al., 2019).

Australian Indigenous languages buck these trends, however. As first pointed out by Dixon (1972), the apical contrast (between apico-alveolar and apico-retroflex) is typically neutralised or may be absent altogether in word-initial position. Australian languages also display patterns of initial phonetic/phonological alternations (Blevins, 2001; Baker, 2014). Initial alternation phenomena among Australian languages are typically phonologically conditioned and involve alternations between phonemes in the inventory, rather than allophones (Baker, 2014). That is, the changes are structure-preserving, as in Welsh, but phonologically-conditioned, as in Spanish, and present a case which is distinct from both. For example, in Wubuy (Heath, 1984), every morpheme which begins in either a continuant consonant or a vowel also has a stop-initial form found after non-continuants (nasals and stops) within words. In (1.3) for example, the verb for ‘race’ begins with a /w/ underlyingly, and appears in that form in the word in (1.3), because the preceding segment is a vowel. But in (1.4), the form of the verb begins in a stop /p/ because the preceding segment is a nasal. This alternation is called ‘hardening’ in (Heath, 1984). ‘Hardening’ is completely productive and exceptionless in Wubuy.¹

Few experimental studies of initial alternations in Australian languages have been undertaken, but there is one processing study of phonological hardening in Wubuy (Bundgaard-Nielsen, Baker & Wang, 2023). Using a cross-splicing preference task like the one employed in the present study, this work suggests that phonological hardening of the kind found in Wubuy may be better accounted for in terms of application of a phonological rule modifying an underlying lexeme than by selection between competing lexemes. This contrasts with what has been reported for Scots Gaelic (Ussishkin et al., 2017) and Welsh (Boyce et al., 1987), where the phenomenon is accounted for as a morphological process.

1.2 Word-initial mutations and alternations in Iwaidja and Mawng

Mawng and Iwaidja are non-Pama-Nyungan languages spoken on Croker Island, an island in the Arafura Sea off the coast of the Northern Territory, Australia, approximately 250 km northeast of Darwin.² Croker Island was the site of the Croker Island Mission (1940–1968). Currently, the only large settlement is the Aboriginal community of Minjilang. According to the Australian Bureau of Statistics 2021 census, 234 people live in Minjilang, of which approximately 90% are Indigenous Australians. Of these, 36% identify as speakers of Iwaidja, 30% as Kunwinjku, and 10% as Mawng (Mawng is primarily spoken in the community of Warruwi on Goulburn Island). Just over 4% of Minjilang’s population identify as speakers of Kuninjku (as with Kunwinjku, a variety of Bininj Gun-wok: Evans, 2003), and another 4% as Burarra speakers. Multilingualism is the norm in the community, and all participants in the present study report speaking at least three languages (Iwaidja, Mawng, and English). Language input domains and usage patterns, however, vary between individuals on Croker Island, depending on family histories and ties. Many speak one or two Indigenous languages as their first language, but there are also individuals who use English very frequently or exclusively, whilst others use English only in domains like school or social or health services. English may also be used as a lingua franca in the absence of a common Indigenous language. Indigenous languages occupy almost all community domains except school and government-related communication. For official community announcements, the default language is usually Iwaidja, and there is no translation into other languages provided (Mailhammer, 2021). This is the only sizable population of speakers of Iwaidja anywhere. Shaw et al. (2020, p. 579) provide a “generous estimate” of around 50 speakers of Iwaidja. Mawng is a somewhat larger language: Singer et al. (2021) estimate around 400 speakers.

Both Iwaidja and Mawng have been described as having historical and grammatically-conditioned initial mutations in verb and noun roots (Evans, 1998; 2000) that differ from Wubuy in being no longer phonologically-conditioned. Examples of the grammatically-conditioned initial mutations in Iwaidja are illustrated in (1.1) and in (1.5)–(1.6), adapted from Mailhammer and Harvey (2018, p. 334). In these examples, there is an alternation between initial /w/ and initial /b/ in the verb stem, but the alternation cannot be predicted on the basis of the phonological environment: In (1.5), /w/ appears after a vowel-final prefix, while we get /b/ in this environment in (1.6). In contrast, /b/ occurs word-initially in (1.5), but following a vowel in (1.6). Therefore, the alternation can only be described in terms of the morphological environment and, in particular, the person/number/gender features of the arguments. Evans (1998) proposes that the alternations have a basis in the historical phonological environment but evidently, they are opaque synchronically and could potentially be accounted for by an approach like that taken for Celtic languages (see, for instance, Ussishkin et al., 2017).

The focus here, however, is that Mawng additionally has been described as having additional synchronic, phonologically-conditioned external sandhi processes affecting initial /ɡ/ at word boundaries (Capell & Hinch, 1970), but the additional presence of widespread mutations in the grammar of both languages makes the question of the importance of the identity of the word-initial segment a potentially more interesting one for these languages. Indeed, speakers of Mawng and Iwaidja have been reported to behave linguistically in a way that makes it difficult to align their phonological systems and speech recognition processes with what is typically assumed of speakers of other languages. This makes the linguistic behaviour—here, word recognition with variable input—extremely important from a theoretical point of view, as well as from a descriptive point of view.

The additional synchronic external sandhi processes described for Mawng affect initial segments at word boundaries. Both the grammars of Mawng (Capell & Hinch, 1970) and Iwaidja (Pym & Larrimore, 1979) include a velar approximant or fricative in the inventory, which can only occur phonemically in word-medial positions in either language. However, word-initial velar stop /ɡ/ in Mawng—but not Iwaidja—is reported to lenite to [w], [ɣ] or [ɰ] following vowels, liquids and glides in the preceding word (Capell & Hinch, 1970, p. 44), as in examples (1.2a) and (1.2c), whereas only the stop is found following nasals and stops as in examples (1.2b) and (1.2d).³ There has been no empirical instrumental study of sandhi in Mawng to date. We note, however, that a recent acoustic and ultrasound study with speakers of Iwaidja (Shaw et al., 2020) reports considerable inter- and intra-speaker variation in the realisation of Iwaidjan words involving /ɡ/ and /ɰ/. They conclude that there is no evidence from production for a phonemic contrast between /ɡ/ and /ɰ/ despite evidence for a bimodal distribution of [ɰ] and [ɡ]: The data suggests that some participants use one form, and others typically another (Shaw et al., 2020, p. 608). The study considered only a subset of the environments in which the velar approximant has been described as phonemic, which further qualifies the conclusions. The fact that Mawng displays synchronic lenition and Iwaidja does not is the critical difference explored in the present studies.

The present studies test the effect of word-initial segmental variability, presumably arising from phonetically-conditioned synchronic lenition processes exemplified in examples in (1.2a) and (1.2c) on word acceptability in Mawng and Iwaidja. We do so by testing speakers’ tolerance of initial segmental alternations of two kinds—onset lenition (stops to approximants) and onset fortition (unattested change of approximants to stops)—in a 2AFC experiment with bilingual Mawng and Iwaidja speakers from Croker Island in the Northern Territory of Australia. The 2AFC paradigm implemented here allows us to examine the way in which surface phonetic alternations influence word activation in speakers. It does so because participant word preference decisions require successful word recognition, and the only thing that differs between the competing selections is the phonetic characteristics of the first segment (if the phonetic difference between forms presented does not matter in terms of word acceptability, neither option will be favoured). The results thus provide novel insights into the importance of first-segment phonetic-phonological fidelity in word recognition (no preference without recognition). The results additionally provide new data on some of the factors that potentially constrain listener tolerance for phonetic variability in word initial phonological segments.

2.1 Stimuli

We recruited one female speaker of Iwaidja and one female speaker of Mawng to produce the stimulus material for the two 2AFC experiments; both speakers were bilingual in Mawng and Iwaidja. The Iwaidjan speaker was aged 50 at the time of the recording and reported speaking Iwaidja, Mawng, English, Kunwinjku, and Kunbarlang. She had some literacy in English, Iwaidja, Mawng and Kunwinjku, and was very familiar with a range of language research activities, including recording of narratives and elicitation of material for acoustic analysis, sociolinguistic interviews, and ultrasound for language research purposes.

The Mawng speaker was 64 years old at the time of the recording. She reported speaking Mawng, Iwaidja, and English, and had some literacy in all three languages. Like the speaker of Iwaidja, she was familiar with a wide range of language research activities, including recording of narratives and elicitation of material for acoustic analysis, sociolinguistic interviews, and ultrasound for language research purposes. A researcher trained in phonetics recorded the speakers in a quiet home in Minjilang on Croker Island. The recording devices used were a Panasonic HC-VX-1 HD camera with Countryman omnidirectional lapel microphones with Sennheiser ew100G3 wireless transmission. All recordings had a 16-bit sampling depth with a sampling rate of 44.1 kHz. Each speaker was compensated for her time with a payment of $50.

The speakers each produced five iterations of two carrier frames (see Table 1). One frame was structured so that the target words followed a vowel (/a/); in the other, the target word followed a stop (/d/ in Iwaidja and /b/ in Mawng). The carrier frames had the same semantic content in both target languages. The speakers also produced five iterations of each depictable noun target in its canonical form in each frame, as well as five lenited iterations of each depictable stop-initial noun in each frame, and five hardened iterations of each continuant-initial noun target: Lenition targets beginning in /b/ were elicited with [w]-onsets; lenition targets beginning in /ɡ/ were elicited with [ɣ~ɰ] onsets (see discussion of elicitation and transcription procedure below and Figures 2a and 2b); and lenition targets beginning in /ɟ/ were elicited with onset [j]. Fortited targets beginning with /w/ were elicited with onset [b], and fortited targets beginning with /j/ were elicited with [ɟ] onsets. The resulting word pairs (a canonical form plus either a lenited or fortited form) are presented in Appendix 1.

Figure 2

Spectrograms 2a and 2b present spectrograms of the target word [ɡabal] ‘flood plain’ in Iwaidja (a) and Mawng (b). Spectrograms 2c and 2d present spectrograms of lenited onset [ɰabal] forms of ‘flood plain’ in Iwaidja (c) and Mawng (d).

The mutated forms of the Iwaidja and Mawng target words were elicited by explaining to the speakers that the different ways of pronouncing the same word were required so that the correct pronunciation of the Iwaidja/Mawng word could be better understood. To support the speakers, a phonetically trained researcher modelled the lenited and hardened forms of the initial segments only (not the target words), and the speakers were then asked to reproduce each target word with the modified onset. No corrections or adjustments were made by the researcher to the pronunciation variants produced by the speakers in either language, and the variants produced were all within the described range of variants for each language respectively.

We assume that the lenited realisations typically transcribed as [ɣ] and [ɰ] are phonetic gradients (more versus less constriction: see also Shaw et al., 2020), and that transcription discrepancies in the published literature on these languages reflect the practical need to select an IPA symbol despite realisations falling on a continuum of greater to lesser constriction. For convenience, we use the approximant symbol [ɰ] throughout, without making assumptions about the degree of constriction of the lenited form. We note, however, that existing research such as Shaw et al. (2020) and Ennever et al. (2017) on the Gurindji language has found very few, if any, fricative productions of lenited stops. In Figure 2, the four spectrograms (2a, b, c, and d) provide example spectrograms of recordings of the Iwaidja and Mawng cognate target word /ɡabal/ ‘flood plain’ in the stop-initial canonical/underlying form and in the lenited word form [ɰabal], elicited from stop-final and vowel-final frames, respectively.

From the recorded nouns, a subset of 14 target nouns for each language were selected for inclusion in the four experiments, as were two sets of minimal pairs as Control Trials (totalling 16 unique trials for each language). The Control Trials were both /b/-/w/ minimal pairs to avoid issues associated with variability in the realization of /ɡ/ in Mawng and Iwaidja. This resulted in a total of 16 word-pairings (see Appendix 1). To generate the stimuli for the experiments, we selected one token of each of the two frames in both Iwaidja and Mawng from iterations 2, 3, or 4 (avoiding list-initial and list-final intonation) as well as one iteration of each target noun (canonical or modified) also from iteration 2, 3, or 4, based on similarity in intonation structure, intensity, and speaking rate. These frames and individual words were then normalised for intensity and the experimental stimuli were generated by cross splicing each target noun (canonical or modified) into each carrier frame in both Iwaidja and Mawng.

2.2. Experimental design

The recorded materials were used to design stimuli for two parallel 2AFC experiments, one in Mawng and one in Iwaidja, testing participant preference for canonical versus mutated noun forms in the two carrier frames (vowel-final versus stop-final) in two prosodic juncture conditions (no pause versus pause inserted). The 2AFC paradigm used for the present study constitutes a preference task—participants are required to pick a preferred stimulus out of two items differing only in the onset characteristics of the target noun: Critical trials differ only in presenting the noun with the underlying/canonical onset versus either lenited or hardened onset, and control trials also differ only in having one continuant and one stop initial target word (same place of articulation: POA), constituting a minimal pair.

Systematic preference for either stop or continuant-initial target words relies on two things: ability to discriminate between stops and continuants with the same POA, and knowledge about the phonological specification of each target word (what is the canonical pronunciation, for instance, of the Iwaidja and Mawng /ɡabal/ ‘flood plain’?). Inability to discriminate between stop and continuant-initial targets, or uncertainty about the phonological specifications of the target words, would lead participants to perform at chance level. Chance level preference patterns, in turn, would be expected from cognate target words (i.e., the control trials), if they indeed activate separate lexical entries, because selection of either alternative constitutes a valid choice.

Each experiment (Mawng versus Iwaidja) was presented in two forms, differing only in whether a silent pause had been inserted before the target noun, presented in counterbalanced order. Each experiment (natural versus pause-insertion) consisted of 16 target words in two carrier frames (see Table 1) in two presentation orders, resulting in 64 individual trials. The order of presentation of the 64 trials was pseudo-randomised, ensuring that consecutive trials did not contain the same target nouns. The pause-insertion condition was created by insertion of 500 ms of silence before the onset of the target nouns. This pause insertion temporally distances the potential phonological conditioning environment from the target word onsets (canonical, lenited or hardened), because external sandhi phenomena are known to be sensitive to prosodic juncture (Kilbourn-Ceron et al., 2020).

Each trial belonged to one of three types (See Appendix 2 for all pairings):

1. Lenition Trials in which the target word was presented in its canonical form and in a lenited continuant-initial form e.g., [baŋɡa] ‘forked stick’ in both Iwaidja and Mawng versus *[waŋɡa], where /waŋɡa/ is not a canonical word in either language;

2. Fortition Trials in which the target word was presented in its canonical form and in an unattested stop-initial (hardened) form, e.g., [wamba] ‘shark’ in both Iwaidja and Mawng versus *[bamba]; and,

3. Control Trials which consisted of minimal pairs of words that differed only in the initial segment: one had an approximant onset, the other a stop onset at the same POA, in a manner similar to the Lenition and Fortition Trials, e.g., [baɡaj] ‘spike/harpoon’ versus [waɡaj] ‘sugar glider’ in Iwaidja and Mawng. In these trials, each realisation arguably activates different entries in the lexicon.

Some of the included target nouns were cognate between Iwaidja and Mawng, for optimal comparison, while others were not. By ‘cognate’ we mean that the forms are identical in each language (whether this is for either reasons of genetic inheritance or for borrowing). One Fortition Trial was cognate (/wamba/ ‘shark’), three Lenition Trials were cognate (/baŋɡa/ ‘forked stick’; /ɟaɾaŋ/ ‘horse’; and /ɡabal/ ‘flood plain’), and one Control Trial was a cognate (/baɡaj/ ‘spike/harpoon’ versus /waɡaj/ ‘sugar glider’), and one a partial cognate (/baɳɖi/ ‘armband’ versus /waɳɖi/ ‘one who is hanging’ [in Iwaidja]⁵; see discussion in Results).

The distribution of Lenition and Fortition Trials also differed slightly between the two languages (see Table 3): The Iwaidjan participants were presented with 12 Fortition Trials (six continuant-initial words plus their non-attested fortited forms); 16 Lenition Trials (eight stop-initial words plus their lenited continuant-initial forms), and 4 Control Trials (two minimal pairs). The Mawng participants were presented with 10 Fortition Trials (five continuant-initial words plus their non-attested hardened forms), 18 Lenition Trials (nine stop-initial words plus their lenited continuant-initial forms), and 4 Control Trials (two minimal pairs; see Table 3).

Each study (Mawng versus Iwaidja, each in the unmodified and the pause-inserted conditions) was delivered to the participants in the form of a PowerPoint presentation. During each study, the participants heard each pair of utterances through headphones from a laptop and at the same time were presented with two line-drawings of faces on the computer screen. For each pair of utterances, the listeners were encouraged to think of the experiment as two Iwaidja- (or respectively, Mawng-) speaking people trying to tell them something, and they were instructed to indicate which of the utterances they preferred. Previous research has demonstrated that framing experimental psycholinguistic research in a socially meaningful narrative improves participant experience and ensures the widest possible range of participants (Bundgaard-Nielsen et al., 2015; Bundgaard-Nielsen, Baker, & Wang, 2023; Bundgaard-Nielsen & Baker, 2020). Limited literacy and computer skills, as well as the need for the stimuli to be socially meaningful, limit the suitability of more traditional experimental approaches, including the use of gating tasks. Participants were allowed to listen to each pair as many times as they liked, and to take as long as they liked before making their decision. Participants were allowed breaks at any point in time.

Our experimental protocol diverges from standard laboratory speech science paradigms in word recognition studies (such as gating tasks, which typically rely on participants having strong independent test-taking and literacy skills: see e.g., Grosjean, 1980) but was implemented to accommodate a wide range of speakers (i.e., different ages, educational backgrounds, familiarity with electronics) as well as to accommodate a testing situation rich in interruptions and social interactions, and to provide a culturally safe environment (Smye et al., 2010). The approach taken here, along with other low-tech and low formality approaches, have been successfully used with both adults and children in other Indigenous speech communities in Australia for psycholinguistic research (Bundgaard-Nielsen et al., 2015; Bundgaard-Nielsen & Baker, 2020; Bundgaard-Nielsen, Baker, Bell, & Wang, 2023; Bundgaard-Nielsen, Baker, & Wang, 2023). These alternative methods are particularly important to consider in communities with a low number of speakers, to ensure that as many speakers of a given language can participate, even when they may face increased barriers to participation, including limited literacy, limited familiarity with electronic equipment and experimental protocols, as well as Western research practices. It is also more accommodating for participants who wish to adhere to local cultural practices that favour interaction in smaller groups over one-to-one interactions with (unfamiliar) outsider researchers.

2.3. Study participants

We recruited 11 L1 speakers of Iwaidja (8 Male; 3 Female), ranging in age from 26 to 88 years of age (M = 49). We also recruited 11 L1 speakers of Mawng (8 Male; 3 Female), ranging in age from 22 to 88 years of age (M = 44). All participants completed a language background questionnaire. All participants, irrespective of which experiment they participated in (Iwaidja or Mawng), reported speaking both Iwaidja and Mawng fluently, as well as Kunwinjku and English to varying degrees of fluency. One participant additionally reported having some competence in Kunbarlang, and another reported some competence in Amurdak, Marrku, and Garig. All but one of the participants also reported regular exposure to Kunbarlang and/or Yolngu Matha. All but one participant were literate in English to some extent, and most reported having very basic Iwaidja and Mawng literacy as well (one participant reported having excellent literacy in both Iwaidja and Mawng). Four participants participated in both the Iwaidja and the Mawng experiments on separate days.

All participants were recruited by word of mouth on Croker Island, and all testing took place in a quiet home on Croker Island. Many of the participants had previously participated in language research. All were compensated for their time and effort by a payment of $50 which is standard in remote communities where living costs are very high and opportunities for paid employment are very limited. The study was approved by the Western Sydney University Human Research Ethics Committee, approval number H14890.

2.4. Predictions

1. Fortition Trials. We would expect both Iwaidjan and Mawng participants to reject all unattested hardened target nouns. We would not expect an effect of phonological frame (preceding stop versus preceding vowel), nor an effect of presence or absence of a pause. We include fortition trials to differentiate between tolerance for lenition and inability to perceptually discriminate between stop and approximants at the same POA: Consistent rejection of fortited approximants (even if lenited stops are acceptable) indicates that participants can perceive the difference between [j] and [ɟ], and [w] and [b]. This also gives us a baseline rejection rate.

2. Lenition Trials. We would expect Iwaidja speakers to be intolerant of segmental alternations of the kind presented in the Lenition Trials in the same manner as those presented in the Fortition Trials, and we would not expect an effect of phonological frame, nor an effect of presence or absence of a pause. Based on reports of synchronic sandhi processes affecting onset /ɡ/ in Mawng, we would expect Mawng speakers to be more tolerant of lenited onset targets than of unattested fortition alternations and more tolerant than the participants in the Iwaidjan experiment. We would also expect an effect of phonological frame, such that lenition would be more acceptable in the vowel-final frame than in the consonant final frame (because this matches the phonological conditioning of lenition described in the literature).

3. Control trials. We would expect the participants to be at chance (50% preference) for 2AFC trials consisting of minimal pairs, as each target word should activate different, equally acceptable, lexical competitors, even though the phonetic/phonological difference between the two words is of the same kind as the pairs that participants are exposed to in the Fortition and the Lenition trials.

4. Prosodic Juncture Condition. We would expect Mawng speakers to be more likely to accept lenited wordforms in the unmodified condition than in the pause-insertion condition. If lenition is a bottom-up, phonetically conditioned sandhi phenomenon, we would expect an effect of pause insertion, such that pause insertion might reduce the acceptability of lenited onsets, as the temporal distance to the conditioning environment is extended (Kilbourn-Ceron et al., 2020).