1.1. Djambarrpuyŋu

Djambarrpuyŋu is spoken by approximately 4,000 people in northeast Arnhem Land (Australian Bureau of Statistics, 2021a). It is a member of the Yolŋu subgroup of the Pama-Nyungan language family (Bowern, 2023), and is one of a small number of Australian Indigenous languages that are used in day-to-day life and learned by children at home (see Koch & Nordlinger, 2014 for discussion). Today, speakers live in a number of communities including Galiwin’ku, Gapuwiyak, Milingimbi, and Ramingining (Wilkinson, 2012; Yunupingu, 1996). Djambarrpuyŋu is now a main community language of these communities, which are not all traditional Djambarrpuyŋu clan and language affiliation areas. Speakers are often multilingual and multidialectal with knowledge of related varieties, other Australian languages, Australian English, and Kriol.

The vowel system proposed by Wilkinson (2012) in her grammatical analysis of Djambarrpuyŋu is symmetrical, consisting of six vowels distinguished by height, backness, and—in the first (stressed) syllable of the word—length. To reflect the tendency for vowels in Australian Indigenous languages to have less extreme cardinal-like articulation and more variation within each quality category, the following symbols are used here: /ɪ, ɪː, ɐ, ɐː, ʊ, ʊː/ (Busby, 1980; Fletcher & Butcher, 2014). The length contrast is exemplified in the following near minimal pairs: weṯi /wɪːʈːɪ/ “wallaby,” wiṯitj /wɪʈːɪc/ “olive python”; gorrum /kʊːrʊm/ “to be high,” gurrum’ /kʊrʊmʔ/ “soft”; and gäna /kɐːnɐ/ “alone,” gana’ /kɐnɐʔ/ “enough” (Wilkinson, 2012). The restriction of contrastive vowel length to the initial syllable (i.e., the stressed syllable) is commonly reported for Pama-Nyungan languages that have a length contrast (Baker, 2014; Dixon, 2002; Fletcher & Butcher, 2014).

A preliminary acoustic study of Djambarrpuyŋu vowel duration from words in isolation confirmed that duration significantly differed between phonemically short and long vowels in the first syllable of CV.CV words with short vowels having a mean duration of 126 ms and long vowels a duration of 211 ms (Jepson & Stoakes, 2015). Wilkinson (2012) notes, however, that the phonetic contrast can be lost in fast connected speech, which results in the collapsing of minimal pairs. She additionally noted that in reduplicated forms, reduplicants do not retain the long vowel, for example, yolŋu’-yulŋu /jʊːlŋʊʔjʊlŋʊ/ “people” (Wilkinson, 2012; see also Heath, 1980 on Ritharrŋu).²

There are 25 consonants in Djambarrpuyŋu: Six places of articulation for stops, for which there are fortis and lenis series (with the former represented by length in the following), /pː, p, t̪ː, t̪, tː, t, ʈː, ʈ, cː, c, kː, k/, and nasals /m, n̪, n, ɳ, ɲ, ŋ/, as well as one trill /r/, three central approximants /w, ɹ, j/, two lateral approximants /l, ɭ/ and glottal stop /ʔ/. The fortis/lenis contrast is yet to be acoustically analysed. Note that the transcription convention adopted here of using the length diacritic for fortis stops does not imply ambisyllabic association, rather it is used to capture that these stops are hypothesised to differ primarily in closure duration, rather than voicing (Butcher, 2004; Round, 2023). Round summarises that languages of Arnhem Land (including Yolŋu languages) with two series of stops usually make the distinction by closure duration, in contrast to languages of Cape York, which primarily use voicing. Further, this convention was adopted in previous phonetic research on the language (e.g., Jepson et al., 2021). Phonologically, stops are limited in where they are considered contrastive in Djambarrpuyŋu—only in morpheme internal intercontinuant position. The contrast between the fortis and lenis stop series has a very low functional load, though near minimal pairs do exist (Wilkinson, 2012). For example: bäba /pɐːpɐ/ “gall,” bäpa /pɐːpːɐ/ “father”; gadharra /kɐt̪ɐrɐ/ “coral,” watharr /wɐt̪ːɐr/ “white”; and räga /ɹɐːkɐ/ “white berry bush,” räkay /ɹɐːkːɐj/ “Eleocharis dulcis.” These specific minimal pairs were not a focus of data collection and the differences between the fortis and lenis stop series are not investigated acoustically in the current paper.

Djambarrpuyŋu allows for syllables that are maximally CV(C)(C)(ʔ); all syllables (and words) are consonant initial with the exception of a very small number of words, predominantly in the “baby talk” register (Wilkinson, 2012, p. 45), for example, anyany /ɐɲɐɲ/ “cute.”

2.1. Contrastive vowel length

Cross-linguistically, contrastive vowel length is most commonly reported with two levels of contrast; generally, “long” vowels are distinguished from “short” (Odden, 2011). A pattern of three levels is rare (Remijsen, 2014 on Dinka). Phonetically, long and short categories are correlated with larger or smaller duration values respectively, with the exact values for each category being language specific (Odden, 2011, p. 465). The raw duration values can vary considerably. For example, in Thai utterance-medial words, short vowels are, on average, 160 ms, and long vowels are 320 ms (Roengpitya, 2001, p. 30), while in Hungarian accented utterance-medial words, short vowels are 54 ms and long vowels are 94 ms (White & Mády, 2008). The ratio or proportional difference between vowel length categories can also show substantial variation. For example, in Marshallese the ratio is approximately 1:2.8 (Choi, 1992, p. 64), but in Czech, the ratio is 1:1.63, averaged across five vowel pairs (Podlipský et al., 2009, p. 134).

Vowel duration is well known to also be conditioned by several other factors, such as metrical strength, accentuation, voicing and manner of the following consonant, vowel quality, syllable structure, number of syllables in the target word, whether uttered in isolation or in an utterance, and the vowel’s position within the target word (e.g., Lehiste, 1970; Port & Dalby, 1982; Turk, 2012). These factors can affect the duration of long and short vowels differently, and consequently, the proportional differences between the categories may also change (e.g., de Jong & Zawaydeh, 2002, p. 62, their Figure 1, on Ammani-Jordanian Arabic and effects of stress; White & Mády, 2008 on Hungarian and syllables per word). Of particular interest here is syllable structure; position within the utterance is also considered.

Cross-linguistically, vowels in closed syllables are found to be phonetically shorter than in open syllables (Maddieson, 1985). This effect has been observed in languages that have contrastive vowel length such as Dutch (Jongman, 1998; Rietveld & Frauenfelder, 1987), Arabic (Broselow et al., 1997; de Jong & Zawaydeh, 1999; Khattab, 2007), and Malayalam (Broselow et al., 1997), as well as those which do not such as Italian (Farnetani & Kori, 1986; Hajek et al., 2007). The position of the word within the utterance is also found to have an effect on segment duration, with vowels in words uttered in utterance-final position having longer duration than comparable vowels in words in other positions (White & Mády, 2008 on Hungarian). For Hungarian, final-lengthening interacts with vowel length category such that final long vowels show a greater effect of position than short vowels, being lengthened approximately 40 ms, while short vowels only lengthened an estimated 21 ms (White & Mády, 2008). See also Paschen et al. (2022) for a discussion of the interaction between vowel length and finality.

Beyond duration, contrastive vowel length is also correlated with spectral differences for monophthongs. Typically, membership in the long category corresponds to being more peripheral in the vowel space, whereas being short corresponds to a more centralised acoustic realisation. This is based on a more general assumption that longer duration of a vowel corresponds with more extreme articulation because articulators are able to move to more peripheral positions, while in shorter periods of time the movement that is possible is restricted (Cho, 2004; Lehiste, 1970, p. 31; Lindblom, 1963, p. 1780). However, vowel length is not always found to affect formant structure (cf. Behne et al., 1996, on Norwegian). And, the prediction does not hold for short versus long vowels in some Australian languages (Fletcher & Butcher, 2014). Of relevance here, however, is Graetzer’s (2012) analysis of vowels in Gupapuyŋu, a Yolŋu variety closely related to Djambarrpuyŋu, in which she found that long vowel centroids tended to be more peripheral in the vowel space in at least one dimension: the pair /ʊ/ and /ʊː/ differed in F2, while the pairs /ɪ/ and /ɪː/ and /ɐ/ and /ɐː/differed in F1 (Graetzer, 2012, pp. 187–188). It is possible that there are also minimal vowel formant differences in Djambarrpuyŋu due to articulatory mechanisms that result in differences which are not sufficient to posit vowels of different qualities.

2.2. Phonetics and phonology of vowel-consonant duration relationships

Not infrequently, a complementary duration relationship between a vowel and the following consonant is described, most often when the vowel is stressed. The duration relationship is inverse, with consonant duration being longer when the vowel is shorter, and shorter when the vowel is longer. This is reported to occur in, for example, Arabic (Khattab, 2007), the Shetland dialect of English (Sundkvist & Gao, 2015; van Leyden, 2004), German dialects (Kleber, 2020), Italian dialects (Baroni & Vanelli, 2000), Norwegian (Behne et al., 1996), Swedish dialects (Elert, 1965; Schaeffler, 2005), Thai (Abramson, 1960), Washo (Yu, 2008), and Guienagati and Ozolotepec Zapotec (Benn, 2021; Leander, 2008). There are a range of phonological, phonetic, and prosodic motivations associated with this type of durational relationship; sometimes the pattern is attributed to the vowel, sometimes to the consonant, and sometimes to another syllable weight or duration requirement.

For Washo, “[p]ost-tonic consonant gemination after short stressed vowels may be viewed as a compensatory response to the monomoraicity of short vowels” (Yu, 2008, p. 517). This interpretation of preserving moraic weight is discussed within phonological theory (Gordon, 2016; Hayes, 1989) and is observed in some Australian languages (Baker, 2014; see also Harvey & Borowsky, 1999 on Warray). For Ngalakgan, an Australian language of the Gunwinyguan family, Baker (2008, pp. 75–82) describes a bimoraic minimum for monosyllabic words which can be achieved through a closed syllable containing a short vowel, or the phonetic lengthening of vowels if the syllable is open. In the moraic theory view of compensatory lengthening, the lengthening of a vowel conditioned by the loss of a coda consonant would not occur if the language’s codas are non-moraic for other purposes, such as stress assignment or minimal word requirements (Gordon, 2016, p. 160). Most Australian languages are described to be weight insensitive regarding stress assignment (Baker, 2014; Fletcher & Butcher, 2014; Jepson & Ennever, 2023), irrespective of whether consonants are moraic or not. There is a preference for minimally bimoraic monosyllabic words in Djambarrpuyŋu, despite a small number of exceptions (discussed in Section 2.3). The possibility of a phonetic bimoraic minimum requirement for stressed syllables is discussed in sections 4.3 and 6.

For the Shetland dialect of English, van Leyden (2004) found that consonants were lengthened in a compensatory manner with the shortening of vowel duration. It was reported that a 100 ms shortening in vowel duration was reflected by a 49 ms lengthening of the following consonant. Van Leyden (2004, p. 39) concluded that, while there is an inverse relationship between vowel and consonant duration in the Shetland dialect, it is strongest for some sets of words considered to constitute pairs, particularly local dialect words.

In Thai (Abramson, 1960; Abramson & Ren, 1990), semivowels and nasals in syllables containing short vowels have longer duration than those in syllables containing long vowels, but again, durational differences are not as great as they are for vowels and do not fully compensate for the difference in duration between long and short vowels: /n/ in /sǐn/ was 150 ms while the /n/ in /sǐːn/ was 110 ms long (Abramson, 1960, p. 132), whereas short and long vowels were on average 85 ms and 165 ms, respectively (Abramson & Ren, 1990). In languages with geminate consonants, such as Makasar for example, singleton /n/ is found to be 100 ms while geminate /n/ is 205 ms on average (Tabain & Jukes, 2016); that is, it is on par with durational differences observed for contrastive vowel length.

When longer vowels are only observed with shorter consonants and shorter vowels with longer consonants, difficulty can arise in determining whether the vowel or the consonant has contrastive length. In Thai and the Shetland dialect of English there is a smaller difference in duration between consonants, leading to vowel length being analysed as the primary determiner of vowel duration and, consequently, determining the duration of the following consonant. However, this is not always so clear cut (e.g., in Swedish Schaeffler, 2005). Central Bavarian presents an interesting case showing the difficulty of determining the motives for varying duration or neighbouring segments (Kleber, 2020). In Central Bavarian, tense (long) vowels occur before lenis stops, and lax (short) vowels occur before fortis stops. Prior to Kleber (2020), there were two proposed phonological solutions to this: Firstly, that the fortis/lenis contrast is phonological while vowel duration is predictable; secondly, that there is a contrast between the complementary length of VC sequences and that prosodic quantity specifies either the vowel or stop to be long, and the other is then predictably short. Kleber concluded that, because lengthening was observed for sonorants (which do not otherwise have a length contrast) and not exclusively for the stops, that vowel length is phonemic in Central Bavarian, and the consonant duration contributes to marking the vowel length contrast. On Washo, Yu (2008) acknowledges that his analysis rests on an assumption of vowel length being contrastive and that the issue is still under debate.

2.3. Segment duration in Yolŋu languages

As mentioned in Section 1, an inverse duration relationship is reported for other Yolŋu languages, with longer vowels being followed by shorter consonants, and vice versa. This type of durational relationship is proposed to be a phonetic feature of most Australian languages that have contrastive vowel length (Dixon, 2002, p. 591). When contrastive vowel length, which had historically been more commonly observed around Australia (Dixon, 2002, p. 639), is lost, consonant duration is hypothesised to be phonologised as being contrastive (Dixon, 2002, pp. 591, 595). Though, for a number of Yolŋu languages, two series of stops are attested, as well as contrastive vowel length (the historical conditions that could account for this are discussed by Dixon, 2002, pp. 604, 611). It appears that this has contributed in part to the diverse analyses for the similar phonetic outcomes in Yolŋu languages, the details for which are discussed here.

Of the varieties and languages discussed below, Djambarrpuyŋu is most closely related to Djapu—both are within the same language group of Dhuwal—while it is more distantly related to Ritharrŋu, Djinang, and Gälpu which are in different language groups (Bowern, 2023; Wilkinson, 2012). In some descriptions, sonorants are the focus of lengthening processes, while in others, obstruents are primarily discussed.

Morphy (1983) reports for Djapu that in disyllabic words where the first/stressed syllable is open, the duration of the following consonant varies according to the length of the stressed vowel. The relationship is predictable and complementary, with short vowels being followed by longer consonants, and long vowels followed by shorter consonants. She illustrates this with the pair waŋa /wɐŋɐ/ “to talk/speak” and wäŋa /wɐːŋɐ/ “home/place” (Morphy, 1983, p. 25). It is argued that the lengthening or “gemination” of the consonant following the long vowel occurs to ensure that the stressed syllable is heavy. In Morphy’s analysis, a syllable is heavy as the result of either a long vowel in the stressed syllable, due to a coda consonant, or, when the following consonant is the onset in the next syllable, the lengthening of that consonant which is consequentially analysed as being ambisyllabic (to fulfil the requirement that all syllables are consonant initial). So, in the above examples, waŋa /wɐŋɐ/ [wɐŋːɐ] is syllabified as [wɐŋ.ŋɐ], while wäŋa /wɐːŋɐ/ [wɐːŋɐ] is syllabified as [wɐː.ŋɐ]. Morphy notes that the durational difference could be analysed as originating from either the vowel or the consonant, but argues that to motivate the lengthening of the consonant is simpler. She suggests there would be no way to motivate lengthening of vowels in words like yolŋu /jʊːlŋʊ/ “person,” which would be a heavy syllable by nature of being closed.

Heath (1980) posits a length contrast for vowels in stressed syllables in Ritharrŋu, though notes that there is a tendency for long vowels to be phonetically shortened in words of three or more syllables as well as in reduplicants (p. 12), as Wilkinson (2012) has also reported for Djambarrpuyŋu (Section 1.1). Lengthened consonants, specifically liquids and nasals, are reported to occur intervocalically when following short vowels. Heath (1980, p. 12) had originally posited a length contrast for consonants in addition to vowels. He provides a discussion of his changing analysis, citing the occurrence of long vowels before consonant clusters, and gives examples such as the word yalu “nest” which was originally transcribed as [jɐlːʊ] (orig. [yallu] in text), explaining consonant duration varies phonetically, without phonological implications. Ritharrŋu is analysed as having fortis and lenis stop series though the effect of vowel length on those consonants is not discussed.

In Djinang (Waters, 1979, 1989), a similar phonetic pattern is described, though the phonological analysis and underlying mechanism differs to Djapu and Ritharrŋu. Waters proposes that vowel length is not contrastive. Rather, vowels optionally lengthen in stressed syllables, especially when the syllable is open (see discussion in Ryan, 2019, p. 36), and that consonants, specifically voiceless stops (i.e., fortis stops in other analyses), geminate following stressed syllables when the stressed vowel is phonetically short.

Wood (1978), on the other hand, proposes for Gälpu (also Gaalpu) that vowel length is determined by a “syllable feature” of length—short or long. Long vowels occur in syllables with the long feature, while short vowels occur in short syllables. Stops lengthen in intercontinuant environments, and lengthening is especially “pronounced” when the vowel is short and the consonant is the onset to the second syllable (i.e., is intervocalic) (Wood, 1978, p. 64). He terms the analysis “a prosodic solution” in contrast to a segmental solution in which vowel length would be analysed as contrastive (i.e., as in Morphy, 1983). These two options are both considered valid analyses by Wood, but he argues that the prosodic solution is better supported as it reflects the analysis of the glottal stop as syllable feature (first proposed by Bernhard Schebeck in a paper in 1974, cited in Wood, 1978), possibly historically associated with stress (Wood, 1978, p. 98).

For Djambarrpuyŋu, Wilkinson (2012) posits contrastive vowel length, but does not mention a durational relationship with the following consonant. Jepson and Stoakes (2015) examined intervocalic consonants in CV.CV Djambarrpuyŋu words spoken in isolation, and found that their duration differed significantly depending on whether they occurred following a phonemically (and durationally) short or long vowel, with consonants following short vowels being 236 ms in duration and consonants following long vowels being 191 ms. This study, also discussed in Section 1.1, supported the phonetic facts of what has been reported for Yolŋu languages: that the vowel in the first syllable of words varies in duration, that the duration of the following consonant varies too, and these are often in a complementary relationship. However, as only CV.CV words were examined, the analysis of segments in words with different syllable structures is necessary to understand the segment duration patterns of Djambarrpuyŋu.

All of these analyses rely to a degree on stress, and a requirement for stressed syllables to be a certain weight (i.e., heavy), or privileged to some degree. That is, contrastive length occurring only in stressed syllables could be interpreted as showing positional prominence, with more segmental contrasts maintained in stressed syllables than unstressed syllables in these languages (Hyman, 2014). On the other hand, as mentioned above (Section 2.2), it is generally assumed that stress is not quantity sensitive in most Australian languages, with most systems having fixed stress either to a word or root morpheme boundary (Baker, 2014; Fletcher & Butcher, 2014; Jepson & Ennever, 2023). The preference for heavy stressed syllables could suggest that some version of a Weight-to-Stress principle, which specifies heavy syllables are stressed, could be at play (Gordon, 2011; Prince, 1990). It is noted that this principle is violated in so much as unstressed syllables can also be heavy (e.g., /ˈɹɐj.mɐl/ “temple, side of head”; /ˈn̪ʊ.mɐn/ “smell, give off smell”), but it could motivate a preference for phonetically heavy stressed syllables. Djambarrpuyŋu allows monosyllabic words of the structure CV, for example /kɐ/ “and,” or the present and today future form of the imperfective auxiliary (Wilkinson, 2012, p. 44). However, there is a preference for minimally bimoraic roots in Djambarrpuyŋu, and from descriptions of other languages with prosodic minimality constraints, function words are observed to be prosodically deficient such as Ixtayutla Mixtec (Penner, 2019). So, it may be that minimality constraints related to word structure apply in Djambarrpuyŋu, and a similar minimum may also apply specifically to stressed syllables in Djambarrpuyŋu.

Considering other motives for lengthening, Solé and Ohala (2010, pp. 608, 611, inter alia) suggest that language-specific secondary properties that covary with a contrastive dimension could serve to enhance the contrast (see also Schertz & Clare, 2020). That could be the case for compensatory lengthening of the type proposed for Yolŋu and other Australian languages, to enhance the contrast between short and long vowels. This too is in conflict with the prevailing understanding of the phonetics of Australian languages. The large consonant inventories with many places of articulation that are found in most Australian languages, it is argued, require a maximisation of acoustic cues to consonant place at the expense of vowel contrasts (Butcher, 2006). This tendency for preserving or enhancing the place information of consonants over enhancing vowels has been termed the “Place of Articulation Imperative” by Butcher (2006) and is phonetically realised as the “lengthening and strengthening” of, in particular, post-tonic consonants in some Australian languages, serving dual purposes of maintaining and possibly enhancing paradigmatic consonant contrasts on the one hand and signalling prosodic prominence on the other (Butcher, 2006; Fletcher & Butcher, 2014). This preference is not observed across the board in Djambarrpuyŋu, but is found in words of two syllables (Jepson et al., 2021).

In Jepson et al. (2021), the duration of intervocalic consonants was measured in words of two to six syllables in length, and post-tonic consonants (i.e., following the stressed vowel) in disyllabic words were found to be significantly longer than post-tonic consonants in words with more syllables and non-post-tonic consonants. The words analysed in Jepson et al. (2021) contained only short stressed vowels. Therefore, it is not entirely clear if the duration of the consonant was conditioned by post-tonic lengthening (a focus of Jepson et al., 2021) or due to following a short vowel, as explored in this paper (for more on post-tonic consonant lengthening in other Australian languages see Butcher & Harrington, 2003a; Butcher & Harrington, 2003b; Fletcher et al., 2010; Fletcher et al., 2015). Disyllabic words are mentioned explicitly in Morphy’s (1983) analysis of vowel length in Djapu, and because they partake in other segment duration processes in Djambarrpuyŋu, it appears that two syllable words could present an interesting and distinct word structure in Yolŋu languages, which may show processes that are otherwise not observed in words of other structures.

2.4. Perception of segment duration and vowel-consonant duration relationships

Vowel duration is the primary cue used by listeners cross-linguistically when categorising vowels as “short” or “long” in perception studies (e.g., German: Lehnert-LeHouillier, 2007, 2010; Tomaschek et al., 2011; Thai: Abramson & Ren, 1990; Lehnert-LeHouillier, 2007, 2010; Roengpitya, 2001; Japanese: Arai et al., 1999; Behne et al., 1999; Kinoshita et al., 2002; Lehnert-LeHouillier, 2007, 2010). Listeners of languages with a phonemic vowel length contrast use their knowledge of the sounds in their language to categorise stimuli into existing categories in categorisation experiments (Holt & Lotto, 2010; Liberman et al., 1957).

However, vowel duration can be one of many cues that lead to the categorisation of the phone to a length category. A cue, albeit the primary one, is neither independent of the influence of other cues on the perception of a phone, nor solely responsible for it. (Repp, 1982). Further, in perception studies (and in day-to-day speech), there are also stimuli with duration values that fall between the typical durational range of “short” and “long” categories found in the production data for a language. Stimuli within this range are ambiguous as they do not provide durational cues that are sufficient for listeners to categorise the phone. In vowel length perception, stimuli containing vowels in the language-specific ambiguous region can be more difficult for listeners to process, which has been reflected in reaction times (Eerola et al., 2012, for Finnish listeners). Further, vowels that are between duration categories are subject to the influence of secondary cues.

Covarying of secondary features may be used as reliable indicators of the primary contrast (Solé & Ohala, 2010, p. 611), and when the primary acoustic cue is ambiguous, secondary cues can influence listeners’ perception (Hillenbrand & Clark, 2000, pp. 3020–3021; Lehnert-LeHouillier, 2010; Whalen et al., 1993). Secondary features are not ignored by listeners when the primary cue is not ambiguous; these cues are used by listeners even when phonologically redundant (Whalen et al., 1993). This type of effect can alter the location of the crossover point between categories. In vowel length perception, spectral information and f0 are most commonly investigated as secondary cues. These cues are used differently across languages in perception, as they are in production (Lehnert-LeHouillier, 2010). Language-specific formant quality changes can also result in varying crossover points in the perception of length contrasts (Abramson & Ren, 1990). This is to say, altering of one cue can be compensated for by altering another cue, thus maintaining the original phonetic percept in a cue trading relation (see Raphael, 2005; Repp, 1982).

Durational information beyond the vocalic element has also been shown to be a cue to vowel length in perception. This can operate between the vowel and the following consonant. For example, in Thai, longer post-vocalic consonant duration was found to condition a higher proportion of short vowel responses than post-vocalic consonants with shorter durations (Roengpitya, 2001). Similarly, van Dommelen (1999) found for Norwegian that the boundary between the categories of long and short for vowels was influenced by the duration of the following consonant. Overall, there was a shift in where the boundary occurred, such that a longer consonant resulted in a later boundary between the vowel categories than a shorter consonant, suggesting longer consonants resulted in perceptual shortening of the vowel.

4.1. Methods

This study is based on a dataset drawn from a larger database produced as part of the author’s doctoral research (in the process of being archived in PARADISEC). The project had ethics approval from the University of Melbourne Human Research Ethics Committee (HREC number: 1544581). Further details of data collection and the database can be found in Jepson (2019a) and Jepson et al. (2021).

4.2. Results

4.2.1. Vowel duration

Figure 1 shows the distribution of values for long and short vowels in the dataset. A clear pattern can be seen in the data: Vowels that are phonemically long (in white) had longer duration values than their phonemically short counterparts (in grey). This observation is reflected in the results of the statistical analysis. Duration of the target vowels was statistically tested in a mixed effects model that included vowel length category (two levels: long, short), vowel quality (three levels: open central /ɐ/, close front /ɪ/, close back /ʊ/), syllable type (two levels: open, closed), utterance frame (three levels: frame initial, frame medial, and frame final), and C2 consonant category (i.e., “consonant category,” two levels: obstruent, sonorant) as fixed effects with an interaction between vowel length category and syllable type, and vowel length and utterance frame. Random effects included random intercepts for speaker and word, as well as by-speaker random slopes for the effect of vowel length category. Table 2 presents a summary of selected post-hoc comparisons.

Figure 1

Duration values (ms) of the six Djambarrpuyŋu vowels. Phonemically short vowels coloured grey, phonemically long vowels coloured white.

Table 2

Summary of mean duration values (ms), standard deviation (ms), minimum duration (ms), and maximum duration (ms) for vowels by phonemic category and length, and ratios between relevant pairs (long:short).

Vowel	n	Mean duration ms (SD)	Min. duration ms	Max. duration ms	Ratio of mean (long:short)
/ɪ/	112	103 (49)	35	340	1.99:1
/ɪː/	80	205 (65)	99	383
/ɐ/	589	102 (37)	22	310	1.92:1
/ɐː/	240	196 (61)	53	378
/ʊ/	294	86 (28)	24	169	2.13:1
/ʊː/	117	183 (57)	73	415
Length
short	991	97 (37)	22	340	2:1
long	437	194 (61)	53	415

Quality had a significant effect on vowel duration (χ²(2) = 16.95, p < 0.001). There was a small difference in duration for the three vowel qualities, between approximately 4 ms and 20 ms (see Table 3); however, only the comparison between the open central and the close back vowels was significant.

Table 3

Selected results from the post-hoc analysis comparing vowel duration (n = 1431).

Comparison	Estimated difference (ms)	Standard error (ms)	p-value
vowel length
short ~ long	–70.3	11.1	<0.0001
vowel quality
/ɐ/ ~ /ɪ/	3.6	6.2	1
/ɐ/ ~ /ʊ/	18.3	4.4	0.0002
/ɪ/ ~ /ʊ/	14.7	6.7	0.092
cons. category
obs ~ son	–33.4	4.9	<0.0001
vowel length:syllable type
V open ~ Vː open	–89.8	11	<0.0001
V open ~ V closed	14.2	5	0.0323
V open ~ Vː closed	–36.6	13.2	0.0631
Vː open ~ Vː closed	53.2	9.2	<0.0001
Vː open ~ V closed	104.1	11.2	<0.0001
V closed ~ Vː closed	–50.8	13.3	0.0046
vowel length:frame
V initial ~ Vː initial	–62.6	11.3	0.0013
V medial ~ Vː medial	–64.7	11.4	0.0009
V final ~ Vː final	–83.7	11.3	0.0001
V final ~ V initial	–1.6	2.2	1
V final ~ V medial	4.1	2.3	1
V initial ~ V medial	5.7	2.3	0.2371
Vː final ~ Vː initial	19.4	3.3	<0.0001
Vː final ~ Vː medial	23	3.5	<0.0001
Vː initial ~ Vː medial	3.6	3.5	1

The category of the following consonant also had an effect on duration (χ²(1) = 40.24, p < 0.0001), with vowels preceding obstruents an estimated 33 ms (SE = 5 ms, p < 0.0001) shorter than vowels preceding sonorants.

The interaction between vowel length category and syllable type had a significant effect on vowel duration (χ²(1) = 14.83, p < 0.001). Figure 2 shows the distribution of duration values for short and long vowels faceted by whether they occurred in open or closed syllables. While the duration of the short vowels in both syllable types was similar, the duration values of the long vowels differed between open and closed syllables, and there was more overlap between the categories in closed syllables; long vowels are an estimated 53 ms (SE = 9 ms, p < 0.0001) longer when in open than in closed syllables. Short vowels, on the other hand, were only an estimated 14 ms (SE = 5 ms, p = 0.0323) longer in open syllables than in closed syllables. Because syllable structure affected long and short vowels differently, the differences between the two vowel length categories differed between the two syllable structures: When the syllable is open, long vowels are an estimated 90 ms (SE = 11 ms, p < 0.0001) longer than short vowels, but 51 ms (SE = 13 ms, p = 0.0046) longer when the syllable is closed. See ratio values for raw duration values in Table 2.

Figure 2

Distribution of duration values (ms) of vowels facetted by syllable type, with the vertical lines representing the group means. Phonemically short vowels coloured dark grey, phonemically long vowels coloured light grey.

Lastly, the interaction between vowel length and position in utterance was statistically significant (χ²(2) = 32.54, p < 0.0001). Differences between long and short vowels remained significant in all frame positions, but long vowels were significantly lengthened when the word was in utterance final position, compared with long vowels in either initial or medial frames. Short vowels did not show an effect of lengthening when the word was in utterance-final position.

4.2.2. Vowel formants

Figure 3 plots Lobanov normalised F1 and F2 values extracted at the vowel temporal midpoint, plotted separately for each speaker. Centroids are plotted with 95% confidence intervals; short vowel centroids are plotted in dark grey and ellipses are represented with a solid line; long vowel centroids are plotted in light grey and ellipses are represented by a dashed line. There is considerable overlap of the short and long vowel pairs for all speakers. However, the centroid for the long vowel in each pair is often more peripheral in the vowel space than the short vowel centroid in either the F1 or F2 dimension. F1 is generally relevant for the open vowel pair /ɐ, ɐː/ with /ɐː/ having higher values than /ɐ/, while F2 is relevant for close vowels /ɪ, ɪː/ and /ʊ, ʊː/, with long vowels having higher or lower values, respectively, than their short counterparts. The long vowels also have smaller ellipses than their short counterparts. It is possible that there is less variability in formant frequency values of these vowels; however, there are also fewer tokens. Mean F1 and F2 values (Hz), and standard deviation values are presented in Table 4.

Figure 3

Mean first and second formant values (z-score; Lobanov normalised by speaker) at temporal midpoints of vowels with ellipses representing the 95% confidence intervals, plotted for each speaker (a–g). Phonemically short vowels are coloured dark grey with full line ellipses, phonemically long vowels are coloured light grey with dashed line ellipses.

Table 4

Summary of mean F1 and F2 frequencies (Hz) and standard deviations at vowel midpoints, for women and men (note: F2 for two /ɪ/ tokens could not be extracted).

	Vowel	n	F1 mean Hz (SD)	n	F2 mean Hz (SD)
Women
	/ɪ/	61	399 (46)	59	2214 (165)
	/ɪː/	41	399 (41)	41	2335 (126)
	/ɐ/	322	687 (115)	322	1588 (262)
	/ɐː/	123	799 (117)	123	1561 (159)
	/ʊ/	158	420 (46)	158	1067 (212)
	/ʊː/	64	437 (56)	64	928 (120)
Men
	/ɪ/	51	373 (68)	51	2106 (223)
	/ɪː/	39	386 (58)	39	2252 (185)
	/ɐ/	263	632 (145)	263	1434 (258)
	/ɐː/	117	723 (141)	117	1404 (147)
	/ʊ/	136	416 (88)	136	1015 (231)
	/ʊː/	53	408 (77)	53	866 (97)

Two models predicted F1 and F2 by vowel length category (two levels: long, short), vowel quality (three levels: open central, close front, close back), and gender (two levels: women, men) as fixed effects with an interaction between vowel length and vowel quality. Random effects included random intercepts for speaker, place of the following consonant, and word. Gender was included to account for variance in the data due to speaker sex, though is not examined in detail here.

Table 5 and Table 6 present selected post-hoc comparisons from the statistical analyses of F1 and F2 values, respectively. For F1, the interaction between length and quality was found to have a significant effect (χ²(2) = 41.11, p < 0.0001). The open vowel pair /ɐ, ɐː/ differed significantly in F1, with F1 being significantly lower for the short category vowel /ɐ/ than the long vowel /ɐː/ by an estimated –97 Hz (SE = 9 Hz, p < 0.0001). For F2, the interaction between vowel length and quality was again found to be significant (χ²(2) = 11.72, p = 0.0028). For the /ɪ, ɪː/ and /ʊ, ʊː/ pairs, F2 was found to differ significantly in the expected directions, with /ɪ/ having significantly lower F2 than /ɪː/ by an estimated –218 Hz (SE = 95 Hz, p = 0.025), and /ʊ/ having significantly higher F2 than /ʊː/ by an estimated 179 Hz (SE = 62 Hz, p = 0.005).

Table 5

Selected results from the post-hoc analysis comparing F1 frequency (Hz) at vowel midpoints (n = 1431).

Comparison	Estimated difference (Hz)	Standard error (Hz)	p-value
length:vowel quality
/ɪ/ ~ /ɪː/	20.51	18.60	0.2741
/ɐ/ ~ /ɐː/	–96.68	9.43	<0.0001
/ʊ/ ~ /ʊː/	–4.18	13.28	0.7535

Table 6

Selected results from the post-hoc analysis comparing F2 frequency (Hz) at vowel midpoints (n = 1429).

Comparison	Estimated difference (Hz)	Standard error (Hz)	p-value
length:vowel quality
/ɪ/ ~ /ɪː/	–218.4	94.9	0.0236
/ɐ/ ~ /ɐː/	37.3	38.7	0.3367
/ʊ/ ~ /ʊː/	179.0	61.8	0.0046

4.2.3. Consonant duration

Duration of the C2 target consonants was statistically tested in a mixed effects model which included vowel length category (two levels: short, long), syllable type (two levels: open, closed, i.e., where the consonant was intervocalic, or the coda, respectively), utterance frame (three levels: frame initial, frame medial, and frame final), and consonant category (two levels: obstruent, sonorant) as fixed effects with a three-way interaction between vowel length category, consonant category, and syllable type. Random effects included random intercepts for speaker and word, as well as by-speaker random slopes for the effect of vowel length category. Mean duration values for consonants and standard deviation values are presented in Table 7. This model was fitted specifically to test the three-way interaction between length of the vowel preceding the consonant, consonant category, and syllable structure.

Table 7

Summary of post-vocalic consonant duration values (ms), standard deviations (ms), and ratios between relevant pairs (after long:after short).

		n	Duration mean ms (SD)	Ratio of mean (after long:after short)
	Vowel length
	short	991	155 (93)	1:1.45
	long	437	107 49)
	Consonant category
	obstruent	319	254 (88)
	sonorant	1109	107 (48)
	Position in syllable
	intervocalic	1025	156 (92)
	coda	403	99 (46)
Summary relevant for interaction between vowel length, syllable structure, and consonant category
	C2 cat.; vowel length
Intervocalic	Obstruent
	after short vowel	261	269 (88)	1:1.37
	after long vowel	27	196 (50)
	Sonorant
	after short vowel	383	129 (54)	1:1.3
	after long vowel	354	99 (42)
Coda	Obstruent
	after short vowel	19	185 (56)	1:1.08
	after long vowel	12	172 (52)
	Sonorant
	after short vowel	328	92 (38)	1:0.94
	after long vowel	44	98 (36)

As shown in Figure 4, obstruents overall had longer duration values than sonorants. Intervocalic consonants were affected by the length of the preceding vowel, with consonants following short vowels being longer than those following long vowels. Consonants in coda position do not show an effect of the preceding vowel’s length.

Figure 4

Duration values (ms) of post-vocalic consonants grouped by category, length of the preceding vowel, and syllable structure. Violins are coloured by preceding vowel length: consonants after short vowels coloured grey, after long vowels coloured white.

These observations were confirmed in the statistical analysis (see selected results in Table 8). There was a significant difference in duration between consonants following short vowels and consonants following long vowels, with an estimated difference of 33 ms (SE = 11 ms, p < 0.005). For obstruents there was a difference of 45 ms (SE = 18 ms, p < 0.015), and 20 ms (SE = 11 ms, p < 0.057) for sonorants. Therefore, without considering the structure of the syllable, a difference due to the length of the preceding vowel is observed only for obstruents. However, there was a significant effect of the three-way interaction between vowel length category, consonant category, and syllables type (χ²(4) = 13.62, p = 0.0086). The duration of consonants was found to differ consistently by preceding vowel length in open syllables, with sonorants following short vowels longer than those following long vowels by an estimated 34 ms (SE = 10 ms, p < 0.036). Obstruent duration was affected in the same way, with obstruents in open syllables following short vowels longer than those following long vowels by an estimated 79 ms (SE = 21 ms, p < 0.008). Duration differences were not significant when the consonants were in closed syllables (p = 1). Position in the utterance frame also had a significant effect on consonant duration (χ²(2) = 16.55, p < 0.001), with consonants in utterance-final words significantly longer than when in utterance-initial or -medial words.

Table 8

Selected results from the post-hoc analysis comparing post-vocalic consonant duration values (Hz) (n = 1431).

Comparison	Estimated difference (ms)	Standard error (ms)	p-value
vowel length
after Vː ~ after V	32.8	11.3	0.0049
vowel length:consonant category
after Vː.obs ~ after V.obs	45.3	18.4	0.0151
after Vː.son ~ after V.son	20.3	10.5	0.0571
syllable structure
open ~ closed	40.8	10	0.0001
vowel length:consonant category:syllable structure
after V.obs.open ~ after Vː.obs.open	78.74	21.08	0.0081
after V.obs.closed ~ after Vː.obs.closed	11.89	29.19	1
after V.obs.open ~ after V.obs.closed	89.32	19.06	0.0002
after Vː.obs.open ~ after Vː.obs.closed	22.48	29.61	1
after V.son.open ~ after Vː. son.open	34.2	9.95	0.0366
after V.son.closed ~ after Vː.son.closed	6.47	16.68	1
after V.son.open ~ after V.son.closed	39.62	8.43	0.0002
after Vː.son.open ~ after Vː.son.closed	11.89	16.11	1

4.3. Summary and discussion

As hypothesised, vowel duration varied consistently with respect to the length categories of short and long. Long vowels, overall, where twice as long as short vowels. The third hypothesis, that intervocalic consonants would vary with the phonological length of the preceding vowel, was also confirmed. The ratio between the two consonant categories following short and long vowels was relatively similar, though overall, obstruents had longer duration than sonorants. One might have thought the putative fortis/lenis contrast would have resulted in less durational variation for obstruents, due to duration most likely being used as a means for distinguishing between the two stop series (Round, 2023). However, it could have also been the case that these fortis stops were lengthened to an even greater degree, as there would be no upper duration limit that could result in an overlap in categories. As lenis stops were not examined here, comments on their behaviour cannot be made. The finding that vowels are, overall, shorter before obstruents than sonorants may reflect a broader segment duration relationship in the language (see Section 4.2.1).

Overall, consonant duration varied to a lesser degree than vowel duration. For both consonant categories, the lengthening does not wholly make up for the shorter duration of short vowels; on average, short vowels were 97 ms and long vowels were 194 ms, while consonants were 155 ms after short vowels and 107 ms after long vowels. That is to say, as in the Shetland variety of English and Washo (van Leyden, 2004; Yu, 2008), while consonant duration is longer after short vowels by approximately 45% (see Table 7, overall ratio), this does not account for short vowels being only half the duration of long vowels. These findings contribute to the conclusion that contrastive vowel length, and not consonant length, remains the most plausible phonological explanation for the observed segment duration values.

The lengthening of the C2 consonant appears to result in the main stressed syllable in Djambarrpuyŋu being heavy, irrespective of the phonological analysis of consonants or vowels (as in Swedish, Thai, Washo, and the Shetland variety of English (Schaeffler, 2005; van Leyden, 2004; Yu, 2008)). The consonant lengthening pattern also does not appear to be conditioned by another word-level requirement, for example, bimoraic minimum, as the minimum was already met in this dataset because all the words were disyllabic. It could be, however, that there is a phonetic bimoraic minimum requirement for stressed syllables. The complementarity between vowel and consonant duration is taken up in the perception study, presented in Section 5.

With respect to the second hypothesis, formants were affected by vowel length; long vowels were more peripheral than short vowels. Close vowels varied in F2 (higher values for /ɪː/ than /ɪ/, lower values for /ʊː/ than /ʊ/), and the open vowel varied in F1 (higher values for /ɐː/ than /ɐ/), reflecting findings reported for Gupapuyŋu (Graetzer, 2012) and tendencies for the relationship between duration and formant structure cross-linguistically (Cho, 2004; Lehiste, 1970, p. 31; Lindblom, 1963, p. 1780).³ The relatively compressed vowel spaces with considerable overlap between short/long pairs is observed in many Australian languages (Fletcher & Butcher, 2014).

The final hypothesis, that syllable structure would affect vowel and consonant duration, was supported. Long vowels were shorter in closed syllables than in open syllables, while short vowels remained relatively constant in duration. This suggests there may be a threshold beyond which vowels cannot be further compressed (Siddins et al., 2013; see also Nooteboom, 1997, pp. 656–658). The effect of syllable structure on long and short vowels consequently affected the ratio between categories (see Table 2). The duration of intervocalic C2 consonants varied inversely with the length of the preceding vowel, while the duration of coda C2 consonants did not vary significantly due to vowel length. In Lebanese Arabic, it has been found that only long vowel duration is affected by following geminate consonants; they have shorter duration compared to when they are followed by singleton consonants, while short vowels remain the same duration in both conditions (Khattab & Al-Tamimi, 2014). That finding, while conditioned by a different process, echoes the Djambarrpuyŋu findings. Syllable association also operates in a different way in Dutch: C2 in CV1C2.CVC words was found to be of significantly longer duration when the V1 was short than when it was long, but in CV1.C2VC words C2 duration did not differ significantly (Jongman, 1998). Jongman argued that the status of the C2 consonant as ambisyllabic or tautosyllabic determines whether lengthening occurs, with only tautosyllabic consonants participating in the lengthening pattern. If the duration variation of the consonants was to enhance the vowel length contrast in Djambarrpuyŋu (in a way described by e.g., Solé & Ohala, 2010), we might expect that consonant duration values in Djambarrpuyŋu followed the Dutch pattern, as there is considerably more overlap in duration between the short and long vowel categories in closed syllables. However, we see no significant duration relationship between vowel and consonant duration when the consonant is coda to the first syllable. The opposing results agree that syllable structure “has a direct influence on the … acoustic realization of individual segments” (Jongman, 1998, p. 219); however, it does not necessarily have the same influence.

5.1. Methods

5.1.2. Stimuli

The minimal pair selected for investigation was waŋa /wɐŋɐ/ “to talk/speak” and wäŋa /wɐːŋɐ/ “home/place.” The selection of this pair was motivated by two factors. Firstly, there is only a small number of minimal pairs that vary only in vowel length in Djambarrpuyŋu, and this minimal pair is cited in linguistic research (e.g., Morphy, 1983) and by Djambarrpuyŋu speakers as demonstrative of the vowel length contrast. Secondly, these items are believed to be common in daily speech.

Due to the absence of an acoustic analysis of the fortis/lenis contrast in the stops, a perception study including stop duration in relation to vowel duration is not possible at this stage (also, a four-way minimal set that contains phonemic short and long vowels with fortis and lenis stops is not known to the author).

Durational measures from a subset of the production data were used to determine the range of segment duration values for the experiment. In those data, the durational range for the short vowel in /wɐŋɐ/ target words was 102 ms to 158 ms, and for the long vowel in /wɐːŋɐ/ target words was 132 ms to 286 ms, while the durational range for the nasal in /wɐŋɐ/ target words was 100 ms to 211 ms, and for the nasal in /wɐːŋɐ/ target words was 67 ms to 129 ms.

A single token of /wɐːŋɐ/ from the production data, spoken by a male speaker, 39 years of age, was selected to create the audio stimuli. This vowel is plotted in the F1–F2 space of all vowels to illustrate the quality of the selected token (Figure 5); it is in the overlapped area for the short and long open vowel counterparts. In this token, the vowel was 181 ms in duration, and the nasal was 83 ms in duration. The steady state portions of the vowel and nasal were selected for manipulation to preserve transitional information that occurred at the segment boundaries. This resulted in the vowel having 97 ms that was transitional from the preceding approximant and into the following nasal, and the nasal having 35 ms that was considered as transitional to/from the surrounding vowels. Therefore, in the token, 84 ms of the vowel and 48 ms of the nasal were manipulated.

Figure 5

Plotted normalised first and second formant values (z-score; Lobanov normalised) at the temporal midpoint of the wäŋa /wɐːŋɐ/ “home/place” token used to create the stimuli for the perception experiment (marked with ×), plotted against all data for all speakers; ellipses represent the 95% confidence intervals.

The duration of the vowel and the nasal segment was manipulated independently. Continua of seven equidistant steps were created for the vowel and for the nasal (following Lehnert-LeHouillier, 2010). Stimuli were created in Praat. The experiment was full factorial in its design; therefore, stimuli contained all combinations of the manipulated vowel steps and nasal steps, resulting in 49 unique stimuli. In the stimuli, vowels ranged between 110 ms and 260 ms, with 25 ms difference between steps. The consonant duration ranged from 90 ms to 200 ms, with 18 ms difference between steps. Due to the full factorial design, some stimuli contained vowel and consonant duration combinations that would infrequently, if ever, occur in natural speech. See Table 9 for the duration of steps in the continua.

Table 9

Duration (ms) of vowel and nasal steps in stimuli. Experimental design was full factorial resulting in a stimuli set in which each vowel step and nasal step combination occurred (total: 49 unique stimuli).

Step	Vowel (ms)	Nasal (ms)
1	110	90
2	135	108
3	160	127
4	185	145
5	210	163
6	235	182
7	260	200

5.1.3. Procedure

The experiment was constructed and presented in OpenSesame (v. 3.1; Mathôt et al., 2012) using a laptop computer. The legacy backend option was selected, which made use of PyGame. Stimuli were presented in randomised order in four blocks that contained 49 trials (i.e., all stimuli occurred once in each block), resulting in each participant completing a total of 196 trials. The first block of trials was originally included as a training block. Its results were consistent with the following three blocks and so responses from the first block have been included in the analysis. Variation due to block is discussed in the results where relevant.

Participants’ categorisation of stimuli was obtained through a forced-choice task in which participants were asked to identify the test stimulus as either waŋa /wɐŋɐ/ “to talk/speak” or wäŋa /wɐːŋɐ/ “home/place.” An image represented each word onscreen (see Figure 6). In each trial, a stimulus was played once, and there was no constraint on the time to respond. Participants made their selection by choosing either the LEFT SHIFT key or RIGHT SHIFT key on the laptop keyboard, which corresponded to the onscreen visual representations of the words. The keyboard prompts remained the same throughout the trials and blocks, as did the arrangement of images onscreen. The participant either made the selection themselves or indicated their selection (by pointing to the left or to the right) to the researcher, who pressed the indicated SHIFT key. Participants used the latter method if they were unfamiliar or not confident with using computers. For this reason, and because participants were not asked to respond as quickly as possible, reaction times are not examined. Responses were collected using the keyboard response collection function in OpenSesame. Between each trial, a fixation dot appeared in the centre of the screen to indicate a new trial was about to commence. Between each block, there was a “break time” screen which allowed participants to have a break of a duration of their choosing. The experiment took approximately 20 minutes to complete, including preliminary discussion about the experiment, signing the consent form, and within-experiment breaks.

Figure 6

Visual representations of waŋa /wɐŋɐ/ “to talk/speak” (left) and wäŋa /wɐːŋɐ/ “home/place” (right) used onscreen in the perception experiment.

Instructions for the task were translated from English to Djambarrpuyŋu by Albena Buyanggirr, a Djambarrpuyŋu speaker and translator. The instructions were presented on the laptop in OpenSesame. The researcher sat with each participant and they read through the instructions together. The participant indicated their willingness to participate in the study and indicated if they had hearing difficulties through buttons onscreen, selected using the laptop trackpad.

Participants were told that the researcher wanted to learn about how Yolŋu (i.e., Indigenous people who speak Yolŋu languages like Djambarrpuyŋu) listened to the sounds in words. Participants were instructed to listen carefully, as each stimulus would be played only one time per trial, and to respond thoughtfully, as the accuracy of their response was important, but not the speed. It was explained that the two SHIFT keys on the laptop keyboard corresponded to the pictures on screen which represented the words waŋa and wäŋa, and upon making a decision, the participant needed to select one of the SHIFT keys to indicate the word they heard. Finally, participants were told the experiment was not a test, and therefore there was no right or wrong answer.

Participants listened to stimuli using Philips SHL3060BK over-ear headphones. Eighteen participants (including the participant whose data were excluded), completed the experiment in an outside location, and two participants were seated inside a house while completing the experiment. Background noises and distractions existed for all participants. If a participant was unable to make a decision or was distracted for a particular trial, the researcher indicated this by using the “P” key to pass the trial. The researcher only used the “P” key following direct instruction from the participant. This response was selected extremely infrequently (n = 6), and those responses were not included in the analysis presented in 5.2.

5.2. Results

Figure 7 shows the proportion of /wɐŋɐ/ and /wɐːŋɐ/ responses to each stimulus, pooled for the 19 participants. Along the x-axis is vowel step, and on the y-axis is the nasal step. Each cell represents one stimulus. The colour of the cell represents the count of /wɐːŋɐ/ responses: A lighter colour reflects more /wɐːŋɐ/ responses, the darker the cell, the fewer /wɐːŋɐ/ responses (i.e., more /wɐŋɐ/ responses). Each stimulus has maximally 76 responses.

Figure 7

Summary of /wɐːŋɐ/ responses to each stimulus pooled for all participants. The more /wɐːŋɐ/ responses to a stimulus, the lighter the cell colour.

Of the 3,718 total valid responses, 1,378 were /wɐŋɐ/ and 2,340 were /wɐːŋɐ/. The responses, therefore, are skewed towards /wɐːŋɐ/ (63%). Possible reasons for this are discussed further in Section 5.3. Overall, participants selected /wɐːŋɐ/ less frequently when a stimulus contained a vowel at step 1 or 2 in the continuum, and more frequently when a stimulus contained a vowel at step 4, 5, 6 or 7 in the continuum. As can be seen in Figure 7, stimuli with vowel step 3 form an intermediate group for which /wɐŋɐ/ and /wɐːŋɐ/ are selected with relatively equal frequency. Nasal step, on the other hand, does not appear to have a consistent effect on listeners’ responses.

5.2.1. Vowel duration

Figure 8 presents the proportion of /wɐːŋɐ/ responses as a function of vowel step. Individual listeners’ response frequencies are plotted by points to provide an indication of the range of responses in the data; points are overlapped so a darker point indicates more overlapping listener responses. There is a sharp categorical boundary between steps 2 and 4, the stimuli either side of which only show a small amount of variation. For stimuli containing vowels of step 5, 6 or 7, there is little sensitivity to the changes in vowel duration. Note also the greater dispersion of individual listener’s responses to the vowel step 3 stimuli compared with other step categories. Taken together, it appears that the crossover point between the short and long vowel categories is around vowel step 3, where participants overall performed at chance (Lehnert-LeHouillier, 2010; Liberman et al., 1957).

Figure 8

Proportion of /wɐːŋɐ/ responses as a function of vowel step. Individual listeners’ response frequencies represented by points.

Post-hoc comparisons between the vowel steps supported the pattern observed in Figure 7 and Figure 8. Comparisons of response to stimuli with vowel duration steps at the higher end of the continuum were not significant: that is, comparisons between vowel steps 5~6, 5~7, 6~7 (p > 0.05) (see Table 10). All other comparisons were significant.

The proportion of /wɐŋɐ/ and /wɐːŋɐ/ responses to each stimulus in each block of trials, pooled for the 19 participants, remained similar for each block, though the statistical analysis showed that there was a significant effect of block on listener’s categorisation (χ²(3) = 28.794, p < 0.0001). Pairwise comparisons revealed that block 1 was significantly different from blocks 2–4, which did not differ significantly from one another. Specifically, more /wɐːŋɐ/ responses were observed in block 1.

Table 10

Selected results from the post-hoc analysis comparing vowel step (n = 3718).

Comparison	z-value	p-value
vowel step
1 ~ 2	–3.950	0.0016
1 ~ 3	–11.775	<.0001
1 ~ 4	–20.789	<.0001
1 ~ 5	–20.646	<.0001
1 ~ 6	–20.405	<.0001
1 ~ 7	–20.316	<.0001
2 ~ 3	–9.048	<.0001
2 ~ 4	–19.555	<.0001
2 ~ 5	–19.231	<.0001
2 ~ 6	–18.965	<.0001
2 ~ 7	–18.871	<.0001
3 ~ 4	–13.509	<.0001
3 ~ 5	–14.292	<.0001
3 ~ 6	–14.117	<.0001
3 ~ 7	–13.948	<.0001
4 ~ 5	–3.686	0.0048
4 ~ 6	–3.803	0.0030
4 ~ 7	–3.474	0.0108
5 ~ 6	–0.176	1
5 ~ 7	0.155	1
6 ~ 7	0.327	1

5.2.2. Consonant duration

Figure 9 presents the proportion of /wɐːŋɐ/ responses as a function of nasal step. It is clear that nasal duration did not have as great of an influence on listeners’ perception compared with vowel duration. However, post-hoc tests showed listeners’ categorisation was significantly affected by nasal step (Table 11). Differences were significant (p < 0.05) for all comparisons that included nasal duration step 7, except 3~7 and 6~7. The comparison between nasal steps 1~6 was also found to be significant. This reflects the decrease in /wɐːŋɐ/ responses observed for nasals at the longer end of the nasal duration step continuum compared with stimuli containing nasals at the shorter end of the continuum in Figure 9.

Table 11

Selected results from the post-hoc analysis comparing nasal step (n = 3718).

Comparison	z-value	p-value
nasal step
1 ~ 2	2.015	0.9215
1 ~ 3	2.986	0.0593
1 ~ 4	1.393	1
1 ~ 5	1.600	1
1 ~ 6	3.913	0.0019
1 ~ 7	4.915	<0.0001
2 ~ 3	1.064	1
2 ~ 4	0.678	1
2 ~ 5	0.409	1
2 ~ 6	2.152	0.6593
2 ~ 7	3.238	0.0253
3 ~ 4	1.744	1
3 ~ 5	1.446	1
3 ~ 6	1.140	1
3 ~ 7	2.196	0.5896
4 ~ 5	0.253	1
4 ~ 6	2.803	0.1064
4 ~ 7	3.908	0.0020
5 ~ 6	2.489	0.2690
5 ~ 7	3.546	0.0082
6 ~ 7	0.954	1

Figure 9

Overall proportion of /wɐːŋɐ/ responses as a function of nasal step. Individual listeners’ response frequencies represented by points.

In Figure 7, it can be seen that for stimuli containing a vowel at either end of the vowel continuum, listeners reliably selected one of the following responses: /wɐŋɐ/ (i.e., short vowel word) when the vowel was at the shortest point in the continuum (vowel step 1), and conversely selected /wɐːŋɐ/ (i.e., long vowel word) when the vowel was at its longest point in the continuum (vowel step 7), irrespective of nasal duration. The vowel step 3 data presents a different case, in which listeners equally often selected a short or long vowel for the first five nasal duration steps. However, at steps 6 and 7, when the nasal was at the two longest points in the nasal continuum, there was a decline in /wɐːŋɐ/ responses (indicated by the darker colour in those cells in Figure 7). That is, when stimuli containing a vowel of step 3 also contained a nasal at the longest or second longest point in the nasal duration continuum, listeners selected the long vowel word /wɐːŋɐ/ approximately 30% of the time. However, the interaction between vowel step and nasal step was not significant (χ²(36, N = 3718) = 38.34, p = 0.36). Nevertheless, there are further trends that can be observed.

Figure 10 presents responses to stimuli that contained nasal step 1 (full black line), nasal step 7 (dashed black line), and all data (thin grey line). If considering stimuli that only contained nasal step 1, listeners’ crossover points between categories occurred earlier than the average across all the data. When stimuli contained a nasal at step 7, listeners’ crossover point between categories occurred later.

Figure 10

Proportion of /wɐːŋɐ/ responses to stimuli containing a step 1 nasal (full black line) and step 7 nasal (dashed black line), as well as across all the data (thin grey line) as a function of vowel step.

5.3. Summary and discussion

It was hypothesised that vowel length perception is categorical in the pair of words /wɐŋɐ/ and /wɐːŋɐ/. The results from the analysis show that listeners’ selection in this forced-choice task were strongly influenced by vowel duration. There was a shift between /wɐŋɐ/ and /wɐːŋɐ/ at around vowel step 3 when considering data pooled across the 19 participants. Vowels at step 3 were ambiguous as to their length category as evidenced by participants performing at chance for stimuli containing those vowels (Lehnert-LeHouillier, 2010). Recall that in the corpus, the duration of the short vowel in /wɐŋɐ/ ranged from 102 to 158 ms (median = 127), and 132 to 286 ms (median = 214 ms) in /wɐːŋɐ/. Therefore, the location for the crossover point between short and long vowel words in perception, around vowel step 3 in the vowel continuum (160 ms), is in line with the production data in terms of the durational ranges for short and long categories.

The duration of the nasal in C2 affected listeners responses in two ways. Firstly, it was hypothesised that the consonant’s duration would have the strongest effect when the vowel’s duration was ambiguous. Figure 7 suggests that consonant duration is used by listeners as a secondary cue to vowel length when vowel duration is in the durationally ambiguous region between short and long (vowel step 3). Secondly, an overall effect of consonant duration on vowel length perception was hypothesised, which is observed by the leftwards and rightwards shift in the vowel category crossover points in Figure 10 for the nasal step 1 and 7 data. Specifically, shorter consonant duration conditioned an earlier crossover between the length categories, while longer consonant duration conditioned a later crossover. There were also fewer /wɐːŋɐ/ responses to stimuli with nasal step 7, even when the vowel was at its longest. Together, these findings suggest listeners made use of secondary cues when the primary cue was ambiguous (Lehnert-LeHouillier, 2010), as well as phonologically redundant phonetic information in their perception of the vowel length contrast in this minimal pair (Whalen et al., 1993). The pattern observed in the figures corresponds to results for Thai (Roengpitya, 2001; see also Abramson & Ren, 1990, on formants) and Norwegian (van Dommelen, 1999), which showed similar movement in the crossover point in vowel categorisation due to the duration of the following consonant.

Because nasal duration did not have a strong effect on listeners’ responses whereas vowel duration showed a pattern consistent with categorical perception, the results of the perception study support the analysis that only vowel length is contrastive in the language. The results from this study do suggest that the inverse lengthening of consonants may perceptually enhance the vowel length contrast for this minimal pair and possibly other words. It is possible that this finding would not be observed in a study comparing words with obstruents. That is because the nasal steps are shorter in duration than the vowel steps in the perception task, and sonorants overall are shorter than obstruents in the language. Both of these factors might mean that the nasal (i.e., sonorant) duration is more of a secondary perceptual cue than if there had been a perceptual comparison between words with obstruents.

With respect to block, recall that it also had a significant effect. One possibility for this result is that listeners became more familiar with the range of durational values contained in the data over time. Inspection of the data showed that listeners’ decisions became more consistent throughout the blocks, especially so for stimuli containing vowel step 1. Overall, responses in block 1 were significantly different from blocks 2, 3, and 4, while blocks 2, 3, and 4 did not differ significantly.

The overall greater proportion of /wɐːŋɐ/ responses may be due to a number of factors, some of which are discussed here. It could be that the continuum did not include vowels of short enough duration. This is supported in the lack of a level state at the lower end of the vowel continuum in Figure 8. Another possibility is that because the original token contained a long vowel and listeners had access to a range of acoustic cues other than duration, these cues conditioned more /wɐːŋɐ/ responses. For example, the small difference in formants may have affected listeners, at least in the trials in block 1, though the manipulated token was within the overlapped areas for short and long vowels (Figure 5). An additional factor that could be utilised by listeners that may be examined in the future is the f0 contour (see e.g., Lehnert-LeHouillier, 2010; Sanker, 2019), which has been found to differ between short and long vowels in Djambarrpuyŋu, but has not been subject to thorough investigation (Jepson, 2019a, pp. 104–109). The presentation of the stimuli could have also been a factor in the skewed responses as the keyboard responses and on-screen images were not counter-balanced throughout the trials or between participants. This may have led to a bias to select the right-side response button (i.e., selecting the long vowel word /wɐːŋɐ/) (Richardson, et al., 2020). These factors will be taken into account in future perception studies with Yolŋu listeners.