Word-level prosodic and metrical influences on Hawaiian glottal stop realization

1.1 Word-level prosodic and metrical influences on the phonetic realization of consonants

Research on the effect of prosodic and metrical boundaries on the realization of consonantal phonetic detail has shown that elements ranging from word-internal boundaries, such as feet or prosodic words, to word edges, to phrase boundaries can influence acoustic and articulatory implementation (e.g., Bombien et al. 2010; Cho 2004; Cho and Keating 2001, 2009; Cole et al. 2007; Fougeron 2001; Fougeron and Keating 1997; Katz 2021; Pierrehumbert and Talkin 1992; Shaw 2007; Sugahara and Turk 2009). Similarly, prosodic prominence, such as syllabic stress or accentual prominence can also affect phonetic implementations of consonants (e.g., Avesani et al. 2007; Bombien et al. 2007; Cho 2006b; Cho and Keating 2009; de Jong 1995; Garellek 2014; Katsika and Tsai 2021; Turk and White 1999). In this paper, we focus on whether prosodic and metrical considerations at the level of the word affect the realization of the phonemic glottal stop in Hawaiian within connected speech. In particular, we focus on whether syllabic stress, prosodic word boundaries, or lexical item (word) boundaries influence the implementation of glottal stop. This study follows up on Davidson (2021), which demonstrated that Hawaiian glottal stops can be produced on a continuum from modal to creaky voicing to a full closure (Davidson 2021; see also Garellek et al. 2021: on the continuum of glottal articulation), but which did not examine whether word-level prosodic factors condition these possible realizations.

The motivation for examining the effect of word-level prosodic and metrical factors on consonant realization comes from research which has argued that in some languages, metrical structure can be signaled by either phonological processes or phonetic implementation. Starting with phonological processes, there is evidence that the foot and the prosodic word are critical domains for explaining the patterning of consonants. For example, it has been argued that the realization of /t/ as a flap in American English is foot-sensitive (e.g., Harris and Kaye 1990; Kiparsky 1979; Turk 1992). Harris (2004) contends that foot structure can account for the phonetic implementation of stops in Danish and Ibibio (see also Akinlabi and Urua 2002); Vaysman (2009) provides a more detailed analyses of Ibibio by arguing that intervocalic consonants cannot lenite if they are at the beginning of a word-internal prosodic word, but crucially, all of the Ibibio accounts attribute the pattern to word-medial metrical boundaries. In order to account for the epenthesis of coda /h/ in Huariapano, Bennett (2013) argues for a maximal foot structure that requires segmental augmentation of the first syllable, whether that syllable is stressed or not. Similarly, Bennett (2018) maintains that the behavior of syllable-initial glottal stops in Kaqchikel is a diagnostic for positing unbounded, iterative recursion of the prosodic words. Vaysman (2009) presents the case of consonant gradation in Nganasan, demonstrating a process of obstruent lenition that occurs only when the consonants are foot-initial, including in word-medial feet, but lenition is otherwise blocked, and it does not correspond to stress placement. Hudu (2022) claims that debuccalization of /k, g, s/ to [ʔ] in Dagbani only occurs when these sounds are foot-medial, and not at the edge of a foot.

While many processes that are used as diagnostics to establish metrical structure in this disparate set of languages are generally considered segmental and are not strictly about the phonetic details of the consonant, it has also been suggested that word-level considerations and metrical structure can also affect how consonants are phonetically realized. For example, at the level of word edge versus word internal phonetic implementation, much of the recent literature on the phonetic properties of consonant lenition (especially in American English) argues that weakening of stops (whether to voiced, spirantized or approximant implementations) cannot be accounted for by a phonological process, but is rather a result of prosodically-controlled changes in gestural duration (e.g., Cohen Priva and Gleason 2020; Katz 2016, 2021; Parrell and Narayanan 2018). In one case where phonetic cues appear to be dependent on foot structure (as opposed to just being word-medial), Shaw (2007) demonstrates that fricative duration in Japanese [tʃ] in foot-initial position, whether word-initial or word-internal, is longer than when it is either foot-medial or extrametrical. Sugahara and Turk (2009) argue that durational lengthening evidence distinguishing Level I (i.e., derivational suffixes such as -al) and Level II (i.e., inflectional suffixes such as -ing, -ed, -s) suffixed words in Scottish English is consistent with analyses that posit different prosodic word structure for these two types of words.

Another word-level factor that has been shown to condition consonant implementation is syllable stress. In languages with aspirated stops such as English, Danish and Lakhota, voice onset time (VOT) has been shown to be longer in stops in the onset of stressed syllables than in unstressed ones (Cho 2006a; Lisker and Abramson 1967; Puggaard 2019; Sprinkle 2022: though not all studies show a significant effect, see Cho and Keating 2009). Cho and Keating (2009) do show that closure duration of both stops and nasal duration consonants in the onset of a stressed syllable in English are longer than for unstressed syllables. In Dutch, the burst duration of the fortis /t/ is longest in unstressed syllables and shortest in stressed syllables, which is interpreted as a fortition effect (Cho and McQueen 2005). A similar effect is found for unaspirated stops in Yakima Sahaptin (Hargus 2005). However, in many other languages with unaspirated stops, such as Spanish, Plains Cree, or Sierra Norte de Puebla Nahuatl, VOT is not affected by prosodic prominence (Hodgson 2021; Kakadelis 2018; Simonet et al. 2014). In some Australian languages, the tonic syllable has an effect on the duration of the following consonant. For example, in Warlpiri and Mawng, C₂ in a word with a shape like ˈC₁VC₂V has a longer duration than the pre-tonic consonant, or consonants that follow an unstressed syllable (Fletcher et al. 2015; Pentland 2004). In a different Australian language, Pitjantjatjara, there are differences in spectral center of gravity in stop bursts depending on whether they are in stressed or unstressed syllables, which is interpreted as an enhancement of the phonological feature [grave] in stressed position (Tabain and Butcher 2015).

In this study, we return to the question of whether word-level metrical and prosodic elements condition the phonetic realization of consonants, specifically with respect to phonemic glottal stops in Hawaiian. In the study of Hawaiian that is the precursor to the current study, Davidson (2021) found that the realization of phonemic glottal stops is conditioned by word position and by whether the flanking vowels are identical or different. In Hawaiian, glottal stops can occur either in intervocalic position within a word (e.g. [poʔe] ‘people’), or contrastively at the beginning of the word (e.g. [ʔaka] ‘laughter’ vs. [aka] ‘shadow’). Full closures were rare overall, accounting for only 7 % of the results. For tokens that are produced with creaky voice, the study distinguished between possible locations and extent of creaky voice: creak between two modal vowels, creak on V1, creak on V2, or creak over the whole vowel-vowel interval. Longer periods of creak, including whole creak and creak on V2, are less likely in word initial position, and whole creak occurs at greater rates when the flanking vowels are identical. Full closures were less likely between identical flanking vowels, but there was no effect of word position on the realization of full closures. However, Davidson (2021) did not include any further metrical or prosodic information beyond word position, such as syllable stress, prosodic word position, or vowel weight which may affect the realization of glottal stop.

Previous work on the factors affecting the realization of phonemic glottal stop is sparse, despite Ladefoged and Maddieson’s (1996: 75) famous quote: “In the great majority of languages we have heard, glottal stops are apt to fall short of complete closure, especially in intervocalic positions.” While DiCanio (2012) does not examine prosodic information in his study of Itunyoso Triqui, he does find a difference in implementation depending on whether the glottal stop is intervocalic within a word, or in coda position at the end of the word. Whereas the intervocalic glottalization is almost never produced as a full closure, the coda glottal stop is more likely to be produced as a full glottal closure. In Mayan languages, such as Uspanteko, the realization of glottal stop as full closure or a period of creaky voice in part depends on whether it is adjacent to a consonant or intervocalic (Bennett et al. 2022). After a consonant, it is frequently a full closure. A study of Cook Islands Māori finds that in conversational speech, the realization of glottal stop depends on dialect. In two of the five dialects that were examined, a period of creaky voice alternated with full closure fairly evenly, while two other dialects had larger rates of no closure or creakiness at all, and the last dialect had mostly full closures (Nicholas and Coto-Solano 2019).

One language in which the realization of glottal stop has been studied in some detail is Maltese. Mitterer (2018) finds that geminate glottal stops are produced as full closures about 80 % of the time, versus less than 20 % for singletons. While it has been assumed that geminates have prosodically greater weight than singletons, which may be predicted to have stronger articulations (i.e. full closures), Mitterer also finds that the influence of phonological quantity cannot be disentangled from the finding that geminate glottal stops also have a significantly longer duration. Mitterer et al. (2019) compares the realization of phonemic glottal stop to non-phonemic glottalization in word-initial (but phrase-medial) position in laboratory speech in Maltese, where the target word followed either a word ending in /a/ or in /m/. In this type of elicitation, phonemic glottal stops are realized as full closure nearly 80 % of the time after /m/, and 50 % after /a/. As for non-phonemic cases, about half are marked with glottalization, and full closures make up about 40 % of the responses after /m/ even for non-phonemic glottalization. Though there is not an explicit manipulation of boundary type in this study, duration of the previous word is used as a proxy for prosodic strength, and there is a correlation between length of the boundary and the likelihood that glottalization is realized as a full closure.

Beyond the level of the word, previous research has shown that glottalization can be used non-phonemically to demarcate prosodic boundaries and indicate degree of strength. In English, word-initial glottalization occurs optionally for vowel initial words, and is more likely to be found when the vowel is stressed than when it is unstressed (Davidson and Erker 2014; Dilley et al. 1996; Pierrehumbert 1995; Pierrehumbert and Talkin 1992). Similar findings regarding the effect of stress have been reported for several varieties of Spanish (Chappell 2013; Michnowicz and Kagan 2016; Trawick and Michnowicz 2019), German (Kohler 1994; Malisz et al. 2013; Pompino-Marschall and Żygis 2011) and Polish (Malisz et al. 2013; Schwartz 2013). Malisz et al. (2013) also report specifically on how the glottal stop is realized, showing that full closures are more likely both in phrase-initial position than medial position, and before stressed vowels in both German and Polish. Using articulatory data, Garellek (2014) found that for sentences produced in a laboratory context, electroglottograph (EGG) contact quotient patterns for both English and Spanish speakers were more consistent with glottalization when a vowel at any prosodic boundary was stressed. In this case, the different degrees of prosodic boundary (utterance initial, IP-initial, ip-initial, and ip-medial) did not lead to significant differences in contact quotient rates.

Taken together, the studies of both phonemic glottal stops and non-phonemic glottalization reveal that in many languages, glottalization can be manipulated in service of indicating prosodic boundaries or prominence. Cross-linguistically, the rates of glottalization vary quite a bit by language, from rarely (e.g., Yucatecan Spanish) to almost obligatorily in some environments (e.g., Polish or German). Likewise, the specific implementation—whether full closure or glottalization—can also be affected by prosodic factors. Garellek et al. (2021) discuss realization of glottalization gestures as part of a phonatory continuum from most open glottis (aspirated or breathy) to most closed (creaky voice or full closure) (see also Gordon and Ladefoged 2001). They argue that while the representation of these gestures may be generally specified for aspiration (as in sounds with breathy voice quality) or glottalization, they are not necessarily specified for the magnitude of the gesture, which can then be determined by other pressures, such as those governing prosodic strengthening (e.g., Cho 2005; Cho and Keating 2001; Fougeron 2001; Fougeron and Keating 1997; Keating et al. 2003) or reduction/lenition (Cohen Priva and Gleason 2020; Gurevich 2011; Katz 2016, 2021; Lavoie 2001).

In the context of the current study of Hawaiian glottal stops, if full closure is considered the extreme of the glottalization end of the continuum, then one possibility is that full closure will be more likely when pressure from prosodic prominence is exerted, such as in the onset of a stressed syllable, or in word initial position, as compared to word medial and unstressed position. To date, the motivation for this possibility comes mainly from the insertion of non-phonemic glottalization, which may behave differently, though evidence from Maltese may also suggest that these types of effects are at play for phonemic glottal stops in that language (Mitterer 2023). However, another possibility is that realization of the phonemic glottal stop could instead reflect the word-internal metrical structure of Hawaiian words, as has been shown for languages like Kaqchikel, Nganasan, Ibibio, Dagbani, Japanese, or Scottish English. These two potential influences on phonetic realization are investigated in this study.

1.2 Glottal stops and metrical structure in Hawaiian

Ka ‘Ōlelo Hawaiʻi, or Hawaiian, is an Eastern Polynesian language spoken today across the eight major Hawaiian Islands, located in the North Pacific Ocean. Hawaiian is currently in the midst of a successful language revitalization effort, following its near demise after the United States military and American businessmen overthrew the Hawaiian Kingdom in 1893 and the new government they imposed banned the use of Hawaiian in schools in 1896 (Wilson 2013). Before this, Hawaiian was the primary lingua franca of the Islands, used broadly in all domains including government, business, and education, as well as at home. After this point, Hawaiian language use began a precipitous decline, and eventually community and family language transfer diminished on most of the Hawaiian islands (Brenzinger and Heinrich 2013; Wilson 2013). Milestones in the revitalization of the Hawaiian language include: (1) the Hawaiian language radio show Ka Leo Hawai‘i (KLH) which ran from 1972 to 1988 and which featured interviews with elders who had learned the language via intergenerational transfer; (2) the establishment of the ʻAha Pūnana Leo preschools in 1983; (3) the lifting of the 1896 ban on Hawaiian as a medium for teaching in the public schools in 1986; and (4) the establishment of Ka Papahana Kaiapuni, a Hawaiian-medium program in the public schools in 1987 (Kawai’ae’a et al. 2007; Warner 2001; Wilson and Kamanā 2001). Today, Modern Hawaiian-medium education can be found from pre-school through PhD-granting programs at the University of Hawaiʻi. Counting the number of Modern Hawaiian speakers today is difficult, but according to the American Community survey from 2009 to 2013, 18,610 respondents reported speaking Hawaiian at home (U.S. Census Bureau, 2009–2013). Since little previous linguistic work has been carried out on either pre-revitalization Hawaiian or Modern Hawaiian, the similarities and differences between the sound systems of these varieties remains largely unknown.

Hawaiian has phonemic glottal stops before vowels in word-initial and word-medial position. Hawaiian has a comparatively small consonant inventory, shown in (1), and a relatively large set of vowels (both short and long) and diphthongs, shown in (2–4) (Elbert and Pukui 1979; Kettig 2021; Ladefoged 2001; Parker Jones 2018; Pukui and Elbert 1986; Schütz 1981):

As noted above, in word initial position, glottal stop is contrastive. In word-medial position, glottal stop can occur before or after any kind of vowel, whether it is monophthongal, diphthongal, short, or long. Examples of various combinations of vowels flanking a glottal stop are provided in (5) and (6), though these options are not exhaustive.

In the stress system of Hawaiian, syllables with diphthongs and long vowels are heavy, and all heavy syllables are stressed, either with primary or secondary stress (Parker Jones 2018; Schütz 1978, 1981; Senturia 1998). The patterns and words in (7) illustrate the six templates that are found for words that only have one stress in Hawaiian (Parker Jones 2010; Schütz 1981, 2010). Parker Jones (2010) considers these templates, first proposed by Schütz (1981), to be possible prosodic word shapes in Hawaiian; we return to this point below.

The Hawaiian stress pattern has been compared to the right-to-left moraic trochee analysis of Fijian (Alderete and MacMillan 2015; Hayes 1995; Kenstowicz 2007), but this analysis is complicated by the fact that long words composed of more than three light syllables do not follow a predictable stress pattern (Parker Jones 2010; Schütz 1978, 1981, 2010). Famously, Hawaiian allows words like [ˌʔe.le.ma.ˈku.le] ‘old man’ and [ma.ˌku.a.ˈhi.ne] ‘mother’, where ‘old man’ has stress on the first syllable, and ‘mother’ has stress on the second syllable. It is typically agreed that the rightmost stress in a word is primary. Whereas Hayes (1995) treats any words in Fijian that did not match his right-to-left trochaic analysis as lexical exceptions, this is not plausible for Hawaiian because the words with five light syllables split about evenly between these two possible stress patterns (Parker Jones 2010; Pukui and Elbert 1986).

Parker Jones (2010) develops a probabilistic context free grammar to determine the likelihood of the prosodic structure of Hawaiian words. Similar to de Lacy’s (2004) analysis of Māori, Parker Jones implements both prosodic words and metrical feet in the parser to capture the analysis of Hawaiian stress, where potential prosodic words correspond to the template shapes in (7) (marked throughout the paper in curly brackets). Metrical feet are bimoraic, consisting of two syllables or one heavy syllable, where the leftmost syllable in the foot is stressed, as shown in (8).

For the purposes of stress assignment, when heavy and light syllables in words with more than three syllables are mixed, the correspondence to the prosodic word shapes in (7) can be ambiguous (Parker Jones 2010). For example, a word consisting of σ_H σ_L σ_H could be either {ˌσ_H σ_L}{ˈσ_H} or {ˌσ_H}{σ_Lˈσ_H} but this can be considered a spurious ambiguity because stress assignment is the same in both cases. However, other combinations of syllables lead to more than one possibility for stress assignment. For example, the sequence σ_H σ_L σ_L σ_H could be realized either as a word with two stresses {ˌσ_H σ_L}{σ_L ˈσ_H}, as in [{ˌmu:.na}{mu.ˈna:}] ‘to eat sparingly’ or with three stresses {ˌσ_H}{ˌσ_L σ_L}{ˈσ_H}, as in {ˌʔo:}{ˌko.le}{ˈhao} ‘moonshine’.

While most analyses of stress in Polynesian languages such as Māori, Samoan, or other accounts of Hawaiian, allow that lexical words can contain one or more prosodic words to ensure appropriate stress assignment (Alderete and MacMillan 2015; de Lacy 2004; Senturia 1998; Zuraw et al. 2014), these analyses are typically focused on reduplication, compounds or affixation which allows them to straightforwardly posit that each morphological constituent contributes and is associated with its own prosodic word. However, Parker Jones (2010) and Schütz (2010) both argue that in Hawaiian, there is not a strong correspondence between prosodic words and morphological constituents, and that a word can contain multiple prosodic words that do not neatly match up to morphological boundaries. In some cases, such as /napoʔo-na/ ‘to set-nmlz’, the nominalizing suffix /-na/ must form part of a prosodic word with material from the stem (i.e. {ˌna.po}{ˈʔo.na}, since the suffix [na] does not meet the minimal word requirement for Hawaiian (Alderete and MacMillan 2015; de Lacy 2004). In other cases, however, there are words with more than three syllables that do not seem to have obvious morphological structure which still require prosodic word boundaries in order to ensure the appropriate stress placement, such as {(ˌpu.le)}{le(ˈhu.a)} ‘butterfly’ (cf. *{pu(ˌle.le)}{(ˈhu.a)}).

We follow Parker Jones (2010) and Schütz (2010) in assuming that multiple prosodic words in a single lexical item can be necessary to capture the observed stress patterns of Hawaiian, whether a word is morphologically complex or not. While we leave it for future work to present a full analysis of the prosodic structure of monomorphemic and multimorphemic words in Hawaiian, and in comparison to other Polynesian languages, we note that other research on Polynesian and Austronesian languages has also speculated about the appropriate higher prosodic level for a lexical item that contains multiple prosodic words. For example, Brown (2015) proposed that the Phonological Phrase is the next level above the prosodic word in Māori, capturing both utterances like [e#{no.ho} _PrWd]_P-Phrase ‘imp-sit’, where the first element does not meet the minimal word requirement, as well as what are considered to be single lexical items such as [{ta.ko.to} _PrWd]_P-Phrase ‘sit down!’. In Samoan, Zuraw et al. (2014) proposed that words that are considered single lexical items can be composed of multiple prosodic words, but they “leave open the question of whether the two prosodic words in (a) combine to form a larger prosodic word, or attach directly to the next level up, such as a Phonological Phrase” (p. 285). Bennett (2018) develops a recursive prosodic word analysis for Kaqchikel in order to account for patterns of glottal stop insertion and degemination. Providing a prosodic analysis of Hawaiian that includes levels above the prosodic word is beyond the scope of this study, so for the purposes of this paper, we will use the cover term “lexical word” or “lexical item” (abbreviated as “word” in the diagrams in (9)–(13), where it is contrasted with PrWd) to account for the dictionary entries/orthographic words in Hawaiian that can contain one or more PrWds.

The goal of this study is to determine whether word-level elements that have been established for the prosodic and metrical system of Hawaiian to date, including lexical word boundaries, prosodic word boundaries, stressed versus unstressed syllables, and heavy versus light syllables, affect whether phonemic glottal stops in Hawaiian are implemented as full closures or as creaky voice. This study uses the data reported in Davidson (2021) to determine whether full closures also correspond to prominent prosodic environments, as they are more likely to for non-phonemic uses of glottalization, or whether the word-internal metrical structure affects realization. We hypothesize that stressed syllables may condition the presence of full closure. As for word position, the analysis in Davidson (2021) did not find greater rates of full closure at the beginnings of words, but this variable is examined here again because it is included in other studies examining prosodic effects on glottalization. The prosodic word boundary is another element which may interact with glottal stop realization, especially since a phonetic cue to prosodic word position may assist listeners in determining the metrical structure of a word that has multiple potential parses. We do not independently examine foot-initial boundaries in the analysis; they are conflated with stress, since the left edge of a foot boundary always coincides with a stressed syllable.

2.1 Participants

The spontaneous speech data used in this study are the same as in Davidson (2021). The talkers are 8 (4 M, 4 F) bilingual speakers of Hawaiian and English who appeared on a radio show called Ka Leo Hawaiʻi in 1972 or 1973. The purpose of the radio show, as explained by the host Larry Lindsey Kimura, was to interview mānaleo, or native speakers of Hawaiian (http://ulukau.org/kaniaina/?a=p&p=history). While the speakers were elderly, an effort was made to choose those speakers who did not have obvious aging-related general creaky or breathy voice (Gorham-Rowan and Laures-Gore 2006; Xue and Deliyski 2001). Transcripts for these speakers are available on the ulukau.org website.

Recordings were made in the studios of KCCN radio in Hawaiʻi on reel-to-reel tape, which were then digitized to mp3s for uploading to the internet. The speakers in this study were in the studio with the host, and not calling in on the telephone. The interview topics ranged from place of birth to families, education, occupations, hobbies such as fishing or sailing, Hawaiian customs such as feasts or making leis, and music. Typically, the host asked questions and provided the speakers with ample time for uninterrupted responses. The interviews ranged in length from 5 to 16 min, with an average of 11 min. The number of glottal stops per speaker is found in Table 1.

2.2 Materials

Glottal stops were identified via the ‘okina in the orthography in the transcript (e.g., Hawai‘i /havaiʔi/, po‘e /poʔe/ ‘people’, ‘umi /ʔumi/ ‘ten’). It has been noted that an ‘okina in the orthography does not necessarily mean that a word is currently pronounced with a glottal stop (Romaine 2002), so any words containing the ‘okina that were not produced with any phonetic instantiation of the glottal stop (as defined by the criteria laid out below) by any of the speakers were removed; this led to the elimination of 2 words (2 total utterances).

The realization of the /ʔ/ phoneme was determined by looking for the presence of a full glottal closure, creaky voice, or other indicator on a spectrogram. The visual criteria we followed were based on descriptions of glottalization events in various previous studies, including Keating et al. (2015), Redi and Shattuck-Hufnagel (2001), Avelino et al. (2011), Gerfen and Baker (2005), Davidson and Erker (2014), and others. The consensus from these papers is that glottalization realized as creaky voice is characterized by a low and irregular fundamental frequency, and damping of the glottal pulses. The focus of Davidson (2021) was on the duration and timing of the glottalization event, so a number of different realizations of creaky voice were catalogued, including a modal voice-creaky voice-modal voice configuration (mid-creak), creak only on the first vowel (V1 creak), creak only on the second vowel (V2 creak), and creak on the whole vowel-vowel interval (whole creak). In other tokens, the glottalization event was realized with a period of complete closure. In a small number of cases, there was no evidence of creaky voice, but rather the token was modally voiced with a dip in intensity that was sometimes accompanied by at least a slight dip in F0. Lastly, some tokens were realized without any acoustic or perceptual evidence of a glottal stop (though other instances of these words were produced by other speakers with glottalization, as discussed above regarding the ‘okina).

In this study, we are not focused on the timing of glottalization, so all of the tokens realized with creaky voice (mid-creak, V1 creak, V2 creak, whole creak) are grouped together into a single category of “Creaky Voice” tokens (N = 502, 66 %). The second category we consider is “Full Closure” (N = 52, 7 %), and the third is “Modal/Intensity Dip” (N = 205, 27 %). The third group includes tokens that were produced with modal voicing only, including those that had a dip in intensity. While an intensity dip in the absence of nonmodal phonation or a full closure could be a cue to a phonemic glottal stop in Hawaiian, there is currently no evidence to confirm this, so all modally phonated tokens are grouped together for this study. Examples of the creaky and full closure realizations are provided in Figure 1 (images all of the creaky realizations are provided in Davidson 2021).

Figure 1:

Different categorizations of glottal stop realization. The glottal stop symbol [ʔ] in the segmentation in each spectrogram indicates the portion corresponding to the different realization type. In the creaky voice realization with whole creak, [ʔ] is included in the interval with the vowel since it is difficult to isolate a portion corresponding solely to the glottal stop. The phonetic transcription in the caption is a closer indication of how the target interval is produced. (a) ‘Full glottal closure’ realization in he ʻumi [he#ʔumi] (‘ten are’). (b) ‘Creaky voice’ realization (modal-creaky-modal) in Heʻeia [heʔ̰eia] (place name) ([ʔ̰] indicates creaky voice flanked by two modal vowels). (c) ‘Creaky voice’ realization (whole creak) in hoʻokahi [ho̰o̰kahi] (‘one only, alone’). (d) ‘Intensity dip’ realization in hoʻokipa [hookipa] (‘hospitality’).

2.3 Prosodic and metrical feature extraction

In addition to identifying whether the glottal stop phoneme is realized as full closure, creaky or modal/intensity dip, we developed a semi-automated pipeline to categorize each word containing a glottal stop for the following properties:

1. Before an unstressed, primary stressed, or secondary stressed syllable: unstressed: [ˈpo.ʔe] ‘people’; primary stressed: [o.ˈʔa.hu] (place name); secondary stressed [ˌʔu.mi.ˌku:.ˌma:ˈi.va] ‘nineteen’

2. Onset of a light or heavy syllable (long vowel or diphthong): light: [ʔe.ˈlu.a] ‘two’; heavy: [ˌʔoi.ˈha.na] occupation

3. Prosodic word initial or prosodic word medial: PrWd-initial: [{ˌha:}{ˈʔa.vi}] ‘to give, grant’; PrWd-medial [{ˈlu.ʔu}] ‘to dive’

4. Lexical word initial or lexical word medial: initial: [ʔo.ˈha.na] ‘family’; medial: [ˈa.ʔo] ‘instruction, teaching’

There were three parts to the pipeline. The first part used a variant of the computational prosodic grammar developed in Parker Jones (2010) to parse the words into their prosodic constituents. A description of the grammar is provided in Appendix A. To illustrate, the place name Oʻahu was normalized and then parsed as in (9). Text normalization, where uppercase letters are converted to lowercase, results in a one-to-one correspondence between symbols in the modern Hawaiian orthography and Hawaiian phonemic analysis (Parker Jones 2010, 2018; Parker Jones and Shillingford 2018). Thus in (9) we show the parse tree for the text-normalized word oʻahu /oʔahu/, followed by the associated bracketed parse which will be important for the next part of the pipeline. In the following parses, SYLL_L and SYLL_H refer to light and heavy syllables, respectively. Stress is not marked by the parser because it is predictable based on foot assignment: the leftmost syllable in a foot is stressed, and the rightmost stress in the word is the primary stress.

(9)

Parse for O‘ahu {o(ˈʔa.hu)}

(WORD (PRWD (SYLL_L (SHORT_VOWEL o)) (FOOT (SYLL_L (CONSONANT ʻ) (SHORT_VOWEL a)) (SYLL_L (CONSONANT h) (SHORT_VOWEL u)))))

Similar bracketed parses for ʻoihana ‘occupation’, hāʻawi ‘to give’, and ʻohana ‘family’ are given in (10), (11), and (12), respectively.

For the second step in the pipeline, we intervened manually to resolve ambiguous parses, as in (13), which shows the two parses for the word ʻelemakule ‘old man’.

(13)

	Two parses for ʻelemakule
a.	{(ˌʔe.le)}{ma(ˈku.le)}
	(WORD (PRWD (FOOT (SYLL_L (CONSONANT ʻ) (SHORT_VOWEL e)) (SYLL_L (CONSONANT l) (SHORT_VOWEL e)))) (PRWD (SYLL_L (CONSONANT m) (SHORT_VOWEL a)) (FOOT (SYLL_L (CONSONANT k) (SHORT_VOWEL u)) (SYLL_L (CONSONANT l) (SHORT_VOWEL e)))))
b.	*{ʔe(ˌle.ma)}{(ˈku.le)}
	(WORD (PRWD (SYLL_L (CONSONANT ʻ) (SHORT_VOWEL e)) (FOOT (SYLL_L (CONSONANT l) (SHORT_VOWEL e)) (SYLL_L (CONSONANT m) (SHORT_VOWEL a)))) (PRWD (FOOT (SYLL_L (CONSONANT k) (SHORT_VOWEL u)) (SYLL_L (CONSONANT l) (SHORT_VOWEL e)))))

Although ambiguities such as these can be resolved in a fully-automatic way, with over 96 % accuracy on unseen data (Parker Jones 2010), we intervened manually in this study to ensure 100 % accuracy. In cases of “spurious ambiguity”, where multiple parses result in the same stress assignment (Parker Jones 2010; Schütz 1981), ambiguity was resolved by following the citation form in the standard Hawaiian dictionary (Pukui and Elbert 1986). An example is the word hōʻoaka ‘to open’, which is parsed by the grammar as either {(ˌho:)}{ʔo.(ˈa.ka)} or {(ˌho:.ʔo)}{(ˈa.ka)}. We follow the dictionary in parsing this word as {(ˌho:)}{ʔo.(ˈa.ka)}.

For the final step in the pipeline, we used regular expression matching to find which bracketed word parses contained a glottal stop phoneme in the following contexts: (a) in a primary or secondary stressed syllable; (b) as the onset of a heavy syllable; (c) in the initial syllable of a PrWd; and/or (d) in the initial syllable of the lexical word. As the initial syllable in a foot is always stressed, case (a) also describes the initial syllable in a metrical foot.

The first feature to extract was whether the syllable with the glottal stop was stressed or not. Words that satisfy this constraint matched one of the two patterns in (14):

Using regular expression matching, where .* matches any substring, the bracketed parses that contain at least one of the two patterns in (14) were classified as glottal stop in stressed syllable. If there was more than one stressed syllable in a word, then the rightmost one was classified as primary stressed, and the other(s) as secondary stressed. All other words were classified as not having a glottal stop in a stressed syllable.

The second feature we wanted to extract concerned heavy syllables:

There is a close relationship between the patterns in (14) and (15), as (15) is a substring of the pattern in (14b). This follows from the fact that heavy syllables are always stressed (Parker Jones 2010, 2018), but stressed syllables are not always heavy – hence, the pattern in (14a). The pattern in (15) will thus match the parse for {ˌʔoi}{ˈha.na} but not for {ˌha:}{ˈʔa.vi}. In one case, the glottal stop is in a heavy syllable ([ʔoi]), in the other, it is in a light syllable ([ʔa]).

To find cases where the phonemic glottal stop was in the initial syllable of a PrWd, we used the patterns in (16):

These three patterns account for cases where the initial syllable is unfooted (e.g., {ʔo(ˈha.na)}), as well as for cases in which the initial syllable of the foot is light (e.g., {(ˌha:)}{(ˈʔa.vi)}) or heavy (e.g., {(ˌʔoi)}{(ˈha.na)}).

Finally, word initial phonemic glottal stops were found using the patterns in (17):

These are almost identical to the patterns in (16), but with the further word-level constraint. The patterns in (17) will match bracketed parses for words like {ʔo(ˈha.na)} and {(ˌʔoi)}{(ˈha.na)}, but not {(ˌha:)}{(ˈʔa.vi)} (see bracketed parse in (11)).

To summarize, the result of the pipeline was to generate prosodic features for each word, where each word was normalized and parsed, disambiguated if necessary, and then matched against a set of regular expressions. This resulted in tabular data for whether a phonemic glottal stop was in a stressed syllable, heavy syllable, PrWd-initial, and/or (lexical) word initial, as illustrated in Table 2.

2.4 Word frequencies

Frequencies for words in the transcripts were obtained using an independent modern Hawaiian text corpus of about 1.6 M words (Parker Jones and Shillingford 2018). This text corpus was normalized and tokenized in Python, using NLTK’s ‘RegexptTokenizer’ (Bird et al. 2009). Tokens were excluded if they contained numerals (0 1 2 3 4 5 6 7 8 9) or punctuation (! " & ’ ( ) , - . / : ; ? [ ] –). All word counts were adjusted using add-1 smoothing before computing base-10 log probabilities, to avoid taking logs of zero.