Northern Raglai voicing and its relation to Southern Raglai register: evidence for early stages of registrogenesis

Abstract

Northern and Southern Raglai are two closely related Austronesian dialects (Chamic branch) spoken in south-central Vietnam. Although they are mutually intelligible, Northern Raglai is described as having a voicing contrast in onset stops, while Southern Raglai is assumed to have replaced the Chamic voicing contrast with a register contrast realized on the whole syllable (but primarily on its vowel). A production study of the two dialects confirms that Northern Raglai preserves a voicing contrast, even if most women exhibit partial devoicing of their voiced stops, and that Southern Raglai has developed a register contrast based on F1 and phonation cues at the beginning of vowels. The weights of the acoustic properties of voicing and register are similar across ages and genders, suggesting that the two laryngeal contrasts are phonetically stable. Even if there is little evidence of change in progress, a close inspection of the Northern Raglai voicing contrast reveals voicing-conditioned modulations of F1 and perturbations of phonation after partially devoiced stops that could be precursors of a register system similar to that of Southern Raglai. We argue that this is a pathway to registrogenesis and speculate about the articulatory laryngeal mechanisms that could trigger registrogenetic changes. Our data also show that the Northern Raglai voicing contrast is unstable in aspirated stops and that voiced aspirated stops typically have a partially voiceless and partially voiced aspiration.

1. Introduction

Raglai is an Austronesian language of the Chamic branch that is spoken in the provinces of Khánh Hòa, Ninh Thuận and Bình Thuận, in south-central Vietnam. While Northern Raglai has been described as preserving a voicing contrast in obstruents (Awơi-hathe et al. 1977; Lee 1966; Tạ 2009), Southern Raglai is reported to have transphonologized obstruent voicing into a typical Southeast Asian register contrast cued by acoustic properties like pitch, phonation and vowel quality (Tạ 2009). The fact that two closely related and mutually intelligible varieties of the same language have come to develop such different realizations of Chamic proto-voicing represents a unique opportunity to understand the phonetic underpinnings of register formation, or registrogenesis, as it could allow us to determine if register is rooted in intrinsic secondary phonetic properties of voicing.

1.1. Register and registrogenesis

Register is a type of phonological contrast commonly attested in Austroasiatic and Austronesian languages of Southeast Asia (Ferlus 1979; Gregerson 1976; Henderson 1952; Huffman 1976). Diachronically, it usually derives from the loss of a voicing contrast in onset obstruents, but it is synchronically realized as a combination of pitch, phonation and vowel quality over syllables or rhymes. Typical register languages oppose a high register, associated with a high f0, a modal phonation and more open vowels, to a low register associated with a lower f0, a breathy or lax phonation and more close vowels. They also often keep remnants of the original voicing contrast: voiced obstruents preserve optional closure voicing in some languages, like Mon, Chrau and Chru (Blagden 1910; Brunelle et al. 2020; Tạ et al. 2022), while in other languages, like Nyah Kur, Lamet, Kuy, Bru and Chanthaburi Khmer, closure voicing is lost and replaced with aspiration (Ferlus 1979; Haudricourt 1965; Huffman 1976; Wayland and Jongman 2002).

These acoustic properties may not all be present or have the same relative weight in all register languages, and they are often maximally distinct at the beginning of the rhyme. An illustrative minimal pair recorded from a Southern Raglai man with a robust register contrast is given in Figure 1.

Figure 1:

Illustration of register contrast in the Southern Raglai minimal pair /mta/ ‘eye’ ∼ /mt̥a/ ‘rich’ as produced by a man born in 1989 (the subscript circle under the /t/ marks the low register). F1 starts lower and F2 higher at the beginning of the vowel of /mt̥a/, and the low register has a slightly weaker relative amplitude in upper frequencies at vowel onset. VOT is also longer in the low register syllable. Contrary to what is attested in other languages, there is no clear f0 difference (red trace) between registers.

A first question that arises when trying to understand the development of register, or registrogenesis, is the motivation for the loss of voicing in obstruents. Why is it so frequent, not only in Mainland Southeast Asian languages (Haudricourt 1965; Matisoff 1973 are seminal references on the issue), but also in typologically unrelated and geographical distant languages (Coetzee et al. 2018; Howe 2017; Pinget et al. 2019 for recent examples)? The most common explanation is the aerodynamic voicing constraint (AVC) or the fact that it is difficult to maintain a continuous transglottal airflow during obstruents because their constriction causes a build-up in supraglottal air pressure (Ohala 1983, 2011).

A second important question is why certain acoustic properties are more likely to take over the contrastive role of obstruent voicing as it is neutralized. The most frequent account is that some of these phonetic attributes are intrinsic coarticulatory properties of onset voicing (Ferlus 1979; Haudricourt 1965; Huffman 1976; Hyman 1976). It has long been recognized, for instance, that f0 is higher at the onset of vowels following voiceless than voiced obstruents (Dmitrieva et al. 2015; Hanson 2009; Kirby and Ladd 2016; Lisker 1986; Ohde 1984; Rousselot 1901–1908). It is also well-established that vowel F1 is lower immediately after voiced stops than after voiceless stops (Esposito 2002; House and Fairbanks 1953; Stevens and House 1963) and there is some evidence that F2 is slightly lower after voiced stops (Cole et al. 2010). A phonetic property like phonation, on the other hand, is more problematic. There is no indisputable mechanism that would explain why breathiness/laxness, and sometimes aspiration, develop in syllables originally headed by voiced stops. One possible explanation is that pharyngeal expansion aimed at overcoming the aerodynamic voicing constraint may stretch the aryepiglottal folds and indirectly force a slight glottal opening (Kingston et al. 1997). Another is that a lowering of the larynx to decrease supraglottal pressure boosts the transglottal pressure drop, causing aspiration or breathiness (Ferlus 1979; Thurgood 2002). A third possible explanation is that the vocal folds may be slackened during the closure phase of voiced stops to facilitate vocal fold vibrations, leading to some coarticulatory breathiness on following vowels (M. Garellek, p.c.). However, as far as we know, these effects have not been instrumentally studied, and the little evidence we have suggests that in languages in which voiceless stops are realized with aspiration, like English and German, it is these stops that are associated with breathiness/laxness rather than the voiced series (Löfqvist and McGowan 1992; Ní Chasaide and Gobl 1987, 1993). In languages in which voiceless stops are not strongly aspirated, like French and Italian, no phonation differences were found on vowels following voiced and voiceless stops (Ní Chasaide et al. 1993). An association between breathiness and stop voicing is attested in Eastern Armenian, but its plain prevoiced stops may derive from a more complex laryngeal setting (Seyfarth and Garellek 2018). A similar association was established experimentally in Tamang and Shanghainese, but the fact that voicing and phonation vary independently across speakers in both languages suggests a non-automatic relation (Gao 2015; Michaud and Mazaudon 2008). In the end, despite an undeniable diachronic connection between voicing and phonation, available crosslinguistic data does not suggest that vowels are systematically laxer or breathier after voiced than voiceless obstruents.

1.2. The Raglai language

According to the Vietnamese census, there were 146,613 ethnic Raglai in 2019. Besides a few hundred individuals raised in other areas of Vietnam, they can all be assumed to be native speakers of the Raglai language (Central Population and Housing Census Steering Committee 2020). A breakdown by province shows that Raglai is primarily spoken in the Vietnamese provinces of Khánh Hòa (57,143 people), Ninh Thuận (70,366 people) and Bình Thuận (17,382 people). The geographical distribution of ethnic Raglai in south-central Vietnam is reported in Figure 2.

Figure 2:

(A) Geographic distribution of Raglai speakers in Ninh Thuận province and neighboring areas; (B) Geographic distribution of Raglai speakers in south-central Vietnam (black box shows area of map A); (C) Distribution of Raglai speakers in Vietnam (black box shows area of map B). The commune is the lowest rural administrative subdivision in Vietnam.

Although the official Vietnamese ethnic classification does not distinguish Raglai groups/dialects, there is significant evidence of dialectal differentiation. A northern dialect, grouping all varieties spoken in Khánh Hòa and those spoken in northern Ninh Thuận, seems well-established (Awơi-hathe et al. 1977; Lee 1966; Tạ 2009). If one adds up all the Raglai speakers living in the province of Khánh Hòa and the district of Bác Ái in Ninh Thuận, this dialect would have about 85,000 speakers. A less well-defined southern dialect that would comprise varieties spoken in southern Ninh Thuận and Bình Thuận is also normally assumed, even if Tạ (2009) suggests that it could be quite heterogeneous. It is unclear, for instance, if the dialect spoken in the hamlet of Cát Gia (now Xóm Bằng, Thuận Bắc district, Ninh Thuận) studied by Lee in the late 1970s is a variety of its own or is a subdialect of Southern Raglai (Lee 1998). The status of the Rai dialect of Bình Thuận also needs to be further explored. In any case, native speakers insist that all varieties of Raglai are mutually intelligible, and we witnessed fluent conversations between speakers of Northern and Southern Raglai on several occasions.

Raglai is in direct contact with Vietnamese, which is spoken to various degrees by the majority of Raglai speakers, but also with two other Chamic languages, Chru and Eastern Cham. Chru is spoken by 23,242 speakers mostly in Lâm Đồng province, just across the border from Ninh Thuận, but also in the Bác Ái district in Ninh Thuận (Brunelle et al. 2020; Central Population and Housing Census Steering Committee 2020). Although contact between Raglai and Chru is limited to specific villages, the two varieties are mutually intelligible. In fact, Southern Raglai as spoken in the village of Ma Nới (Ninh Sơn, Ninh Thuận), one of the dialects studied in this paper, is in close contact with the Chru variety spoken in the village of Proh (Đơn Dương, Lâm Đồng) and the two varieties appear to only exhibit minor lexical differences. Some Raglai varieties spoken in the provinces of Ninh Thuận and Bình Thuận are also in contact with Eastern Cham, a language spoken by a little more than 100,000 speakers in the provinces of Ninh Thuận and Bình Thuận, but intelligibility seems more limited than with Chru. None of the two varieties studied in this paper has significant contact with Cham.

Two Raglai varieties are investigated in this paper. The first one is the Northern Raglai dialect spoken in the district of Phước Đại, in Northern Ninh Thuận. As other Northern Raglai varieties, it is described as preserving a voicing contrast in onset stops (Awơi-hathe et al. 1977; Lee 1966; Tạ 2009). As shown in Table 1, stops can be either voiced unaspirated [b, d, ɡ], voiced aspirated [b^h, d^h, ɡ^h], voiceless unaspirated [p, t, k], voiceless aspirated [p^h, t^h, k^h] or voiced implosive [ɓ, ɗ]. Voiced aspirated stops are rare (appearing in less than a dozen words in Awơi-hathe et al. 1977), but they deserve special attention as the existence of true voiced aspirated consonants has been questioned in previous literature due to their typological rarity (Ladefoged 1971). Recent studies have shown that “voiced aspirated” consonants are actually stops with a voiced closure followed by a breathy (voiced) release (Berkson 2019; Islam 2019; Seyfarth et al. 2018). However, Faytak et al. (2020) provide evidence that Yemba voiced aspirates have fully voiced closures, are truly aspirated at the release (voiceless aspiration) and are followed by breathy vowels.

Table 1:

Consonantal inventory of Northern Raglai (adapted from Lee 1966).

p	t	c	k	Ɂ
b	d	ɟ	ɡ
pʰ	tʰ	cʰ	kʰ
bʰ	dʰ	ɟʰ	ɡʰ
ɓ	ɗ	ʄ
	s			h
m	n	ɲ	ŋ
	l, r

The second variety described in this paper is the Southern Raglai dialect spoken in the district of Ma Nới. This dialect is described as having lost voicing in plain stops and aspirated stops but as preserving it in implosives (Tạ 2009). The functional role of the voicing contrast in plain onset stops was apparently taken over by a register contrast on the following vowel, as in other varieties of Southern Raglai and in Cát Gia Raglai (Lee 1998; Tạ 2009). Voiced aspirated stops, on the other hand, merged with voiceless aspirated stops without registral developments (Tạ 2009). The consonant inventory of Ma Nới Raglai is otherwise identical to that of Northern Raglai (Table 1).

In several languages, like Cham and Khmer, register spreads allophonically through onset sonorants (Friberg and Hor 1977; Huffman 1976; Thurgood 1999). In Austroasiatic and Chamic, this usually occurs in sesquisyllables, words whose final stressed syllable, the main syllable, is preceded by a reduced unstressed syllable, the presyllable (Butler 2014; Thomas 1992). For example, in formal Eastern Cham, the main syllable of the word /talah/ “lost” is realized with the acoustic properties of the high register, just like its presyllable, whereas in the word /t̥alah/ “tongue”, the low register of the presyllable, which is marked with a subscript circle, spreads to the main syllable in the surface form [t̥al̥ah] (Brunelle 2009). It is unclear if there is register spreading in Raglai dialects, but its existence would be evidence for a well-established register system.

1.3. Research questions

In this paper, we try to answer the following questions:

1. Are previous descriptions of onset stop voicing in Raglai varieties accurate? Do we find onset stop voicing in Northern Raglai? If yes, how is it realized? Are the Southern Raglai counterparts of Northern Raglai voiced stops really devoiced?

2. Is there any acoustic evidence for the existence of register, be it contrastive or allophonic, in Raglai varieties? If there is, is it equally found after plain and aspirated stops and does it spread to subsequent sonorant-initial syllables?

3. If there is evidence for register in Northern and Southern Raglai, do its acoustic properties have comparable magnitudes? Could these properties be attributed to onset voicing, synchronically or diachronically?

4. Are the voiced aspirated stops of Northern Raglai really voiced?

2. Materials and methods

To address the research questions above, acoustic and EGG recordings of Northern and Southern Raglai speakers were carried out in Ninh Thuận province. In the next subsections, we provide details about the participants (Section 2.1), the wordlist and recording procedure (Section 2.2) and the acoustic and statistical analysis (Section 2.3).

2.1. Dialects and participants

Recordings were made in the province of Ninh Thuận in April 2019. Northern and Southern Raglai data were collected in the communes of Phước Đại and Ma Nới, respectively (see map in Figure 2). Twenty-two speakers of Northern Raglai were recorded (11 women, 11 men). They were all born in the commune of Phước Đại and were aged between 22 and 65 at the time of recordings (mean: 42, standard deviation: 12). All were fluent in Vietnamese and two spoke it at home (along with Raglai). Five of them had resided outside the commune of Phước Đại for more than a year, but only two had ever lived outside the province of Ninh Thuận (one had resided five years in Lâm Đồng and one eight years in Đắk Lắk). Twenty-two speakers of Southern Raglai were also recorded (12 women, 10 men). All were born in Ma Nới and were aged between 22 and 68 at the time of recording (mean: 42, standard deviation: 15). They were all fluent in Vietnamese, but only spoke Raglai at home. Nine of them had previously resided outside Ma Nới for more than a year, but only three had lived outside the province of Ninh Thuận (one had spent 10 years in the neighboring province of Lâm Đồng and two had worked in Saudi Arabia for two years).

2.2. Wordlists and recording procedure

Wordlists designed to measure the acoustic properties of register in various contexts were created. We tried to include words with all combinations of the vowels [iː, ɛː, aː, ɔː, uː] with the coronal and velar onsets [t, tʰ, d, dʰ, ɗ, s, l, n, r, k, kʰ, g, gʰ, ŋ] (recall that voiced aspirated stops are no longer found in Southern Raglai). Data elicitation sessions conducted before the recordings suggest that the realization of voicing and register in these two places of articulation does not differ qualitatively from that of labials and palatals. Monosyllabic words without codas were preferred; when they were not available, we tried to find sesquisyllables with open main syllables and monosyllables with sonorant codas (which have a limited effect on laryngeal settings). Some sesquisyllabic words with main syllables starting in [l, r] were also included to determine if there was register spreading. In total, the Northern and Southern Raglai wordlists included 63 and 60 words, respectively. As the lexica of the two dialects are not identical and as words common in one dialect can be rare in the other, the wordlists differ significantly (only 22 words appear in both lists).

Four repetitions of each word were recorded. The frame sentences used for the two dialects are given in (1–2).

(1).

Northern Raglai frame sentence

/kɔw    ɗəː     ___   kaː    ɲuː    hməː/

I           say    ___   for   3ps    hear

‘I say __ for him/her to hear.’

(2).

Southern Raglai frame sentence

/kɔw    ɗəːp    ___    kaː    saɁaːj               tʰɔw/

I           say     ___   for    older.sibling   know

‘I say ___ for him/her to know.’

Variants of these frame sentences were accepted if they sounded natural. For instance, pronouns and kinship terms were often adapted to the relative age of the participant and experimenter. Sentences were randomized by repetition blocks in SpeechRecorder on a laptop computer. As few participants are literate in Raglai, words were read in Vietnamese by the experimenter (the third author) and had to be translated in Raglai by the participant. They were then inserted and pronounced in the frame sentence. Recording sessions took 30–60 min, depending on the speaker.

Four signals were acquired into a laptop computer through a 4-channel Steinberg UR44 preamp: 1) a high-quality audio channel recorded with a Shure Beta 53 microphone, 2) an EGG signal recorded from a Glottal Enterprises EG2-PCX, 3) a laryngeal height channel acquired through the EG2-PCX, and 4) a moderate-quality audio signal recorded through a Behringer ECM8000. Only the high-quality audio and EGG signals are used in this paper.

All the audio recordings used in the paper are available from the Pangloss and Nakala collections (https://pangloss.cnrs.fr/corpus?lang=en&mode=normal and https://nakala.fr/collection/11280/874bf292).

2.3. Acoustic and EGG analysis

After removing disfluent sentences and sentences disrupted by background noise, a total of 5,465 Northern Raglai words and 5,130 Southern Raglai words were annotated in Praat Textgrids. Several important acoustic events were marked based on spectrograms and audio waveforms (see Figure 3): the sonorant preceding the target syllable (ps), the closure phase of the onset or the full duration of fricatives and sonorants (cl), the opening phase (op), which starts at the release of the closure in words with onset stops or at the onset of the vowel in other words and ends at vowel offset, and the coda (cd), when applicable. Laryngeal events were also annotated based on the EGG signal (second channel in Figure 3): the onset of voicing after the beginning of the closure (ov), the cessation of voicing (cv) when voicing stops before the release because of the aerodynamic voicing constraint, and the resumption of voicing (rv) if voicing is interrupted and resumes.

Figure 3:

Annotation in Praat of a token of the Northern Raglai word /dɔl/ ‘to win, to be the best’ by a female speaker born in 1985, including the spectrogram (top), EGG signal (middle) and TextGrid (bottom) containing delineations of the preceding sonorant (ps), closure phase (cl), opening phase (op), coda (cd), onset of voicing (ov), cessation of voicing (cv) and resumption of voicing (rv).

Several acoustic measurements were extracted using PraatSauce, a Praat-based application for spectral measures based on VoiceSauce (Kirby 2018; Shue et al. 2011). Temporal measures included onset stop closure duration, vowel duration, and stop VOT. Spectral measures included f0 (pitch), F1 and F2 (vowel quality), cepstral peak prominence (CPP) and several spectral slope measures after correction for formant frequencies and bandwidths (Iseli and Alwan 2004), marked as H1*–H2*, H1*–A1*, H1*–A3*, at every 1 ms in the following vowel. As phonation measures are highly correlated, we only report one spectral slope measure, H1*–H2*, but complement it with CPP, a robust indicator of noise (Garellek and Esposito 2021) and open quotient, a measure that we extracted from the EGG signal using PraatDet, a Praat implementation of PeakDet (Kirby 2020; Michaud 2007). We report the open quotient as calculated with the Howard hybrid method, in which the closing instant of each glottal cycle is the time of its maximum derivative and its opening instant corresponds to the time at which the glottal signal crosses 3/7 of the range defined by the maximum and minimum points of a glottal cycle (Howard 1995). The derivative method for open quotient calculation (dEGG) was also calculated but is not reported here as it was less robust due to frequent double opening peaks in some speakers (Henrich et al. 2004). The open quotient results of seven female subjects (five Northern and two Southern Raglai) were excluded because their EGG signal was too weak to yield regular glottal waveforms.

Since the window used for acoustic analyses had a duration of 25 ms, f0, formant measures and spectral results made during the first and last 12 sampling points of each vowel are not reported, as they would be unduly affected by adjacent consonants. Outliers and tracking errors were also filtered out. Local tracking errors were detected by z-normalizing f0, F1 and F2 by speaker and calculating derivatives for each of these normalized measures (in z-scores). All measures whose derivatives were above 0.5 z or below −0.5 z were removed. Then, in order to detect tracking errors over longer portions of the signal, mean f0, F1, F2 and open quotient values were computed for all combinations of speaker, vowel and voicing/register. Any measure distant from the mean by more than three standard deviations was considered an outlier and removed. All spectral slope measures dependent on excluded f0, F1 or F2 values were also removed. The proportions of removed values for the two dialects are given in Table 2.

Table 2:

Proportion of the data removed as outliers and tracking errors before analysis, by relevant phonetic property.

	f0	F1	F2	H1*–H2*	Open quotient
Northern Raglai	3.3%	3.9%	2.4%	6.4%	1.3%
Southern Raglai	3.4%	4.7%	3.7%	7.3%	1.2%

To facilitate comparison of acoustic properties across speakers, f0, F1, F2, H1*–H2* and CPP measures were z-normalized a second time after the data was cleaned. Then, to facilitate interpretation of the data, these z-scores were converted into familiar scales using the means and standard deviations of all the speakers (mean of all speakers + z-score * mean standard deviation of all speakers).

2.4. Statistical analysis

To test the statistical significance of effects, mixed models were conducted in R using the lmerTest package (Kuznetsova et al. 2017) on VOT and on the secondary properties of voicing at various time points during the vowel. Fixed main effects were voicing (Northern Raglai) or register (Southern Raglai), place and vowel. Only two-way interactions were included as three-way interactions often relied on single tokens and resulted in overfitting. Random effects included by-subject and by-word random intercepts. Models were simplified by dropping non-significant interactions with the lowest F-values in the ANOVA models one at a time if the resulting ANOVA had 1) a lower Akaike information criterion (AIC) score than the previous model, or 2) a higher or equal, but not significantly different AIC.

To limit the effects of potential tracking errors on statistical results, we report statistical models for the acoustic properties of vowels averaged over three 10-ms-long time windows (0–10, 50–60 and 100–110 ms, respectively). It should also be noted that we did not attempt to fit models measuring the significance of acoustic differences conditioned by Northern Raglai aspirated stops because there were too few tokens of voiced aspirated stops in our wordlist. Our description of aspirated stops is limited to observed data plotted in figures.

Cohen’s d’s were used as a proxy for the cue weights of the different acoustic properties of voicing and register (Cohen 1988; Clayards 2008; Brunelle et al. 2020; Schertz and Clare 2020 for a review of different cue weighting indicators). They were obtained by calculating the difference between the vowel-weighted and subject-weighted means of each level of each cue during the first 10 ms of the vowel following plain stops, divided by the vowel-weighted and subject-weighted standard deviation of each cue. A large absolute Cohen’s d indicates a large separability between two distributions and provides evidence that a given cue could bear some level of phonological contrast.

3. Results

Northern and Southern Raglai results will be presented in parallel, to facilitate comparison between dialects. In Section 3.1, we illustrate the realization of voicing in stops. In Section 3.2, we compare the realization of vowels following the voicing/register contrast in the two dialects. We finally report the relative weight of each acoustic property involved in the voicing and register contrasts in Section 3.3. In the rest of this section, we will refer to Northern Raglai and Southern Raglai as NR and SR, respectively.

As will become obvious in a few pages, Northern Raglai preserves the voicing contrast in onset stops while Southern Raglai has replaced it with a register contrast, confirming previous descriptions. For the sake of clarity, we will label the laryngeal contrasts accordingly from this point on.

3.1. Onsets

A comparison of stop VOT in the two dialects supports the claim made in previous descriptions that NR preserves a robust voicing contrast that has largely been neutralized in SR. In the left panel of Figure 4, the majority of NR plain voiced stops have a negative VOT, with a mean of −72 ms (cf. Table S.1 in the Supplementary materials). In contrast, the right panel shows that low register stops, their correspondents in SR, overwhelmingly have a positive VOT, with a mean of 14 ms (cf. Table S.2 in the Supplementary materials). NR plain voiceless stops and their SR high register counterparts, on the other hand, share a moderately positive VOT (mean of 8.6 ms for NR and 12.5 ms for SR). A closer inspection of the SR register contrast in plain stops further reveals that when tokens with closure voicing are excluded, low register velar stops have a significantly longer VOT than their high register counterparts (RegisterLow*PlaceVelar: β = 10 ms, t = 4.66, p = 0.04), but that no such effect is found in coronal stops (cf. Table S.4 in the Supplementary materials).

Figure 4:

Voice Onset Time in Northern Raglai (left) and Southern Raglai (right) stops.

The NR voicing contrast and the SR register contrast do not significantly affect closure duration, but a closer look at closure voicing confirms that the two Raglai dialects have markedly different voicing contrasts. Rather than a simple voicing contrast in which voiced stops have a negative VOT, we found a variety of realizations of the latter, which we divide into three main categories. This is illustrated in Figure 5 with three realizations of the voiced stop /d/ in the word /məda/ ‘rich’ by the same young female NR speaker. The top panel of the figure is a typical example of a stop with a negative VOT. We categorize all stops with a negative VOT as voiced, even if their closure is not fully voiced. The mid panel is an example of a stop with a positive VOT and marginal perseverative voicing (called “bleeding” in Davidson 2016). We categorize stops with a positive VOT as devoiced if their bleeding does not exceed 30% of the closure duration (which corresponds to the rounded-up maximum bleeding found in phonologically voiceless stops). Finally, the bottom panel is an example of a stop that is voiced over most of its closure, but in which voicing is interrupted before the release, probably because of an air pressure build-up in the oral cavity (the aerodynamic voicing constraint first described by Ohala 1983), and only resumes 20 ms after the burst. Stops combining a portion of closure voicing that is not mere bleeding with a positive VOT are categorized as having a voiceless release.

Figure 5:

Spectrograms of three realizations of the Northern Raglai voiced stop /d/ in the word /məda/ ‘rich’. Top panel: Fully voiced closure; Mid panel: Voiceless closure (with limited perseverative voicing from preceding vowel); Bottom panel: Voiced closure with voiceless release and lag VOT.

A breakdown of NR voiced stops and SR low register stops by voicing category, in Figure 6, reveals that a characterization of plain stops in terms of VOT only is insufficient. One can see in the left panel that most NR women have a non-negligible proportion of voiceless releases (and some devoiced stops) in their voiced stops, while NR men rarely do. By contrast, in SR, women (except the oldest one, F49) almost exclusively produce devoiced low register stops, whereas most men preserve traces of voicing. Despite different baselines, the two dialects share a common pattern: voicing appears less prevalent in women.

Figure 6:

Breakdown of NR voiced stops and SR low register stops by category and speaker. On the x-axis, subjects are identified by gender (F/M) and year of birth. NR F84 and F84′ are two women born the same year.

Implosives and aspirated stops mostly behave as expected. Figure 4 confirms that implosive /ɗ/ has a negative VOT. What differentiates it from non-implosive /d/ is its higher voicing amplitude, as shown in Figure 7. Aspirated stops tend to have a large positive VOT but the difference in closure voicing between NR voiceless and voiced aspirates is not as robust as anticipated (recall that SR has neutralized this contrast and only preserves high register aspirates). While the closure of voiceless aspirates is systematically voiceless, the proportion of the closure that is voiced is highly variable in voiced aspirates, as can be seen in Figure 8. Older speakers preserve closure voicing more systematically than younger ones, and young men devoice more than young women. This could be an indication that the voicing contrast wanes in a socially structured manner, but we have too few voiced aspirates in our wordlist to fully explore this question (5 words × 4 repetitions per speaker).

Figure 7:

Voicing amplitude during the closure of /d/ and /ɗ/. Thick lines represent means, thin lines represent individual observations. There are few tokens of Southern Raglai /d/ as it is normally devoiced.

Figure 8:

Northern Raglai voiced aspirated stops. Left: Mean proportion of voicing during closure, by Year of Birth and Sex. Right: Proportion of each voicing type, by speaker.

3.2. The effect of the laryngeal contrast in onsets on following vowels

In this section, we look at the effect of NR voicing and SR register on several acoustic properties associated with these laryngeal contrasts at the beginning of vowels. As the acoustic effects of voicing and register are largely localized in the initial portion of the vowel, we only show the first 200 ms in figures and report statistical models in the text (and in the Supplementary materials). For each acoustic property, we first report on voicing/register-conditioned differences after plain stops and NR aspirated stops (recall that SR has lost the voicing contrast in aspirates). We then look at possible register spreading effects by comparing the acoustic properties of sonorant-initial syllables that are register-neutral (either because they are monosyllabic or because they are preceded by a presyllable that is itself headed by a sonorant) with those of sonorant-initial syllables that are preceded by presyllables headed by voiced/low register onsets and by voiceless/high register onsets. We finally look at the acoustic properties of vowels following the fricative /s-/ and the implosive /ɗ-/ to determine if they seem to pattern with one of the two plain onset stop series (voiced/low or voiceless/high).

A look at f0 after different consonants reveals that it does not play a contrastive role across speakers. The small differences found after voiced and voiceless plain stops in NR and after high and low register plain stops in SR in Figure 9 seem to go in the opposite direction from expectations, but only approach significance immediately after NR voiced coronal stops (Tables S.5–S.10 in the Supplementary materials). A greater difference, in the expected direction this time, is apparently found after voiced and voiceless aspirates in NR, but there are too few words in our sample to test it statistically. The f0 of vowels following sonorants is not affected by the voicing or register of the preceding syllable or the absence of a preceding syllable in either dialect (Table S.11–S.12 in the Supplementary materials). Other obstruents behave as expected, with the fricative /s/ conditioning a much higher f0 than the implosive /ɗ/ (which has a f0 comparable to that of plain stops).

Figure 9:

Normalized f0 of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

Figure 10 shows that the NR voicing contrast in plain stops conditions a small difference in H1*–H2* in following vowels, but this difference is only significant between the 50th–60th ms of the vowel (VoicingVoiced: β = 1.7 dB, t = 2.57, p = 0.04; Table S.14 in the Supplementary materials). Register seems to affect H1*–H2* more dramatically after SR plain stops, but this effect does not actually reach statistical significance because of large interspeaker variation (Tables S.16–S.18 in the Supplementary materials). Note however that another spectral slope measure, H1*–A3*, is significant across speakers at the beginning of the vowel (at 1–10 ms, RegisterLow: β = 6.2 dB, t = 4.8, p = 0.039). There is no clear difference between voiced and voiceless aspirates in NR, but as expected, both series are followed by a high degree of breathiness, as are the aspirates of SR. As for the H1*–H2* of vowels following onset sonorants, it is not clearly affected by the voicing/register of the preceding syllable, but it starts a bit higher in neutral syllables in both languages (Tables S.19–S.20 in the Supplementary materials). Other obstruents behave as expected: the fricative /s/, which is produced with an open glottis, conditions a high H1*–H2* at the onset of following vowels, but implosives, which presumably have a tightly closed glottis, seem to start on a relatively low H1*–H2*.

Figure 10:

Normalized H1*–H2* of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

The other acoustic indicator of phonation reported here, CPP, is shown in Figure 11. It is slightly lower immediately after voiced stops in NR (VoicingVoiced: β = −1.58 dB, t = −2.50, p = 0.047; Table S.21 in the Supplementary materials), with a much stronger effect in velars (VoicingVoiced:PlaceVelar β = −4.80 dB, t = −5.39, p = 0.002; Table S.21 in the Supplementary materials), but is not affected by register in SR (Tables S.24–S.26 in the Supplementary materials). No meaningful CPP differences appear to be conditioned by voicing/register in aspirated stops and sonorants (Tables S.27–S.28 in the Supplementary materials) but sonorant-headed vowels start on an overall higher CPP than obstruents, probably because of the former’s more neutral glottal setting, and aspirate-initial syllables begin with a low CPP that seems to last longer into the vowel than in plain stops because aspiration disrupts regular vocal fold vibrations.

Figure 11:

Normalized CPP of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

The last phonetic indicator of phonation reported in this study, the open quotient (calculated with the Howard method), yields mitigated results, just like H1*–H2* and CPP. In Figure 12, there is a weak, but significant difference in open quotient immediately after NR plain voiced and voiceless stops (VoicingVoiced: β = 0.03, t = 3.14, p = 0.017; Table S.29 in the Supplementary materials). The difference in open quotient after high and low register plain stops appears much greater in SR, but due to individual variation, it only reaches significance between 50 and 60 ms (RegisterLow: β = 0.03, t = 3.00, p = 0.023; Table S.33 in the Supplementary materials). There is no apparent open quotient difference in NR vowels following voiceless and voiced aspirated stops, but open quotient is higher after glottally abducted aspirated stops in general. No voicing or register difference is found in sonorants either (Tables S.35–S.36 in the Supplementary materials). Finally, open quotient is higher after the glottally-abducted fricative /s/ and lower after the glottally-adducted /ɗ/ than after plain stops in NR, but unexpectedly, the fricative /s/ and the implosive /ɗ/ both fall between high and low register plain stops is SR. The latter may suggest that /ɗ/ does not preserve its full glottal adduction in SR, where it no longer contrasts with a voiced coronal stop.

Figure 12:

Normalized open quotient of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

Turning to vowel quality, we see in Figure 13 that F1 is affected by voicing and register. In NR, F1 is significantly lower immediately after plain voiced stops than voiceless stops in vowels /aː/ and /ɔː/ (VoicingVoiced: β = −159 Hz, t = −7.90, p < 0.001; Table S.37 in the Supplementary materials). In contrast, differences in F1 between SR registers at vowel onset are large regardless of the vowel (RegisterLow = −246 Hz, t = −5.20, p = 0.04; Table S.40 in the Supplementary materials) and persevere throughout the first 100 ms (Tables S.41–S.42 in the Supplementary materials).

Figure 13:

Normalized F1 of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

There appears to be weak voicing-conditioned differences in F1 after aspirated stops in NR, but there are too few tokens of voiced aspirates to test this statistically. In SR, F1 is noticeably lower in sonorant-initial syllables preceded by low register or register-neutral presyllables than in those preceded by high register presyllables (RegisterLow: β = −146 Hz, t = −3.80, p = 0.001; RegisterNeutral: β = −95 Hz, t = −3.59, p = 0.002; Table S.44 in the Supplementary materials), but no such effect is found in NR (Table S.43 in the Supplementary materials). The implosive /ɗ/ and the fricative /s/ do not seem to pattern in a coherent way across vowel qualities.

F2 variations are plotted in Figure 14. There is no significant effect of NR plain stop voicing on the F2 of following vowels, except for /ɛː/ (Tables S.45–S.47 in the Supplementary materials), due to an unexpectedly low F2 in the word [gɛːj]. In SR, however, vowels following low register plain stops start with a significantly higher F2 (RegisterLow: β = 163 Hz, t = 4.79, p = 0.002, Table S.48 in the Supplementary materials), a difference that is still significant between 50 and 60 ms and between 100 and 110 ms (Tables S.49–S.50 in the Supplementary materials). NR aspirated stops do not seem to condition F2 differences except perhaps in vowel /ɔ:/, but as there is a single word with a voiced aspirate before that vowel, no conclusions can be drawn. Sonorants, fricatives, and implosives do not condition detectable F2 modulations in either dialect (Tables S.51–S.52 in the Supplementary materials).

Figure 14:

Normalized F2 of vowels following Northern Raglai (top) and Southern Raglai (bottom) onsets. Thick lines represent means, thin lines represent individual observations.

The most relevant results of this subsection can be summarized as follows. The NR voicing contrast only has a limited effect on the acoustic properties of following vowels. NR plain voiced stops condition a lower F1 in the first 50 ms of vowels /aː/ and /ɔː/, but this effect is weak after other vowels. CPP, open quotient and H1*–H2* results provide evidence for a marginally laxer phonation in following vowels, but this effect, even if significant, is not necessarily large enough to be perceptible. No significant effect of voicing on f0 and F2 could be detected. NR voiced aspirates, on the other hand, appear to have a slightly lower f0, but there are too few of them in our dataset to analyze this difference statistically.

The SR register contrast affects vowels in a more salient manner. Vowels following low register stops have a robustly lower F1 and a significantly higher F2 over at least their initial 110 ms. That F2 is higher in low register stops, which were historically voiced, appears to contradict previous results on voicing-F2 coarticulation (Cole et al. 2010) and could suggest that register-conditioned F2 patterns are not due to a direct transphonologization coarticulatory F2. Phonation measures (H1*–H2*, CPP, open quotient) are large when we look at averages across speakers, but are not consistently significant, reflecting an important individual variation that is further addressed in Section 3.3. No effect of register on f0 was detected.

There is no evidence of phonologized register spreading through sonorants in either dialect, although there may be some weak F1 coarticulation in SR. As for non-contrastive obstruents, /s/ conditions a relatively breathy phonation (higher H1*–H2* and lower CPP) and a high f0 in both dialects, while /ɗ/ appears to roughly pattern like the plain voiceless stops of NR and the high register stops of SR.

3.3. Individual cue weights

In order to weigh the relative importance of each acoustic property in the voicing and register contrasts and to inspect cue weight variation across individuals, Cohen’s d’s were computed (see Section 2.3 for more details on their calculation). A small Cohen’s d indicates a weak acoustic weight whereas a large positive or negative Cohen’s d indicates a large one. Cohen’s d’s for each property and participant are given in Figure 15, where a positive Cohen’s d shows that the voiceless/high register series has higher values than the voiced/low series and a negative Cohen’s d shows the opposite.

Figure 15:

Distribution of Cohen’s d’s associated with Northern Raglai plain stop voicing (left) and Southern Raglai registers after plain stops (right). Participants are ordered by year of birth and sex on the X-axis (F84 and F84′ are two different female speakers born in 1984).

A comparison of NR and SR shows that VOT is by far the dominant cue of the NR voicing contrast, whereas F1 is the dominant cue of the SR register contrast. There are no consistent secondary cues in NR, but some speakers do seem to make use of cues other than VOT on an individual basis. In contrast, phonation (H1*–H2*, CPP) is a non-negligible secondary cue in most, but not all, SR speakers, and does not systematically reach significance in the speaker sample as a whole. Moreover, there is no evidence that this variation in the use of phonation is structured by age or gender or correlated with the use of F1.

The other striking distribution when looking at cue weights in Figure 15 is that in NR, Cohen’s d’s for VOT are on average much higher in men than women. This is due to the fact that many NR women exhibit a tendency to stop devoicing that is primarily embodied by a high prevalence of voiced stops with voiceless releases, as discussed in detail in Section 3.1. If these women cannot depend on phonetic voicing as reliably as other speakers to distinguish the NR plain stops voicing contrast, one wonders if they could make more use of other cues. We investigated possible differences between groups by looking at the distribution of all the acoustic cues reported in Section 3.2 in the production of Devoicers, the eight women who have voiced stops with voiceless releases in Figure 6, and Voicers, the other speakers. The only acoustic property for which there is a significant difference between the two groups is CPP. As illustrated in Figure 16, devoicers produce significantly more noise at vowel onset immediately after voiced stops, an effect that is even stronger in velars (VoicingVoiced: β = −2.97 dB, t = −2.97, p = 0.14; VoicingVoiced:PlaceVelar: β = −4.77 dB, t = −3.35, p = 0.007 in Table S.53). Voicers, by contrast, do not exhibit a significant CPP difference between the voiced and voiceless series (Table S.54 in the Supplementary materials).

Figure 16:

CPP differences after phonologically voiced and voiceless stops in Northern Raglai in Devoicers and Voicers.

This forces us to revise the results obtained for NR phonation in Section 3.2. It appears that in this dialect, there are large CPP differences between phonologically voiced and voiceless stops in devoicers, but not in other speakers. In other words, NR speakers whose voicing contrast is unstable exhibit a greater reliance on phonation than other speakers.

4. Discussion/conclusion

Our discussion of the significance of the results follows the research questions raised in Section 1.3. Sections 4.1–4.2 respectively address the nature of the voicing contrast in onset stops (Question A) and review evidence for the presence of register (Question B) in the two Raglai varieties. In Section 4.3, we discuss how the register contrast of SR could have evolved from a voicing contrast like that of NR (Question C). We conclude with a brief description of the voiced aspirated stops of NR (Question D) in Section 4.4.

4.1. Onset stop voicing

Our results confirm the claims made in previous descriptions that NR preserves a voicing contrast in onset stops while SR does not (Awơi-hathe et al. 1977; Lee 1966; Tạ 2009). NR voiced stops typically preserve robust closure voicing and a negative VOT, but 13% of them exhibit a voiceless release (like that illustrated at the bottom of Figure 5), which means that their closure voicing ceases a few milliseconds before the burst and resumes a few milliseconds after it. This type of partial devoicing is found in most women (8/11 exhibit it) but is almost unattested in men. In contrast, the SR cognates of NR voiced stops, the low register stops, are typically voiceless, even if a minority of speakers occasionally produces them with closure voicing.

4.2. Register

Rather than a voicing contrast, SR has a register system. In this dialect, Chamic proto-voiced stops have been transphonologized into a low register primarily realized through vowel quality and, for most speakers, secondarily modulated by phonation. At their onset, low register vowels are significantly more close (i.e. have a lower F1) and slightly more fronted (i.e. have a higher F2) than high register vowels, an effect that is more marked in open than close vowels. Speakers who exhibit secondary phonation also have a moderately breathier vowel quality in the low register (lower CPP and higher H1*–H2*). However, SR register does not condition consistent f0 differences.

In contrast, NR only has marginal registral cues. Vowel quality differences conditioned by onset voicing are limited: F1 is significantly higher at the onset of /ɔː, aː/ following voiced stops and F2 is not affected. Unexpectedly, f0 is not a secondary property of NR voicing either, contrary to what is found in most European languages (Coetzee et al. 2018; Dmitrieva et al. 2015; Hanson 2009; Kirby et al. 2016; Lisker 1986; Rousselot 1901–1908; Ohde 1984). This suggests that voicing may not be realized in the exact same manner in NR and in these languages. It is possible, for example, that the increased longitudinal tension that hinders vocal fold vibrations in voiceless stops in English and Dutch is not found in NR (Löfqvist et al. 1989). The only acoustic property that is robustly affected by onset stop voicing in NR is phonation. When all speakers are aggregated, there is only a slightly laxer phonation after voiced stops: the difference in spectral balance is consistently below 2 dB, which is negligible compared to what is found in other register languages (e.g., Brunelle et al. 2020; Tạ et al. 2022). However, if we look at variation across speakers, we find that the women who partly devoice their voiced stops have a dramatically breathier phonation on the following vowel (i.e., a CPP lower by about 3 dB in coronals and 8 dB in velars), a difference that is not found in other speakers. There is therefore evidence for the development of a conspicuous registral property, phonation, in the very NR speakers who have an unstable voicing contrast.

We found no evidence for the rightwards spreading of the acoustic properties of voicing and register in either dialect. There is a slight tendency for F1 to be higher in sonorant-initial syllables preceded by high register presyllables in SR, but it would be better accounted in terms of vowel-to-vowel coarticulation than as a phonological process of register spreading. We could not find evidence of a generalization of the register contrast to the non-contrastive obstruents /ɗ-/ and /s-/ either. These facts suggest that the SR register system is still intrinsically linked to plains stops and has not started generalizing to new environments.

The fact that SR register does not make use of contrastive pitch and has not developed suprasegmental properties like register-spreading suggests that it is different from tone, even if its phonetic precursor, onset obstruent voicing, is instrumental in the development of both complex and binary tone systems (Haudricourt 1961; Svantesson and House 2006). While other register systems are more tone-like, including that of closely related Eastern Cham (Brunelle 2005, 2006), this would suggest that register and tone are related, but distinct contrasts.

Before concluding this subsection, it must be noted that the SR register system is very similar to that observed in Chru, a neighboring and closely related language (Brunelle et al. 2020). SR and Chru both have register systems in which F1 is the primary property and phonation plays a secondary role. However, contrary to what was found in SR, a significant proportion of Chru speakers preserve a redundant voicing distinction. This, along with important lexical similarities between Chru and the SR variety reported here, suggests that Raglai and Chru may be part of a Chamic dialectal continuum, and that the autonyms used by Chamic communities may not match dialectal boundaries.

4.3. From voicing to register

There are two ways in which the acoustic perturbations caused by the NR plain stop voicing contrast on following vowels could be precursors to the formation of a register system like that of SR. The first one is a direct phonologization of the F1 modulations caused by stop voicing, possibly through a larynx lowering or a tongue-root advancement maneuver meant to facilitate vocal fold vibrations by increasing transglottal airflow. This poses an empirical problem: even if voicing-conditioned F1 variations are robust across speakers in NR, they only significantly affect /ɔː/ and /aː/. If the ancestor of SR ever resembled present-day NR, we would have to hypothesize that the F1 differences found in these low vowels after voiced stops were exaggerated and generalized to other vowels as voicing started to become an unreliable contrastive property, leading to the development of a F1-based register.

The other possible precursor of register observed in NR is the breathiness found after voiced stops in women who exhibit signs of unstable voicing. If SR registrogenesis started out with the same kind of phonation perturbation, we would need to hypothesize that it spread from women to men but would have no clear explanation for why SR register is now largely based on vowel quality. Registrogenesis scenarios in which an initial phonation difference triggers changes in vowel quality have been proposed, but without a clear phonetic motivation (Huffman 1976; Wayland and Jongman 2002). A possible explanation could be a reorganization of register cues caused by the perceptual relation between breathiness and vowel quality (Esposito et al. 2019; Kingston et al. 1997; Lotto et al. 1997). Another possibility is that the transphonologization of F1 and phonation occured simultaneously without being causally linked: the phonologization of independent but redundant secondary cues may even have facilitated the shift away from voicing.

Since we know that women normally lead sound change, an account in which phonation differences are grounded in the frequent devoicing found in their speech is appealing. However, this account raises two important questions for which we do not have clear answers. The first one is why NR women would exhibit a stronger tendency to devoicing than NR men. Although there is evidence that women devoice obstruents more than men (Bayley and Holland 2014; Brunelle et al. 2020; José 2010; MacKenzie 2018; Michnowicz and Planchón 2020; Smith 1978; Van Alphen and Smits 2004; Verhoeven et al. 2011), no systematic survey of the question has to our knowledge been conducted and no clear physiological mechanism seems to account for this asymmetry between sexes. Women produce voicing with glottal gaps more frequently than men, but this should not necessarily result in devoicing (Hanson 1995; Patel et al. 2012; Yamauchi et al. 2014). Besides, while it has been proposed that air pressure may build up faster in women’s smaller supraglottal cavities (Ting 2015), this account depends on the debatable assumption that there is less of a sexual asymmetry in subglottal than in supraglottal pressure. The second question is the nature of the phonetic mechanism linking devoicing to a breathier phonation on the following vowel. An obvious explanation would be the presence of a glottal gap allowing a noisy air leakage during phonation in women’s speech. This would be a parsimonious account since a glottal gap was also invoked as a possible cause for devoicing, but in the absence of articulatory data, we only have circumstantial evidence for it. In any case, data from NR devoicers confirms the presence of breathiness at vowel onset following a subtype of voiced stops, those with a voiceless release (cf. Figure 16). To our knowledge, it is the first time that the existence of such a phonation perturbation, that has long been postulated as an essential step in registrogenesis (Ferlus 1979; Haudricourt 1965; Huffman 1976; Thurgood 2002), is found in a language that preserves a robust voicing contrast. More importantly, these results suggest that the breathiness associated with the low register in many Southeast Asian languages may not be a transphonologized secondary property of voicing but a consequence of devoicing.

4.4. Voiced aspirated stops

While the voicing contrast is well-preserved in NR plain stops, it seems more unstable in its aspirated stops. Canonical voiced aspirates appear to have somewhat devoiced closures with partially voiceless and partially voiced aspiration. Devoicing typically occurs towards the end of the closure and leads into a voiceless aspiration at the release. As aspiration merges into the vowel, it becomes more voiced and the following vowel is strongly breathy. Thus, aspirated voiced stops in NR do not have the fully voiced closures and the entirely voiceless aspiration found in Yemba “true-voiced aspirates” (Faytak et al. 2020), nor is their aspiration entirely voiced as in “breathy-voiced” stops described elsewhere (Berkson 2019; Islam 2019; Seyfarth et al. 2018). Alongside these canonical voiced aspirates, it appears that some speakers, and especially younger ones, realize voiced aspirated stops with fully voiceless closures (Figure 7) without compensating this devoicing by phonetic modulations on following vowels. If our limited results are representative of voiced aspirated stops in general (and it should be kept in mind that we have too few tokens of voiced aspirates to conduct a quantitative analysis), they would be evidence that the NR voicing contrast in aspirates is gradually being neutralized without compensation, just as it was in SR. Given how few minimal pairs depend on a voicing opposition in aspirates in NR, this would not impact the lexical structure of the dialect dramatically. In fact, most Chamic languages have already neutralized voicing in aspirates without compensation (Brunelle and Jensen in press).

Now that we have established the relative role of the acoustic properties involved in the production of NR voicing and SR register, an obvious next step that was not possible because of the recent COVID-19 pandemic would be to conduct a perceptual experiment to test the role of each of these properties in the perception of the contrast. It would be especially interesting to determine if listeners are sensitive to cues used in the other dialect and actively use them in cross-dialectal communication.

Supplementary Material

Supplementary Material

Appendix: Word lists

Table 1.1:

Northern Raglai word list.

N. Raglai (IPA)	Vietnamese (Quốc ngữ)	English
ti: (tlwaʔ)	theo	to follow
tɛ:m	tem	stamp
mata:	mắt	eye
patɔ:	dạy	to teach
tu:	tủ	wardrobe, closet
di:l	ghi nhớ	to remember dearly
mada:	giàu	rich
dɔ:l	thắng	to win, to be the best (phys. activity)
hadu:	bao nhiêu, mấy	how many, how much
kaɗi:	vụ án, vụ việc	case, affair
ɗɛ:l	cạn	shallow
caɗa:	nứt	fissure, crack
kaɗɔ:	nhím nhỏ	porcupine
(bi)ɗu:	nổi	to float
thi:l	chặt cây cho bằng	to level, to smoothen
thɔ:	cầm, giật	to seize, to pull down
thu:	khô	dry
bidhi: (atɔw)	lễ bỏ mả	k.o. funeral ceremony
kadha:	việc giải quyết mẫu thuận vợ chồng	family meeting to solve domestic strife
dhɔ:	chân dốc, thung lũng	foothills
dhu:l	con dơi	bat
tasi:	lược	comb
kasɛ:	nói bậy bạ khi say rượu	to talk stupidly when drunk
sa:	một	one
(aŋi:n) tasɔ:	bão nhỏ	small storm
su: (ʔea:)	cái xô	pail
ni:	này	this
nɛ:	nè	hey
ina:	con thú mẹ	mother animal
anɔ: (ʔea:)	gánh	to carry water
anu:	loại chuột, con dũi	k.o. rodent
alɛ:	cây le	k.o. bamboo
ula:	rắn	snake
lɔ:	Trung Quốc	Chinese
lu:	lắm, nhiều	very
(patɔ:) tali:	đá bàn	bedrock
kalɛ:	cờ lê	wrench
pila:	trồng (rẫy)	to plant (in dry field)
kumlɔ:	câm	dumb, mute
ɟalɛ:	cái cuốc	pickaxe
bala:	ngà	tusk
galɔ:	lỗ	hole
bilu:	mầu nâu	brown
rɛ:	ngứa ngoài da	skin rash
ara:	vịt trời	wild duck
rɔ:	nuôi con vật	to raise animals
ru:ʔ	lấy sức	to collect one’s strength
aŋi:n	gió	wind
(Ɂja:) laŋa:r	(nước) trong	clear (water)
tuki:	gạc, sừng	antler, horn
kɛ:	con nít	young child
ka:	hàm	jaw
kakɔ:	bệnh lác (da)	scaly skin disease
iku:	đuôi	tail
khi:n (ŋaʔ)	dám (làm)	to dare
ukha:	rễ	root
(laŋiʔ) khɔ:	trời tạnh	cleared sky
khu:ŋ	cũng	also
pagi:	ngày mai	tomorrow
gɛ:j	bo bo, cao lương	barley
paga:	hàng rào	gate, fence
(tuki:) agu:ʔ	queo (sừng trâu)	pointing down and curved (of buffalo horn)
ghu:	cháy, nhóm lửa	to light kindle

Table 1.2:

Southern Raglai word list.

S. Raglai (IPA)	Vietnamese (Quốc ngữ)	English
kati:r	dung nham	lava
patɛ:l	ly bự	big glass
mata:	mắt	eye
patɔ:	dạy	to teach
tu:	tủ	wardrobe, closet
kadi:	ấm nước	teapot
mada:	giàu	rich
dɔ:r	chôn	to bury
hadu:m	bao nhiêu, mấy	how many, how much
kaɗi:	vụ án, vụ việc	case, affair
(ta:) rəɗɛ:	cà chua	tomato
kaɗɔ:	nhím nhỏ (hoẵng)	porcupine
ɗu:	đồ đạc, tài sản	things, belongings
pathi: (atɔw)	lễ bỏ mả	k.o. funeral ritual
kətha:	chuyện	case, affair
thɔ:	cầm giữ, kéo	to hold something stretched or tense
thu:	khô	dry
tasi:	lược	comb
(bləw) məsɛ:	lông tơ	down, fluff
sa:	một	one
sɔ:r	con dơi	bat
su:m	bụi cây	pollen
ni:	này	this
nɛ:	nè	hey
na:	đó, đấy	there (medial)
anɔ:ŋ	gánh	to carry on shoulder
anu:	con dũi	k.o. rodent
li:r	làm cho đầy	to fill
alɛ:	cây le	k.o. bamboo
ala:	rắn	snake
ralɔ:	thịt	meat
lu:	lắm, nhiều	very
tali:	đá bàn	bedrock
pala:	trồng (rẫy)	to plant (in dry field)
kamlɔ:	câm	dumb, mute
ɟalɛ:k	cái cuốc	pickaxe
bəla:	ngà	tusk
bəlu:	mầu nâu	brown
ari:	nhích	to shift
rɛ:	xui/vết thương	unlucky/wound
ra:w	rửa	to wash
rɔ:	chuồng	cage
ru:	chung	together
aŋi:n	luồng gió, luồng hơi	gust of wind
aŋa:n (mta:)	tên gọi	name
palaŋɛ:	nhìn xa	to look afar
rəŋɔ:	lá thơm, ngò gai	sage (plant)
taki:	gạc, sừng	antler, horn
pakkɛ:	tắc kè	gecko
laka:	vết sẹo	scar
aku:	đuôi	tail
khi:r	kín	tight
khɛ:	đỏ	red
akha:r	rễ	root
khɔ:	kho	to braise
pəgi:	ngày mai	tomorrow
rəgɛ:j	khéo léo	handy
paga:	hàng rào	gate, fence
gɔ:r	cán	handle
gu:l	tròn	round

Supplementary Material

The online version of this article offers supplementary material (https://doi.org/10.1515/phon-2022-2019).