Redundant voicing and register in Mnong Râlâm

2.1.1 Procedure

Twenty-three native speakers (12 women) of Mnong Râlâm born between 1944 and 1992 (mean: 1975, sd: 15) were recorded in March 2019. They were all originally from the commune of Yang Tao, except for one man who was born in Buôn Ma Thuột, the province capital, but moved back to Yang Tao at the age of four years old. They were all fluent in Vietnamese, and 21 of them also spoke Rade, a Chamic language that is the lingua franca used between minority groups in the province of Đắk Lắk. All spoke primarily Mnong in their daily lives, except one who was married to an ethnic Vietnamese man and spoke Mnong and Vietnamese equally. Five participants had lived in other regions of Vietnam (Buôn Ma Thuột, Hồ Chí Minh City, Kontum) for periods of time ranging from two to eleven years.

Sixty real words comprising target syllables made up of combinations of all dental and velar onsets /t, k, tʰ, kʰ, d, ɡ, ɗ, s, n, ŋ, r, l/ and the five vowels /iː, ɛː, aː, ɔː, uː/ were selected. The full wordlist is given in Appendix I. Open syllables were preferred; when they were not available, syllables closed by sonorants were chosen. Target words were either monosyllables or sesquisyllables ending in target syllables. Sequisyllables are disyllables in which the final stressed syllable has the full array of contrasts while the initial syllable, or presyllable, has a reduced inventory (Thomas 1992).

While the register contrast is normally limited to plain obstruents, it must be noted that many Austroasiatic and Chamic languages have a process of register spreading in which sonorant-initial syllables acquire the register of the preceding syllable or stop (Brunelle and Phú 2019; Friberg and Hor 1977; Huffman 1967; Thurgood 1999). For example, Eastern Cham /t̥anaw/ ‘pound’, where the subscript circle represents the low register, is either realized as [t̥an̥aw] or [n̥aw], which is evidence that register has generalized from stop-initial syllables to sonorant-initial ones. In order to test if this happened in Mnong Râlâm, we included in the wordlist sonorant-initial words (called register neutral sonorants below), as well as sonorants-initial syllables preceded by voiced/voiceless stops or by presyllables headed by voiced/voiceless stops (called low and high register sonorants below).

Target words were read four times in a fixed frame sentence, randomized in SpeechRecorder (Draxler and Jänsch 2004). As few participants could read Mnong, the first author provided the words in Vietnamese to the participants, who had to translate them in Mnong and pronounce them in the frame sentence in (1). A female native assistant born in 1988, H Hê Bkrông, was present during all the recordings and provided clarification in Mnong whenever necessary.

Recordings were conducted in the Yang Tao communal house, a wooden building located off a dirt road, chosen because it was relatively quiet and offered good recording conditions. Two important signals were acquired during the experiment: an audio signal recorded with a Shure Beta 53 headworn microphone, and an EGG signal recorded with an electroglottograph Glottal Enterprises EG2-PCX. They were channeled through a Steinberg UR44 preamplifier and recorded in SpeechRecorder on a PC laptop. The EGG signal was used to track the onset and offset of voicing; all other measures were obtained from the audio signal. Two additional signals, a larynx height tracking channel recorded with the EG2-PCX and a back-up audio channel recorded with a low-quality microphone, will not be reported here.

All recordings made for this study are available on the Pangloss collection (https://pangloss.cnrs.fr/corpus/Mnong_R%C3%A2l%C3%A2m) and annotated files can be downloaded from the Nakala archive (https://nakala.fr/10.34847/nkl.a35d1if0).

2.1.2 Data analysis

Target words were annotated in Praat Textgrids. The beginnings of the consonant and of the open phase of the rhyme (i.e., the release of the closure in stops and the beginning of the vowel in fricative- and sonorant-headed syllables) as well as the endpoints of the vowel and coda (if any) were labeled based on the acoustic signal. The onset of voicing, as well as any cessation and resumption of voicing were labelled based on the EGG signal. Important landmarks are illustrated in Figure 2.

Figure 2:

Annotation of target word /diːr/ ‘to avoid’. Top: spectrogram; middle: EGG signal; bottom: acoustic landmarks (ps, previous sonorant; cl, closure; op, open phase; cd, coda; ov, onset of voicing; cv, cessation of voicing; rv, resumption of voicing).

Automatic measures were made using Praatsauce, a Praat implementation of VoiceSauce (Kirby 2018; Shue et al. 2011). Besides the duration of each interval and the time of each laryngeal landmark, spectral measures (f0, H1-H2, H1-A1, H1-A3, CPP, F1, F2, F3) were made at each millisecond of the vowel. The effect of formants on harmonic amplitudes were then corrected using the formula in Iseli and Alwan (2004): adjusted measures will therefore be reported with stars (H1*-H2*, H1*-A1*, H1*-A3*). Since the measurements windows were 25 ms wide, the initial and final 12 measures of the vowels were excluded as they are spuriously affected by measures made during the preceding and following segments.

Outliers were excluded in two ways. To detect local tracking errors, f0, F1, F2 and F3 were speaker-normalized into z-scores, and derivatives of these z-scores were obtained. All derivatives exceeding −0.5 z and 0.5 z were deemed to indicate abrupt jumps and were excluded as spurious measures. Global tracking errors were then excluded by calculating means and standard deviations for f0, F1, F2 and F3 for each combination of speaker, vowel and voicing/register and by deleting any measure exceeding the mean by more than three standard deviations. All tracking errors were left blank and any spectral measure (H1*-H2*, H1*-A1*, H1*-A3*) derived from an excluded f0 or formant value was also excluded. The proportion of data removed for each indicator is given in Table 2.

Data recorded from different speakers were normalized to allow their aggregation in charts and in statistical models. Each acoustic measure was z-normalized by speaker. To facilitate interpretation of the results, z-scores were then converted into familiar scales using means and standard deviations obtained from all speakers (weighted mean of all speakers + z-score × mean standard deviation of all speakers), but statistical results are equivalent to those that would be obtained directly from z-scores.

2.1.3 Statistical treatment

We evaluated the statistical significance of the voicing/register contrast and of other linguistic factors on relevant phonetic properties by fitting linear mixed models to the data using the lmerTest package in R (Kuznetsova et al. 2017). The phonetic properties chosen as dependent variables include the VOT of plain stops and the mean f0, H1*-H2*, H1*-A1*, H1*-A3*, CPP, F1 and F2 over the initial 10 sampling points (10 ms) of vowels in syllables with plain stops. As will be shown in §2.2.2, this portion of the vowel consistently contains the most significant differences between voicing/register, as register effects in vowels immediately follow the onset rather than spreading uniformly over the entire vowel. The fixed effects used in the models were voicing/register, vowel quality, and place of articulation of onsets, as well as all two-way interactions of these fixed factors (three-way interactions were omitted as they led to overfitting). Random slopes for both subject and word were also incorporated. These maximal models were progressively simplified from the top down by iteratively dropping the interaction with the smallest F-value in the model’s ANOVA as long as the resulting model had a significantly lower Akaike information criterion (AIC) score than the preceding model, or a higher or equal but not statistically significant AIC difference. It is important to note that we did not attempt to run statistical models on the acoustic properties of words with onsets other than plain stops, either because they do not contrast in voicing/register or, in the case of sonorants, because there were not enough target words to fit meaningful models.

In mixed models conducted on vocalic properties, the intercept was set to the high register (or voiceless) dental onset and the vowel /ɔː/ (i.e., target syllable /tɔː/). The reason for choosing an open mid vowel is that it allows easy comparisons with phonetically close and open vowels. Back /ɔː/ was preferred to /ɛː/ because it had fewer gaps in the wordlist (cf. word list in Appendix 1). The predicted effects of vowels other than /ɔː/ on dependent variables can be computed from the model summaries in Appendix II by adding the estimates of their main effects or interactions to those of the reference vowel /ɔː/.

We used Cohen’s d’s to determine the weight of each acoustic property in the realization of the voicing/register contrast in the first 10 ms of the vowel (Brunelle et al. 2020, 2022], 2024]; Clayards et al. 2008; Cohen 1988; Tạ et al. 2022). Cohen’s d’s were computed by dividing the difference between the vowel-weighted and subject-weighted means of each property for each voicing/register category by its standard deviation. A large absolute Cohen’s d (>|0.8|) indicates a substantial difference between the two distributions under investigation, suggesting they could contribute to maintaining the contrast.

2.2.1 Laryngeal contrasts in onsets

The distribution of VOT in onset stops is reported in Figure 3. Voiceless stops have a short positive VOT (mean = 15.1 ms), as expected, but voiced stops have an unexpected bimodal distribution. Most voiced stops exhibit a strong negative VOT (mean = −21.3 ms) characteristic of true voicing, but a significant proportion also have a positive VOT that is statistically undistinguishable from that of voiceless stops (see Table 3, Appendix II). Unsurprisingly, aspirated stops have a long lag VOT (mean = 47.2 ms). Implosives generally have a long lead VOT (mean = −62 ms) but a non-negligible minority have a slightly positive VOT (42/262, right-hand peak in Figure 3). The reason for this positive VOT is unknown, but they are for the most part prenasalized.

Figure 3:

VOT distribution in different types of stops.

To better understand the bimodal VOT distribution of plain voiced stops, they were broken down into three types: fully voiced stops have voicing throughout their closure; devoiced stops either have voiceless closures or carryover voicing from the previous sonorant over less than 30 % of their closures; voiced stops with a voiceless release have voicing over more than 30 % of their closure but are devoiced before the burst. An example of a voiced stop with a voiceless release was already given in Figure 2.

The breakdown of phonological voiced stops by phonetic realization is given in Figure 4. Few voiced stops are fully devoiced and full devoicing is more prevalent in women than men (yet only F1987 devoices more than half of her voiced stops). Partial devoicing is much more common: most voiced tokens have a voiceless release, especially in women. Tokens with fully voiced closures are rare in women but common in men, especially older ones. Overall, women tend to exhibit more devoicing, full or partial, than men. Since a similar sex asymmetry was obtained in neighboring languages studied with a comparable methodology, it is likely that this partly reflects anatomical differences (Brunelle et al. 2020, 2022], 2024]; Tạ et al. 2022), but the magnitude of the age and sex differences is much greater in Mnong Râlâm. Finally, despite their smaller supraglottal cavity, velars undergo less devoicing than coronals (Coronals: voiced 34.6 %, voiceless release 57.2 % voiceless 8.2 %; Velars: voiced 51.9 %, voiceless release 37.3, voiceless 10.7 %). This tendency is found irrespective of sex or following vowel.

Figure 4:

Proportion of phonological voiced stops that are fully devoiced, have a voiceless release or are fully voiced, by speaker. Speakers are organized by sex (F/M) and year of birth.

Importantly, the duration of voiced stop closure does not seem to be affected by their type of voicing. Fully voiced tokens have a shorter duration (mean: 73 ms) than devoiced stops (mean: 89 ms) and voiced stops with a voiceless release (mean: 89 ms), but this is not statistically significant in a mixed model in which Place, Vowel and Voicing Type and their two-way interactions are included as fixed effects, and Subject and Word as random effects. Inspection of the data reveals that durational differences are attributable to a combination of men’s faster speech rate and of the high prevalence of fully voiced stops in their speech.

The upshot is that Mnong Râlâm voiced stops by and large preserve a voicing contrast characterized by the presence of vocal fold vibrations during the closure, even if partial devoicing is frequent.

2.2.2 Registral properties on vowels

Now that it has been established that closure voicing is still present in onset stops, let us examine the following vowels to determine if there is any evidence that register properties have developed in Râlâm. For the sake of simplicity, we will use the terms high and low registers to refer to the two contrastive series in the rest of this section, without any a priori assumption about the phonological nature of the contrast.

In Figure 5, the f0 of vowels following various types of onsets is reported over their first 200 ms. F0 is marginally higher at the very beginning of the vowel following high register plain stops, but this is only significant in velars (RegisterLow: β = −9.63 Hz, t = −0.908, p = 0.415, RegisterLowPlaceVelar: β = −53.992 Hz, t = −7.327, p = 0.002, see Table 4 in Appendix II).

Figure 5:

Normalized f0 of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

Differences between words with initial sonorants (neutral), and syllables with onset sonorants preceded by different registers are also small: unexpectedly, a low f0 seems to be found on sonorants preceded by voiceless stops or high register syllables. F0 starts very high after fricative /s/, is a bit higher before aspirated than before plain stops and is comparable to that of low register plain stops after implosives.

Voice quality differences between the low and high registers are small and are not captured by all voice quality indicators. In Figure 6, H1*-H2* differences are visible in the first 100 ms after plain stops, with low register stops being a bit breathier or laxer, but they do not reach statistical significance (RegisterLow: β = 2.772 dB, t = 0.753, p = 0.466, see Table 5 in Appendix II). Differences between high and low register sonorants appear even smaller (neutral sonorants pattern with low register ones). Vowels following aspirated stops have a high H1*-H2* indicative of their irregular airflow, while those following fricative /s/ and implosive /ɗ/ exhibit a slightly higher and slightly lower H1*-H2* than neutral sonorants, reflecting their moderately open and constricted glottal settings. Overall, H1*-A1* exhibits the same general tendencies as H1*-H2*, but effects are weaker (see Table 6 in Appendix II).

Figure 6:

Normalized H1*-H2* of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

H1*-A3*, another spectral slope indicator, exhibits larger and more significant register differences in Figure 7, with a significantly steeper slope in the low register after plain stops (RegisterLow: β = 6.095 dB, t = 3.986, p = 0.003, see Table 7 of Appendix II). However, these results should be taken with a grain of salt as 15.8 % of H1*-A3* measures had to be excluded because of tracking errors. H1*-A3* differences are more limited after the different types of sonorants, even if low register sonorants seem a bit breathier than high register ones. Just as for H1*-H2*, H1*-A3* values are very high after aspirated stops, and they seem to indicate some breathiness after fricative /s/ and a tense voice after implosive /ɗ/, which reflects the effect of their respective laryngeal settings.

Figure 7:

Normalized H1*-A3* of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

The last voice quality indicator is cepstral peak prominence (CPP), which correlates positively with periodic glottal cycles and negatively with spectral noise. We see in Figure 8 that it is slightly higher in the high register for the first 50 ms after plain stops, indicating a more modal phonation. This difference does not reach statistical significance in dentals but is more robust in velars (RegisterLow: β = −1.826 dB, t = −2.282, p = 0.087, RegisterLow:PlaceVelar: β = −2.418 dB, t = −4.327, p = 0.013, see Table 8 in Appendix II). Note that similar results are obtained even if we run models with values averaged from measurement windows spanning later portions of the vowel (20–30 ms and 30–40 ms).

Figure 8:

Normalized CPP of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

There are no clear CPP differences between the three types of sonorants either. As for aspirated stops, fricatives and implosives, they all have a low initial CPP, indicating a perturbation of regular phonation at vowel onset because of the laryngeal settings of these consonants.

Turning to formants, F1 differences between registers are robust after plain stops, as can be seen in Figure 9. Overall, low register vowels have a lower F1 than high register vowels following plain stops (RegisterLow: β = −222.563 Hz, t = −3.036, p = 0.039, see Table 9 in Appendix II). Differences between registers appear greater in open and mid vowels in Figure 9 but these additional effects are not significant. Register differences in F1 are negligible after sonorants, except in /aː/. Aspirates, fricative /s/ and implosive /ɗ/ seem to mostly pattern with high register plain stops.

Figure 9:

Normalized F1 of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

There are also systematic register differences in F2 after plain stops (Figure 10). Low register vowels have a higher F2 over their first 50 ms (RegisterLow: β = 154.073 Hz, t = 5.054, p = 0.001, see Table 10 in Appendix II). Register differences in F2 appear negligible in sonorants, and aspirates, fricative /s/ and implosive /ɗ/ mostly pattern with high register stops.

Figure 10:

Normalized F2 of the first 200 ms of vowels following various types of onsets. Thick lines represent means, thin lines individual observations. Aspirates /tʰ, kʰ/, the implosive /ɗ/ and the fricative /s/ do not contrast in register and are included for comparison.

2.2.3 Production weights

The aggregated results presented above conceal a certain amount of interspeaker variation. Figure 11 therefore provides individual cue weights for each acoustic property after plain stops. As can be seen in the top panel, onset voicing is still a crucial phonetic property in Mnong Râlâm. The heaviest cue is either VOT or the proportion of closure voicing (duration of the voiced closure/total closure duration) in all participants except F1987. However, women have much weaker voicing contrasts than men, especially older ones. Moreover, while there is still a robust voicing contrast in all speakers, register properties have non-negligible weights. F1 is the strongest register cue overall, but most speakers also have large weights for F2 and voice quality. The role of f0 seems much more limited, its weight gravitating around 0.

Figure 11:

Cohen’s d’s of each acoustic property associated with voicing and register. Top panel: all properties, with MainVQ indicating the Cohen’s d’s of the strongest voice quality property of each speaker. Bottom panel: detailed Cohen’s d’s of the four measured voice quality properties. Speakers are organized by sex (F/M) and year of birth.

Closing in on voice quality cues, in the bottom panel of Figure 11, we see that there is important interspeaker variation as to what cue has the heaviest weight. H1*-H2* is an important voice quality cue for most speakers but is especially robust in younger speakers. CPP also plays a role, especially in men, while H1*-A1* is stronger in women (H1*-A3* is similar to H1*-A1* but a bit weaker in most speakers). As medial surface thickness of the vocal folds and resting glottal opening are modeled to be good predictors of spectral noise and spectral slope measures (Zhang 2016) and are correlated with sex (Hanson 1995; Patel et al. 2012; Yamauchi et al. 2014), anatomical differences between genders are likely to play a role in some of this variation. However, in the absence of research linking actual laryngeal production with spectral measures, the exact mechanisms involved remain speculative. In any case, Figure 11 shows that the mixed models of the previous section, that are based on aggregated data, slightly underestimate the role of voice quality modulations in the register contrast.

2.2.4 Trade-off between closure voicing and acoustic properties of register

As the results above show that voicing and register are largely redundant in Râlâm, it is important to determine if they are correlated and in what direction. In Figure 12, the main acoustic properties of register are plotted as a function of the proportion of closure voicing in individual voiced stops. When pooling all vowels together, there is no meaningful correlation between the proportion of voicing and H1*-A3* and F2, but the other four indicators show significant correlations. Effects are rather small and vary a bit by vowel, but overall, a larger proportion of closure voicing is associated with acoustic properties characteristic of the low register (a lower f0, a higher H1*-H2*, a lower CPP and a lower F1). This indicates that robust closure voicing co-occurs with salient low register properties. As such, there is no trade-off between closure voicing and register in Mnong Râlâm. Register does not replace voicing; rather, robust voicing is accompanied by robust register cues.

Figure 12:

Acoustic properties of register after phonologically voiced stops, averaged over the first 10 ms of vowels. Black regression line represents correlations for all vowels, while colored regression lines show correlations by vowel.

2.2.5 Summary of acoustic results

The results presented above portray a language in which plain voiced stops show significant signs of devoicing but in which full devoicing is still rare. In fact, closure voicing appears to be the dominant cue distinguishing the two stop series in all but one speaker (F1987). However, despite this limited devoicing, there is clear evidence of registral developments. Register is mostly realized through formant differences, with F1 weighing much more than F2: vowel onsets are higher and more fronted after voiced than voiceless stops, resulting in moderate ongliding. This ongliding appears slightly greater in open and mid vowels than in close ones, but is not as dramatic as in other register languages. Impressionistically, it is audible in low register /aː/ (which is realized as [^εaː]) but is more discrete in other vowels.

Other acoustic properties show more subtle patterns. Voice quality differences are not negligible, but individual variation makes them appear less significant than they actually are in the statistical models of §2.2.2 because the cues that best reflect them vary across speakers. CPP is more important in men, while H1*-A1* and H1*-A3* weigh more in women. H1*-H2* differences, on the other hand, seem to be age-dependent and are more robust in younger than older speakers. Finally, f0 is lower after voiced velar stops than voiceless velar stops, but there is no significant difference after dental stops.

Registral differences after other onsets were not tested statistically but there is limited evidence for register spreading through sonorants. Measures made in vowels following aspirates, the fricative /s/ and the implosive /ɗ/ confirm that they have particular laryngeal settings. Aspirates and /s/ are produced with a wider glottis than plain stops and sonorants, resulting in a breathier voice quality, while /ɗ/ is produced with slightly tighter glottis than high register stops.

3.1.1 Stimuli

Stimuli were synthesized with Klattgrid synthesis implemented in Praat (Boersma and Weenink 2010). Two minimal pairs were synthesized: /taː/ ‘at’ ∼ /daː/ ‘wild duck’ and /tuː/ ‘tomb’ ∼ /duː/ ‘to return’. As the production results showed largely redundant voicing and register, /d/ will be used to refer to the low register plain voiced stop in the rest of this section, without presuming of the perceptual weight of the various properties.

We manipulated the acoustic parameters that had the heaviest weights in the acoustic study: voice quality, F1, F2 and voicing. We did not manipulate f0, as it is not significant in dental stops and as adding a fifth synthesized acoustic property would have made the experiment too long. Stimuli were synthesized with a middle-aged male voice and were based on the natural productions of a male speaker born in 1987. Practically, this means that formant targets are lower than the averages presented in §2.2.2 and that we prioritized H1*-H2* and CPP as voice quality cues, as they have the heaviest production weights in younger men. Vowel duration was set to 300 ms. Three-step continua were generated for each property using the following parameters:

1. Voicing. Three laryngeal settings were generated: (1) a stop with full closure voicing (70 ms), (2) a voiced stop with a voiceless release: voicing over the first 40 ms of its closure, 30 ms of voicelessness at the end of the closure and 10 ms of aspiration after the release, and (3) a stop with a 70 ms voiceless closure and 10 ms of aspiration after the release.

2. Voice quality. Voice quality was manipulated by varying the open phase of the glottal waveform (or open quotient, OQ, in Klattgrid), a parameter that affects spectral tilt. At vowel onset, the breathy step had an OQ of 0.58, the middle step had an OQ of 0.5 and the modal step had an OQ of 0.43. All 3 synthesized steps then reached an OQ target of 0.5 at 150 ms. Since the Mnong Râlâm low register is not very breathy, it was not necessary to boost the breathiness of low register stimuli by adding breathiness amplitude, contrary to what was done in identification studies conducted on other languages by the same team (Brunelle et al. 2020, 2024]; Tạ et al. 2019).

3. F1. For /aː/, targets at vowel onset were 500, 625 and 700 Hz. All /aː/ stimuli then reached 700 Hz at 150 ms and gradually dropped to 650 Hz at vowel end. For /uː/, targets at vowel onset were 400, 450 and 500 Hz. All /uː/ stimuli then dropped to 375 Hz at 150 ms and remained stable until vowel end.

4. F2. For /aː/, targets at vowel onset were 1,400, 1,450 and 1,500 Hz. All /aː/ stimuli then reached 1,450 Hz at 150 ms and gradually dropped to 1,400 Hz at vowel end. For /uː/, targets at vowel onset were 1,200, 1,275 and 1,350 Hz. All /uː/ stimuli then dropped to 950 Hz at 150 ms and remained stable until vowel end.

5. f0. Pitch did not vary across stimuli. It started at 120 Hz and dropped linearly to 110 Hz at vowel end in all syllables.

All possible combinations of these acoustic values were synthesized, yielding 81 stimuli (3 voicing steps X 3 voice quality steps X 3 F1 steps X 3 F2 steps). Spectrograms of stimuli representing high register /taː/ and low register /daː/ are given in Figure 13. The reliability of the synthesis parameters was checked by measuring the stimuli with PraatSauce (Kirby 2018). The distribution of the stimuli along the most relevant acoustic dimensions is given in Figure 14.

Figure 13:

Spectrograms of sample stimuli. Top: Stimulus mirroring natural productions of high register /taː/ (targets at vowel onset: OQ 0.43, F1 700 Hz, F2 1,400 Hz). Bottom: Stimulus mirroring natural productions of low register /daː/ with voicing (targets at vowel onset: OQ 0.58, F1 550 Hz, F2 1,500 Hz).

Figure 14:

Mean values of the acoustic parameters manipulated in the stimuli used for the identification experiment. Top panel: /taː∼daː/. Bottom panel /tuː∼duː/. The ribbons show one standard deviation above and below the mean.

3.1.2 Participants and procedure

A total of 53 participants took part in the identification experiment (32 women, 21 men). They were all born between 1944 and 2003. None of the participants reported hearing problems, and none were obvious to the experimenters (even if it can be assumed that older listeners have normal aged-related hearing loss). Besides Mnong, all listeners spoke Vietnamese and 33 spoke Rade with variable fluency. They were all born in the district of Yang Tao, except one male participant who was born in the neighboring city of Buôn Ma Thuột. Eight of them had lived in other regions of Vietnam for more than a few months (Buôn Ma Thuột, Tây Ninh, Bình Dương, Hồ Chí Minh City). Three female participants were excluded because they were not able to complete the experiment. Out of the 50 listeners, 18 had participated in the production experiment three years earlier.

The procedure took place in a community house where up to three participants completed the task simultaneously. The experiment was conducted in OpenSesame on Mac or PC laptops and sounds were played through Sennheiser 280D PRO headphones. Participants had to listen to stimuli and to identify them by pressing one of two keyboard buttons associated with images representing response choices. There were three training phases: one with three repetitions of the two stimuli most closely mirroring natural productions (with feedback), one with five repetitions of the same two near-natural stimuli (without feedback), and one with ten random stimuli. Training responses are not reported here. Participants then had to identify each set of eighty-one stimuli three times, in alternating blocks. As few participants could read Mnong, visual instructions accompanied by Vietnamese text were provided.

3.1.3 Analysis

Mixed logistic regressions were used to analyze the identification results, by syllable. The dependent variable was the response provided by participants. The fixed effects were the types of voicing (negative, voiceless release and voiceless) as well as voice quality (OQ), F1 and F2 steps. Random slopes for each main effect by participant were also included. Models were simplified using the same top-down approach as in §2.1.3: interactions were dropped one by one, starting with that with the lowest F-value, as long as the resulting models had a lower Akaike information criterion (AIC) score than the previous model, or a higher or equal but not significantly different AIC.