Acoustic correlates of Burmese voiced and voiceless sonorants

1.1 Burmese inventory

Burma, also known as Myanmar, is a country in Southeast Asia with a population of approximately 50 million people, of which over 30 million are Burmese speakers (Bradley 1996). Burmese, a Tibeto-Burman language, has 34 consonants, as shown in Figure 1, and two sets of vowels, as shown in Figure 2 (Watkins 2001). The first set of vowels consists of eight monophthongs that are not nasalized and appear in open syllables. The second set of vowels consists of eight monophthongs and diphthongs that are either nasalized in an open syllable (except for [ɛ]) or not nasalized in a closed syllable. Of these vowels, [a] is the most common. Burmese is also a tonal language with four attested tones: high, low, creaky, and killed (Watkins 2000).

Figure 1:

Burmese consonants, with voiced-voiceless sonorant pairs highlighted. Figure based on Watkins (2001).

Figure 2:

First set of vowels (left) that are non-nasal in open syllables, and second set of vowels (right) that are nasal in open syllables (except for ɛ) or non-nasal in closed syllables. Figure based on Watkins (2001).

1.2 Voiceless sonorants

Burmese has 12 sonorant consonants with contrastive voicing: [m, m̥, n, n̥, ɲ, ɲ̥, ŋ, ŋ̥, l, l̥, w, w̥] (Watkins 2001). The voiceless sonorants have been described as having two parts: a voiceless portion followed by a voiced portion (Bhaskararao and Ladefoged 1991; Dantsuji 1984). Unlike some languages with voiceless sonorants, the phonemic status of Burmese voiceless sonorants is not questioned (Ladefoged and Maddieson 1998; Watkins 2001).^[1] Work on Burmese voiceless sonorants so far has focused on the nasal sonorants [m̥, n̥, ɲ̥, ŋ̥] (Bhaskararao and Ladefoged 1991; Dantsuji 1984), or some subset of the nasal sonorants (Chirkova et al. 2019; Ladefoged 1971). Only Maddieson (1984) considers [l̥] in addition to the nasal sonorants, but this analysis focuses on fundamental frequency (F0), ultimately finding that F0 as measured at the post-sonorant vowel onset is lower for voiced sonorants and higher for voiceless sonorants, in line with cross-linguistic findings that voiceless segments tend to be associated with a higher F0 (Kingston 2011). Other studies on Burmese voiceless sonorants consider durational measurements (Bhaskararao and Ladefoged 1991; Dantsuji 1984), oral and nasal airflow measurements (Bhaskararao and Ladefoged 1991; Chirkova et al. 2019), and EGG measurements (Chirkova et al. 2019). While these works have helped to further our understanding of Burmese voiceless sonorants, they focus primarily on the temporal and aerodynamic aspects of voiceless nasals. Therefore, there remains a gap in the literature which can be filled by a more thorough acoustic analysis of Burmese voiceless sonorants.

The analysis of Burmese sonorants can be contextualized by comparing them to accounts of voiced and voiceless sonorants in other languages. Cross-linguistically, voiceless sonorants are found most commonly in Tibeto-Burman languages, such as Burmese, Mizo, and Angami (Bhaskararao and Ladefoged 1991; Chirkova et al. 2019), and Indo-European languages, like Icelandic and Welsh (Bell et al. 2021; Jessen and Pétursson 1998).^[2] Previous work has found some consistent acoustic correlates with voiceless sonorants. For example, voiceless sonorants tend to have a longer duration (considering the voicelesss and voiced portions together), as in Khonoma Angami, where voiceless nasals have been found to be 7–10 % longer than their voiced counterparts, (Blankenship et al. 1993), and in Icelandic, where voiceless sonorants have been found to be an average of 42 % longer (Bombien 2006). Voiceless sonorants also tend to have a higher spectral tilt than their voiced counterparts, as in Icelandic, where voiceless sonorants have a higher H1–H2, signifying a more breathy production (Bombien 2006). They also tend to have more spectral noise, as has been observed both in Iaai (Maddieson and Anderson 1994) and in Icelandic, where voiceless sonorants have more frication (Bombien 2006).^[3] Taken together, these studies suggest that there are some common cross-linguistic features of voiceless sonorants. However, voiceless sonorants tend to have very little acoustic information segment-internally, instead being largely identifiable based on adjacent segments. This is similar to how many correlates of contrastive breathy sonorants in Marathi are realized on the following vowel, rather than on the sonorant itself (Berkson 2019), or how stop voicing contrasts in English are correlated with duration and F0 on adjacent vowels (Kingston 2011; Lisker 1972).

In addition to cross-linguistic accounts of voiceless sonorants, historical reconstructions also provide important information for the study of voiceless sonorants. Blevins (2018) has argued that voiceless sonorants have their origins in sonorant-SG clusters, which ultimately phonologize to become a contrastive phoneme. Put another way, most voiceless sonorants, including Burmese voiceless sonorants (Dantsuji 1984), historically were clusters consisting of an /h/ + sonorant or sonorant + /h/, which were eventually reanalyzed as a single phoneme, with all previous environment restrictions lost. Thus, SG gestures seems to be a key element of voiceless sonorants, but existing acoustic accounts do not focus on this aspect.

Taken together, it seems that perhaps Burmese voiceless sonorants are better described as having a SG period. This is compatible with both historical reconstructions and existing descriptions of the voiceless sonorants. This is also in accordance with the cross-linguistic link between preaspiration (which results from a SG gesture) and sonorant devoicing. For example, Hansson (2001) found that Scandinavian languages with preaspiration also invariably have sonorant devoicing and Stevens and Hajek (2004) replicated this finding with Sienese Italian, which was been found to have preaspiration with voiceless geminate stops in addition to sonorant consonant devoicing before voiceless stops. In both the Scandinavian languages and in Sienese Italian, the voiceless sonorants are realized with some degree of aspiration, which Stevens and Hajek (2004) argue “confirms the phonetic similarity between sonorant devoicing and preaspiration,” and which Hansson (2001) argues could be evidence for preaspiration and sonorant devoicing being a unified and general process. However, at least in the case of Sienese Italian, sonorant devoicing and the corresponding realization of aspiration is not ubiquitous, and indeed only occurs 85 % of the time (Stevens and Hajek 2004). The variable realization of preaspiration has been called ‘non-normative preaspiration’ (Helgason 2002), and it has been noted for a variety of contrasts. For example, it has been observed with Norwegian and Central Standard Swedish stop voicing contrasts (van Dommelen 1998; Helgason 2002), Itunyoso Trique and Sienese Italian stop gemination contrasts (DiCanio 2010; Stevens and Reubold 2014), and Central Standard Swedish fricative voicing contrasts (Gordeeva and Scobbie 2010). It is therefore possible that the [−voice] period described in Burmese could be an instance of a non-normative SG gesture, similar to preaspiration. However, if the SG gesture is non-normative, voiceless sonorants will not consistently be produced with turbulent airflow. The implications of this are discussed further in Section 1.4.

1.3 Acoustic correlates of voicelessness and SG gestures

In order to acoustically characterize Burmese voiced and voiceless sonorants, a variety of different acoustic correlates must be considered. In line with previous work on laryngeal contrasts (Berkson 2019; Garellek 2020) and voiceless sonorants (Blevins 2018; Bombien 2006), duration, H1–H2, strength of excitation, harmonics-to-noise ratio, and F0 are considered with respect to Burmese sonorants. The remainder of this section provides an outline of what each of these measures indicates and how they are expected to pattern for voiced and voiceless sonorants.

1.4 Predictions

Given the information laid out so far, predictions can be made for the stated research questions. With respect to the first question (what acoustic correlates distinguish voiced and voiceless sonorants?), acoustic correlates were chosen that are likely to capture the most acoustic information on voiceless sonorants based on previous accounts: duration, H1–H2, SoE, HNR, and F0. The correlates of these measures and the predictions for how they will correspond to voiceless sonorants compared to voiced sonorants are laid out in Table 1. Measurements will be taken both on the sonorant itself, and on the preceding and following vowels. This will allow for a deeper understanding of where correlates of voiceless sonorants can be realized. Given findings from Blevins (2018) and Berkson (2019), it is likely that some correlates of voicing category will be realized on the sonorant and the vowel following the sonorant. It can also be anticipated that a period of voiceless turbulent airflow will generally occur immediately preceding the voiceless sonorant, in line with previous descriptions of Burmese, e.g. Dantsuji (1984), who noted a period of voiceless nasal frication before voiceless nasals, and in line with descriptions of voiceless sonorants in other languages (Bell et al. 2021; Blevins 2018; Bombien 2006). If Burmese voiceless sonorants are indeed better thought of as having a SG gesture, then we might expect both correlates related to the SG gesture (shorter duration, lower HNR) and voicelessness (lower SoE, higher F0) to appear.

To answer the second question (are there multiple realizations of voiceless sonorants and what acoustic correlates distinguish them?), the first task is to establish the possible realizations of voiceless sonorants, particularly with respect to the SG gesture. If the SG gesture is always present, then it can be concluded that the period of frication is normative in Burmese, with no further investigations necessary. If, however, it is non-normative, then a deeper investigation of the cues present for voiceless sonorants with the SG gesture and voiceless sonorants without the SG gesture is called for. To preview the results, data suggests that there are two realizations of voiceless sonorants in Burmese. Voiceless sonorants are realized with a SG gesture 78 % of the time, but 22 % of the time no SG gesture is present.

There are a few possible explanations for what could be happening when the SG gesture is absent from voiceless sonorants. First, it is possible that there is gestural shifting, such that the SG gesture is timed completely concurrently with the sonorant or an adjacent vowel, and thus it cannot be segmented as distinct from the voiced portion of the sonorant or an adjacent vowel. If the SG gesture is shifting, we would expect to see correlates of SG, such as a lower HNR, on one of the adjacent segments. A second option is that the phonologically voiceless sonorants without the SG gesture are mispronunciations, and are articulated exactly the same as voiced sonorants. If this is the case, we would not expect significant differences in any correlates for voiced versus voiceless non-SG sonorants. Third, it is possible that the SG gesture is absent, but other correlates of voicelessness remain. This would mean correlates associated with the SG gesture (such as a lower HNR or a difference in duration) would not be significantly different from voiced sonorants, but those associated with voicelessness (such as higher F0 and lower SoE) would be different.

2.1 Stimuli

The acoustics of Burmese voiced and voiceless sonorants were examined in a production study with speakers of Burmese. Target words were all monosyllabic minimal pairs consisting of every combination of 8 consonants ([w, w̥ , l, l̥, ŋ, ŋ̥, m, m̥]), 2 tones (high and low), and one nucleus vowel ([a]).^[5] In order to ensure that target words were as phonologically controlled as possible, the tone, consonant onset, and vowel were balanced, but words were not balanced based on syntactic category. Target words can be found in Table 2. Each target word appeared in 9 different carrier phrases, 3 for each phrasal position (phrase-initial, phrase-medial, and phrase-final), but in order to limit differences caused by prosodic effects, and because prosodic position was not the focus of the current study, only target words in phrase-medial position were considered for the present analysis. Phrases were controlled so that each phrase had the same number of syllables, as shown in Table 3, and target sonorants were preceded and followed by [a]. It is worth noting that this resulted in some phrases that were strange or syntactically marked, but participants had no problem reading the phrases fluently due to the transparency of Burmese orthography. Furthermore, all stimuli were checked by a speaker of Burmese and there is no evidence that the content of the phrases impacted the pronunciation of target words or adjacent segments. Altogether, these combinations resulted in 48 target phrases, such as [ŋa la hãʊ̃] ‘my old mule.’ Once all the stimuli were compiled, three lists were created with the stimuli randomized in different orders. Participants were randomly assigned to one of the three lists, for a total of three participants per list.

Table 2:

Target words differing in terms of consonant, voicing, and tone.

Table 3:

Carrier phrases used to present stimuli.

2.2 Participants

Nine speakers of Burmese between the ages of 19 and 42 were recorded, and all were living in New York. Table 4 shows demographic information about participants, including their age at the time of the 2021 study, their gender, and the number of years they had been living in the United States. All participants reported normal speech and hearing as well as literacy in Burmese. Participants were contacted primarily through personal channels and Facebook groups and were reimbursed $25 for participating in the hour-long study. No participants were excluded, although occasionally participants skipped a phrase, resulting in uneven N values between participants.

2.3 Procedure

Phrases were presented to participants in Burmese orthography one at a time in a slideshow. Participants were instructed to read the phrase to themselves first, and then out loud once they had familiarized themselves with the phrase. They were told that they were not reading complete sentences and that they should read the phrases as if they were responding to a friend, in order to encourage a more casual speaking style. Participants were given a break after every 10 phrases. All recordings were taken in a sound-attenuated booth using a Zoom H4n Pro recorder and a head-mounted mic. The recordings were done with a bit depth of 16-bit for a sampling rate of 44,100 Hz.

2.4 Labeling and measurements

After collecting the data, target words were labeled in Praat (Boersma and Weenink 2015). These words were further segmented for the preceding vowel, sonorant (further subdivided into SG portion, where relevant, and voiced portion), and following vowel, as shown in Figure 3. The SG portion (hereafter “SG”) was segmented according to the beginning and end of clear frication in the spectrogram and waveform. The voiced portion (hereafter “sonorant”) was segmented separately in order to get reliable acoustic measures, and segmentation was done on the basis of the beginning and end of the second formant. This is in line with measurements in Bhaskararao and Ladefoged (1991), where the voiceless and voiced portions are measured separately. Finally, vowels were segmented based on the beginning and end of the second formant in order to reliably attain voice quality measures. The categories are referred to hereafter as “preceding vowel,” “SG,” “sonorant,” and “following vowel.”

Figure 3:

Segmentation of the phrase [ŋa l ◦ a w ã і ∽ ] ‘my round sword’ with SG.

The averaged acoustic measures for each of these labeled portions were obtained using VoiceSauce (Shue 2010). The measures considered in the analysis were duration, H1–H2, SoE, HNR < 3500 Hz, and F0, in line with recent work by Berkson (2019) and Garellek (2020). Averaged data was considered rather than time-course data in order to establish a baseline acoustic description of voiceless sonorants and to allow for linear discriminant analysis classification.

Linear mixed-effects regressions were carried out in R using the lme4 package to determine how accurately the voiced and voiceless sonorants can be classified using these acoustic correlates (Baayen et al. 2008). Linear mixed-effects models consider how the combination of fixed and random effects correlate with a given variable. This method has been regularly used in the phonetics literature to examine topics such as Marathi breathy sonorants (Berkson 2019) and Spanish vowel nasalization (Bongiovanni 2021). The two tones and four sonorants were collapsed in these models to allow for model convergence and interpretable results. Data was collected from multiple tones and consonant onsets in order to gather a more representative sample of Burmese sonorants without the experiment becoming too onerous for participants.

A linear discriminant analysis (LDA) was also run in R using the MASS package lda() function to determine the acoustic correlates that were best able to predict the voicing status of the sonorant (Venables and Ripley 2002). LDA is a statistical classification tool that, given a set of features, finds a linear combination of those features that successfully distinguishes a set of predefined variables. This methodology has been increasingly used in the phonetics literature, with researchers using it to study topics such as Yoloxóchitl Mixtec fricatives (DiCanio et al. 2020) and the complex phonation system of !Xóõ (Garellek 2020). This methodology is well-suited to address Burmese sonorants, as the key correlates of voiceless sonorants are not yet clear. It seems likely that Burmese voiced and voiceless sonorants differ in terms of a variety of acoustic correlates, which have different weights in terms of their importance in classifying the sound.

3.1 Acoustics of Burmese sonorants

Spectrograms of Burmese sonorants, shown in Figure 4, reveal one typical realization of voiced sonorants and two common realizations of voiceless sonorants. An inspection of the top left spectrogram reveals that Burmese voiced sonorants have formant patterns typical of voiced sonorants cross-linguistically. Voiceless sonorants, on the other hand, are generally produced with a two-part structure: a SG gesture followed by the voiced portion of the sonorant, as shown in the top right spectrogram. This results in a realization like [^hla]. However, in 22 % of voiceless tokens, there was no evidence for a SG gesture, as shown in the bottom spectrogram. The proportion of voiceless sonorants with a SG gesture to voiceless sonorants without a SG gesture is largely consistent between participants.

Figure 4:

/la/ ‘mule’ (top left), / l ◦ a/ ‘sword’ (top right) with SG, and / l ◦ a/ ‘sword’ (bottom) without a SG gesture. All are spoken by the same participant.

The overall averages of each acoustic correlate for voiced and voiceless sonorants and the overall averages of each acoustic correlate for voiceless SG and voiceless non-SG sonorants are reported in Table 5 and Table 6, respectively. There is a high degree of consistency within and across speakers for how voiced sonorants are realized, and by-participant details are reported in the Appendix (Table 12). The average duration of the SG period (when present) was 100.74 ms.

3.2 Distinguishing voiced and voiceless sonorants

Linear mixed-effects models were run with the goal of determining whether voicing is predictive of differences in the acoustic correlates. Models were run on data from the preceding vowel, the sonorant, and the following vowel. These models took the acoustic correlate (duration, H1–H2, SoE, HNR, and F0) as the dependent variable, voicing (voiced or voiceless) as a fixed effect, and participant and word as random intercepts. The baseline for all the models was a voiced sonorant. The output of the models is summarized in Table 7. The coefficients (β and z) reflect how much change is indicated when the sonorant is voiceless compared to voiced. The results suggest that both the sonorant and following vowel have acoustic correlates that distinguish voiced and voiceless sonorants. Measures for different voiceless sonorant realizations can be seen in Table 6.

First, voiced sonorants (121 ms) are significantly longer in duration compared to the voiced portion of voiceless sonorants (75.5 ms). Without the SG portion, the voiceless sonorants are on average 45.5 ms shorter than the voiced sonorants. However, if the average 100.74 ms SG gesture were included, the voiceless sonorants would be on average 55.24 ms longer than the voiced sonorants. In fact, when duration includes both the SG portion and the voiced sonorant portion, there is a significant difference between voiced and voiceless sonorants such that voiceless sonorants are longer (β = 40.21, z = 5.96, p < 0.01). This is in line with previous work on Burmese.

F0 as measured over the sonorant also has significant differences, such that the voiced sonorant portion (192 Hz) have a significantly higher F0 than phonemically voiced sonorants (171 Hz). This is consistent with previous literature that has found a cross-linguistic correlation between voiceless phonemes and raised F0 (Kingston 2011).

Finally, SoE as measured over the vowel following the sonorant was found to be significantly different depending on if the sonorant was voiced or voiceless. Vowels following voiced sonorants (SoE = 0.006) had significantly higher SoE than vowels following voiceless sonorants (SoE = 0.005), which suggests that voiced sonorants are associated with a greater degree of voicing. It is interesting to note that this difference is only significantly different when measured on the following vowel, and not on the sonorant. This could be due to the sonorant being a poor carrier of this acoustic correlate, as has been discussed in previous literature on voiceless sonorants (Blevins 2018).

Finally, a linear discriminant analysis (LDA) was carried out with the goal of determining how accurately a model can distinguish between the two phonological categories and what the most important cues were for distinguishing them. The LDA was given information about average duration, H1–H2, SoE, HNR, and F0 for the preceding vowel, sonorant, and following vowel of each token. It was also provided with information about the duration of the SG gesture, which was 0 if there was no SG gesture. This resulted in the LDA having information about whether a sonorant was voiced or voiceless (based on the production as intended), and a total of 16 acoustic correlates. After the model was trained on 80 % of the data, it was tested on 20 % of the data to determine how accurately it was able to distinguish between voicing categories. It had a success rate of 86.7 % in accurately categorizing a given token as either voiced or voiceless, correctly categorizing voiced tokens 99.4 % of the time, and correctly categorizing voiceless tokens 73.6 % of the time. Within the voiceless sonorants, voiceless sonorants with a SG gesture were correctly categorized 93.7 % of the time, while those without a SG gesture were correctly categorized only 5.4 % of the time. All coefficients of the linear discriminant are reported in Table 8, with negative values reflecting a correlate that is more predictive of voiced sonorants, and positive values reflecting a correlate that is more predictive of voiceless sonorants. Taken together, these results suggest that the correlates examined in the present study are adequate for accurately and consistently distinguishing voiced and voiceless sonorants. Of these correlates, the duration of the SG period is the most important. However, other correlates may also play a role; most notably, F0 of the sonorant, the duration of the sonorant, and SoE as measured on the following vowel, which were the same cues that were found to statistically differentiate voiced and voiceless sonorants in the linear mixed-effects models. It is also worth noting that acoustic correlates measured on the preceding vowel were consistently found to be among the least useful to the model, suggesting that the preceding vowel does not hold much acoustic information for sonorant voicing.

3.3 Variable realizations of voiceless sonorants

Having understood the acoustic correlates that distinguish Burmese voiced and voiceless sonorants, the next step is to carry out a closer analysis of voiceless sonorants. As demonstrated in Figure 4, voiceless sonorants do not surface in only one way. 78 % of the time they surface with a SG gesture, but 22 % of the time they do not, with remarkable between-speaker consistency and no significant difference based on gender (a SG gesture was present 75 % of the time for women and 82 % of the time for men). As established in Section 1.4, there are a few possibilities for the realization of voiceless non-SG segments: (1) the SG gesture is mistimed, and shifts to an adjacent segment, (2) there is a mispronunciation in which a voiced sonorant is produced, or (3) the SG gesture is absent but other laryngeal correlates of voicelessness remain. The results of linear mixed-effects regressions are reported below with the goal of providing clarity on this issue.

Linear mixed-effects models were run for each acoustic correlate with a three-way divide for voicing (voiced, voiceless SG, and voiceless non-SG) as a fixed effect. Once again, these models used data from the sonorant and adjacent vowels, and took the acoustic correlate (duration, H1–H2, SoE, HNR, and F0) as the dependent variable, and participant and word as random intercepts. A post-hoc test considering the differences between the three voicing categories was carried out with the emmeans package in R using the Holms correction for multiple comparisons. The post-hoc results are reported in the Appendix (Table 13) and support the significant results found in the linear mixed-effects regressions. Average values of acoustic correlates for voiceless SG and voiceless non-SG sonorants are presented in Table 6.

The output for voiceless SG sonorants is summarized in Table 9, where voiced sonorants were the baseline. Duration, HNR, and F0 as measured over the sonorant are significantly different between the two categories, and SoE as measured over the vowel following the sonorant is significantly different between the two.

The output for voiceless non-SG sonorants is summarized in Table 10, where voiced sonorants were the baseline. The sonorant had significant differences in F0, and the following vowel had significant differences in SoE. This suggests that while the correlates associated with SG gestures are no longer significant, correlates of voicing, such as F0 and SoE, remain relevant.

Finally, the output comparing voiceless SG sonorants and voiceless non-SG sonorants are summarized in Table 11. Factors were re-leveled, such that voiceless SG sonorants were the baseline, allowing for a direct comparison between the two. The models show that the sonorant itself has significant differences in duration, HNR, and F0 between the two categories, and F0 as measured over the following vowel is also significantly different.

Duration of the sonorant is significantly different between the voiceless SG sonorants and the two other voicing categories, such that the voiced portion of voiceless SG sonorants (61.3 ms) is significantly shorter than voiceless non-SG sonorants (124 ms) and voiced sonorants (121 ms), as shown in Figure 5. This is likely due to the fact that the SG period results in a shorter voiced portion. However, even when duration includes the voiceless SG portion and the voiced portion, there is still a significant difference between voiceless SG sonorants and the other two categories, such that voiceless SG sonorants are longer than both voiced (β = −51.22, z = −8.15, p < 0.01) and voiceless non-SG (β = −49.07, z = −6.910, p < 0.01) sonorants.

Figure 5:

Duration measured over the sonorant. In this figure and the following figures, lines in the box plot represent the median, dots represent the mean, and stars represent a significant difference.

HNR measured over the sonorant also differs significantly, such that the voiced portion of voiceless SG sonorants (45.1 dB) have a significantly lower HNR than either voiced sonorants (48.5 dB) or voiceless non-SG sonorants (49.6 dB), as can be seen in Figure 6. This suggests that voiceless SG sonorants are less modally voiced than either of the other voicing categories. Interestingly, this is the only result that is significant for the voiced versus voiceless SG comparison but not for the general voiced versus voiceless comparison. The lack of a significant difference for voiced versus overall voiceless comparison is likely due to the fact that voiceless non-SG sonorants do not have the SG gesture, and therefore don’t have the lower HNR associated with noise. Since a portion of the voiceless sonorants do not have a SG gesture, it could impact the overall voiced versus voiceless HNR comparison.

Figure 6:

HNR measured over the sonorant.

F0 measured over the sonorant was significantly different for all three voicing categories, with voiceless SG sonorants (195 Hz) having the highest F0, followed by voiceless non-SG sonorants (183 Hz), and finally voiced sonorants (171 Hz), which had the lowest F0. This is reflected in Figure 7. Both voiceless sonorant variants have a higher F0 than the voiced sonorant, in line with the cross-linguistic pattern of voicelessness being associated with a higher F0. The difference observed between the voiceless sonorants, on the other hand, closely mirrors patterns observed in other languages, where F0 is higher with SG segments than modally voiced segments (Ladd and Schmid 2018). This result helps confirms the idea that F0 is correlated with both voicing and SG, and that these two processes can be separate.

Figure 7:

F0 measured over the sonorant.

Turning to the acoustic correlates on the following vowel that are significantly different between voicing categories, SoE is significantly higher for voiced sonorants (SoE = 0.0060) than either voiceless SG sonorants (SoE = 0.0050) or voiceless non-SG sonorants (SoE = 0.0052), as shown in Figure 8. Put another way, the vowel has significantly higher amplitude voicing following a phonemically voiced sonorant as compared to a phonemically voiceless sonorant. There are no significant differences in SoE between the voiceless SG and voiceless non-SG sonorants for the following vowel.

Figure 8:

SoE measured over the following vowel.

Finally, there is a significant difference in F0 measured over the following vowel between voiceless SG sonorants (173 Hz) and voiceless non-SG sonorants (168 Hz), but not voiced sonorants (164 Hz), as shown in Figure 9. This finding reflects the same pattern observed for F0 as measured over the sonorant for voiceless SG and voiceless non-SG sonorants, and suggests that, in line with previous work, segments with a SG gesture tend to have a higher F0 than non-SG segments, even if there is no corresponding difference in phonological voicing category (Ladd and Schmid 2018).

Figure 9:

F0 measured over the following vowel.