Voice register in Mon: experiments in production and perception

2.1.1 Participants and data collection

We collected acoustic and electroglottographic (EGG) recordings from 17 Burma Mon speakers (6 females and 11 males). They were born between 1970 and 2002 (21–53 years old at the time of the experiment). Participants were from different townships in Mon state: 7 speakers from Paung township, 3 speakers from Mudon township, 2 speakers from Thanbyuzayat township, 2 speakers from Ye township and 1 speaker from Chaungzon (see Figure 1). All participants were native speakers of Mon, with Mon being their primary language. Additionally, they were fluent in Burmese, and all possessed at least a basic conversational proficiency in Central Thai.

Figure 1:

Map of townships in Mon state by the authors using Mon State Village Tract Boundaries MIMU v9.4 (Myanmar Information Management Unit 2023).

During the recording sessions, participants were presented with target words in Mon script, accompanied by corresponding Burmese translations, on a laptop screen. Speakers were instructed to produce the target words within a frame sentence, as shown in 2.1.1. Each target word was repeated three times, except for one speaker (M01) who provided two repetitions and another speaker (F05) who provided only one repetition.

The target words consisted of attested Mon monosyllabic words or polysyllabic words with target syllables in final position. The word list was verified with native Mon speakers. Target words and syllables were selected to contain all possible combinations of coronal and velar stop and nasal onsets, as well as high and low registers. We excluded words with voiceless sonorants to limit the duration of the task. We sought a balance between target words occurring in open and closed syllables, with nasal, stop, /h/, and /ʔ/ codas. While prioritizing words with an /a/ nucleus, in cases where such words were not available, we opted for substitutes containing other non-high vowels such as /e ɤ o oi ɛa ɔ ɑ/. It is noteworthy that the vowels /o/ and /ɔ/ in open syllables exhibit variation in vowel height across speakers, regardless of register. For example, some speakers raised /o/ to /o̝/ and /ɔ/ to /ɔ̝/. We therefore analyzed /o/ and /ɔ/ in open syllables separately from those in closed syllables. The final list included 55 words (see Appendix A).

Recordings were made with the SpeechRecorder software (Draxler and Jänsch 2004). The audio signal was captured direct to disk at a sampling rate of 44.1 kHz through a USB digital audio interface connected to a head-mounted microphone. Simultaneously, the EGG signal was captured via a EGG-D800 device from Laryngograph Ltd. The recording sessions took place in a quiet room at the data collection site.

2.1.2 Annotation

All recordings were reviewed by research assistants to identify speech errors, assess recording quality, and verify the presence of the intended target words. Of the 2,640 initial recordings, 21 were excluded due to speech errors. The remaining 2,619 audio recordings were force aligned using the MAUS language-independent model (Schiel 1999). Subsequently, the TextGrids generated by MAUS were manually corrected as needed using Praat (Boersma and Weenink 2020).

After segment intervals were identified, additional annotations were made to mark the onset of voicing and the release of onset closure of the target words. Initially, the onset of voicing was annotated based on the Harmonic-to-Noise Ratio (HNR) for the 0–500 Hz frequency range during the intervals between the target onset and vowel. The HNR values were extracted using PraatSauce (Kirby 2018). The onset of voicing was identified as the first window where the HNR exceeded zero, which served as an indicator of glottal vibration. The annotations of voicing onset were manually reviewed and corrected as needed. Subsequently, the release of onset closure was manually annotated. We did not observe any tokens with pre-voiced onsets where voicing cessation occurred before the release of closure.

2.1.3 Acoustic measurements

In this paper, we focus our analysis on seven acoustic parameters: Voice Onset Time (VOT), spectral center of gravity (CoG) of the release burst, fundamental frequency (f0), the first and second formants (F1 and F2), the corrected difference between the first harmonics and the harmonics at the third formant (H1*-A3*), and Cepstral Peak Prominence (CPP). These measurements were chosen because they have been found to correlate with phonation type and register contrasts in various languages, such as Khmu Rawk (Abramson et al. 2007), Nakhonchum Mon (Abramson et al. 2015), Northern Raglai (Brunelle et al. 2022), Chong (Di Canio 2009), Gujarat, White Hmong (Esposito 2012), Jalapa Mazatec (Garellek and Keating 2011), and Ngãi Giao Chrau (Tạ et al. 2022), among others.

Voice Onset Time (VOT) was determined by calculating the time difference between the annotated closure release and the onset of voicing. Additionally, we computed the Center of Gravity (CoG) of the voiceless stop burst. Following the methodology outlined in Chodroff and Wilson (2014), the recordings were resampled at 16 kHz, pre-emphasized, and high-pass filtered at 200 Hz. For each burst, we generated a smoothed spectrum by averaging the squared amplitudes of five 64-point Fast Fourier Transform (FFT) spectra obtained from 3 ms Hamming windows with a 1 ms time step. The first window was centered at the annotated closure release. If the lag between the annotated closure release and the onset of voicing was shorter than the duration required for five windows, we adjusted the number of spectra to fit within this interval. CoG was then calculated from the smoothed spectra using the formula:

Here, f _i represents the frequency of the ith spectral component, A _i represents the amplitude of the ith spectral component at frequency f _i , and N represents the total number of frequency bins in the spectrum.

In addition to VOT and CoG, acoustic measurements were extracted from the annotated files using PraatSauce (Kirby 2018), a set of Praat scripts for extracting spectral measures modeled after VoiceSauce (Shue et al. 2011). Measurements were made every millisecond over the entire recording using a 25 ms analysis window. Fundamental frequency (f0) was estimated using Praat’s autocorrelation method, with a speaker-specific f0 range determined according to the approach proposed by De Looze (2010). Initially, a broad f0 range of 75–400 Hz was utilized for the first pass. Subsequently, the first (Q1) and third (Q3) quartiles of f0 values for each speaker were calculated. In the second round of f0 tracking, the f0 floor for each participant was set to Q1 × 0.75, while the f0 ceiling was set to Q3 × 1.5.

Formant frequencies (F1, F2, and F3) were estimated using the Burg LPC algorithm with a tenpole filter and a Gaussian-like analysis window. The formant ceiling was set to 5,000 Hz for male speakers and 5,500 Hz for female speakers. We also measure the H1*-A3*, the difference between the amplitude of the first harmonic (H1) and the most prominent harmonic of the third formant (A3), corrected for the effects of vocal tract resonance using the method of Iseli and Alwan (2004). H1*-A3* has been shown to correlate with the open quotient of the glottis (Bickley 1982) and the velocity of vocal fold closure (Stevens 1977). A larger difference between the first harmonics (H1) and the higher harmonics (e.g., A3) suggests a breathier voice quality. Cepstral Peak Prominence (CPP) is a measure of the periodicity of vocal fold vibrations (jitter, shimmer). CPP was calculated according to the method described by Hillenbrand et al. (1994) as implemented in PraatSauce. Higher CPP values have been found to correlate with a more modal voice quality (Garellek 2019; Garellek and Esposito 2023; Seyfarth and Garellek 2018).

Spurious measurements of f0, F1, F2, and F3 were removed if they deviated by three standard deviations from the mean computed for each combination of speaker, vowel, and register. This process resulted in the removal of approximately 1 % of data points. Following this, H1*-A3* measures at time points where f0 and F1/F3 had been excluded were also discarded.

2.1.4 EGG signal

The EGG signals were processed using praatdet (Kirby 2020), a Praat script designed for EGG signal processing. The EGG signals underwent high-pass filtering with a 40 Hz pass frequency and a 20 Hz smoothing cutoff. Additionally, we specified the f0 minimum and maximum thresholds to be the speaker-specific range used for f0 tracking of acoustic signals (see Section 2.1.3). The smoothing window size parameter was set to k = 20 preceding and following each period. The smoothed signals were used to calculate the open quotient (OQ) using the method proposed by Howard (1995), with the amplitude threshold set to 3:7 of the EGG cycle’s peak-to-peak amplitude.

2.1.5 Normalization and smoothing

Our objective was to investigate whether acoustic measures and OQ during vowel intervals differed across registers. To facilitate comparison across speakers, f0, F1, F2, H1*-A3*, CPP, and OQ were converted to z-scores by speaker. Statistical analyses were performed on the z-scored values. However, for illustration purposes, we convert z-scores back to the original scales based on the grand mean and mean of standard deviations of all speakers (grand mean + z-score × mean standard deviation).

We time-warped each trajectory to an equal length of 50 data points using MATLAB’s linear interpolation function interp1(). Since OQ data points were not sampled uniformly, the corresponding time steps for each OQ data point were also included as sample points in the linear interpolation process. The time-normalized trajectories were then further smoothed using MATLAB’s smoothdata() function, employing a Gaussian-weighted average over a window of 15 data points. This smoothing was applied to reduce ‘wiggliness’ caused by inaccurate or noisy measurements, as seen in Figure 2. These time normalization and smoothing steps are crucial for the subsequent functional principal components analysis (FPCA), as discussed in Gubian et al. (2015). For instance, CPP is inherently noisy, and the expected difference across registers pertains to the overall height of the trajectories rather than the noise variation in the trajectories. Smoothing helps prevent overfitting in the FPCA.

Figure 2:

An example of trajectory smoothing for a high register token. Red dashed lines represent the time-normalized trajectories before smoothing, while black lines represent the time-normalized trajectories after smoothing.

2.1.6 Functional PCA

Acoustic and OQ trajectories were analyzed using functional principal components analysis (FPCA) (Ramsay and Silverman 2005) followed by linear mixed-effects modeling (Bates et al. 2015). We elected to use FPCA for three reasons. First, it allows us to model entire trajectories, rather than having to select arbitrary regions of the contour. Second, it produces easily interpretable parameterizations when compared to alternative methods such as GAMMs. Third, it permits the analysis of multidimensional dynamic trajectories, considering the changing shapes of multiple measurements simultaneously, such as vowel formants (F1 and F2) and voice quality measures (CPP, H1*-A3*, and OQ). For an introduction and overview of FPCA in a phonetics context, see Gubian et al. (2015).

FPCA provides a data-driven parameterization of one or more input trajectories. For the case of a single f0 trajectory, this parameterization can be expressed with the following equation:

Here, f0 _i (t) represents the continuous function of f0 trajectories with i as the token index and t as the continuous normalized time variable, μ _f0(t) is the mean f0 trajectory, PCk _f0(t) are the K principal components (PCs) derived from the entire f0 trajectories (k = 1, …, K), for instance, PC1 _f0(t) is the first principle component derived from f0 trajectories, and s _k is the score that modulates PCk differently for each f0 trajectory. Each PC captures an independent mode of variation in the data, and each accompanying score controls the direction and magnitude of that variation. The significance of the scores can then be assessed using a linear mixed-effects regression in the usual manner. FPCA is similar to other functional methods of curve parameterization, such as the discrete cosine transform or growth curve analysis, but has the advantage that the PC curves are optimized for a particular dataset.

A major advantage of FPCA is that it can model multidimensional trajectories by combining tuples of functions like those in Eq. (2) (Happ and Greven 2018). Each dimension (trajectory) is parametrised by its own set of PCs, but the same set of PC scores is shared across dimensions. This means that each PC score jointly influences the shape of multiple dimensions, which is useful to model covariance between dimensions.

In this paper, we consider two multidimensional trajectories, one for vowel quality measures and one for voice quality measures. To analyse differences in vowel quality, we use as input trajectories for FPCA paired formant tracks, F1 _i (t) and F2 _i (t), defined as continuous functions representing F1 and F2, respectively. This can be expressed with the following equations:

The formulas for these paired functions are similar to the function for f0 trajectories, but both functions share the same s _k,i scores. In other words, a single s _k,i modulates both F1 and F2 for each token. In a similar fashion, we use FPCA to analyze voice quality measures by considering our three voice quality measures – CPP, H1*-A3*, and OQ – as three dimensions of the same phenomenon.

FPCA was applied to 2,004 tokens with voiceless stop onsets. Tokens with aspirated stop, preglottalized, and sonorant onsets were excluded from the analysis due to an imbalance in token numbers. The mean trajectories for each measurement across all onset categories, presented in Appendix B, suggest that registers are realized similarly across these categories, despite some differences in the magnitude of the effects.

2.1.7 Statistical modeling of FPCA scores

Once all trajectories of all measurements were fitted with FPCA, we obtained the PC-scores s ₁ and s ₂, corresponding to the scores of PC1 and PC2, respectively. These PC-scores capture variations in the curves, irrespective of the register. We focused on the first two PCs because the PC scores of these higher components did not show any relationship with register distinctions. To investigate whether the scores differ across registers, the PC-scores s ₁ and s ₂ were modeled using linear mixed-effect regressions with the lmerTest package in R (Kuznetsova et al. 2017).

In these models, s ₁ and s ₂ for each set of measurements were included as dependent variables. The independent variable was register (high vs. low) for the models of PC-scores of f0 and voice quality. For the vowel quality models, vowel and its interaction with register were also included as independent variables. Subject was included as a random intercept and slope for register. The R model syntax is given in Eqs. (4a)–(4c).

In Section 2.2, we present plots of the reconstructed trajectories based on model predictions with 95 % confidence intervals to provide an overview and demonstrate the significance of the estimates. These reconstructed trajectories were derived from the estimated marginal means of PC-scores, computed through post-hoc tests of the models showing significance using the emmeans package in R (Lenth 2022). In addition to model predictions, we reported both marginal and conditional R ² values for each model, calculated using r.squaredGLMM from the MuMIn package (Bartoń 2024). Marginal R ² represents the variance explained solely by the model’s fixed effects, while conditional R ² represents the variance explained by the entire model, including both fixed and random effects.

The full model output is available in Appendix C.

2.2.1 Onset voicing

Before delving into the analyses of register effects on the vowels, we first checked whether registers were also associated with differences in onset voicing. Descriptive statistics for Voice Onset Time (VOT) are presented in Table 3 and illustrated as histograms in Figure 3. Within the categories of voiceless stops and aspirated stops, no observable effect of register on VOT was evident. Notably, in Figure 3, the VOT of aspirated stops in the high register appears slightly longer than in the low register. However, this difference is attributable to the presence of /k^h/ in the high register and its absence in the low register. As shown in Table 2, /k^h/ generally exhibits a longer VOT compared to /t^h/. The expected difference in voicing across onset categories was observed, with voiceless stops exhibiting short positive voice lags, aspirated stops showing long positive voice lags, and voiced stops demonstrating voice leads. We identified just 19 tokens (out of 2,004) of voiceless stops exhibiting voice leads. It is worth noting that 13 of these tokens belonged to the high register, while the remaining 6 tokens belonged to the low register.

Figure 3:

VOT distributions.

To investigate whether registers were associated with differences in onset voicing, we also examined the Center of Gravity (CoG) of burst, which has been linked to onset voicing by Chodroff and Wilson (2014). Descriptive statistics for CoG are presented in Table 4 and depicted as a boxplot in Figure 4. We focus on tokens with voiceless stops /t, k/ onset followed by the vowel /a/, which did not exhibit voice lead (total of 799 tokens). Within place of articulation, we did not observe any effect of register on CoG. The only difference in CoG observed was across place of articulation, with velar stop onset displaying a lower CoG than alveolar stop onset.

Figure 4:

CoG of burst for different levels of register.

In sum, the analysis of VOT and CoG distributions indicates that registers are not associated with acoustic differences in the onset voicing contrast.

2.2.2 Vowels: f0

Figure 5 illustrates the shape variation of f0 captured by PC1 and PC2. Each panel represents the effect of each principal component. The dark solid line depicts the mean f0 trajectories, μ _f0(t), with PC-scores s _k equal to zero. The color-coded curves represent the curves reconstructed from different s _k based on Eq. (2). Blue curves indicate the trajectories reconstructed from positive PC-scores, while red curves indicate those reconstructed from negative PC-scores. The s _k values used in Figure 5 are equally spaced, ranging between −1 (red) and +1 (blue) standard deviations (σ _k ).

Figure 5:

First two PCs of normalized f0 during target vowel interval pooled across speakers.

The first two principal components explained 83.9 % and 13.9 % of the variance in all f0 trajectories, respectively. PC1 captures the height of f0 trajectories. A positive s ₁ corresponds to a lower f0 than the average f0 trajectory, whereas a negative s ₁ corresponds to a higher f0. On the other hand, PC2 captures the shape of the f0 contour, whether it is falling or rising. When s ₂ is positive, the f0 trajectory becomes rising, while a negative s ₂ reflects falling f0 contours.

Figure 6 illustrates the s ₁ and s ₂ scores for high and low registers. The figure shows that the high register exhibits lower s ₁ values compared to the low register, while the difference in s ₂ is not as pronounced.

Figure 6:

Violin plots of the s ₁ and s ₂ PC-scores of f0 by registers. Red dots represent the mean of each category.

This observation is also supported by the results of the linear mixed-effect regressions. Table 5 presents the predicted s ₁ and s ₂ scores based on the register. The effect of register on s ₁ is significant: s ₁ is higher for the low register, as indicated by the positive estimated effect size (β ₁), which is the difference in s ₁ between high and low register. This finding indicates that the high register has an overall higher f0 contour than the low register.

Although the effect of register on s ₂ is also significant, the marginal R ² is very low, with the fixed effect of register explaining only 3 % of the variance in s ₂. In contrast, the conditional R ² is much higher, indicating that the entire model, including both fixed and random effects, explains 50 % of the variance. The substantial difference between marginal and conditional R ² suggests that the variation in s ₂ is primarily captured by the random effect, which is the speaker. In other words, s ₂ represents the variation across speakers rather than across registers. This suggests that the rising or falling contours of f0 are specific to individual speakers and not integral to the realization of registers.

The reconstructed f0 trajectories from the estimated marginal means of s ₁ are shown in Figure 7. The high register consistently exhibits a higher f0 than the low register throughout the entire trajectory.

Figure 7:

Reconstructed normalized f0 trajectories from the estimated marginal means of s ₁ for different registers.

2.2.3 Vowels: voice quality measures

Figure 8 illustrates the shape variation of the three voice quality measures, CPP, H1*-A3*, and OQ. The first two principal components explained 41.8 % and 25.7 % of the variance in all voice quality measure trajectories, respectively. For PC1 (Figure 8, left panel), an increase in s ₁ corresponds to a decrease in CPP and an increase in H1*-A3* and OQ. Conversely, a decrease in s ₁ corresponds to an increase in CPP and a decrease in H1*-A3* and OQ. In other words, s ₁ captures breathiness: high s ₁, indicating low CPP, high H1*-A3*, and high OQ, signifies breathy voice quality.

Figure 8:

First two PCs of normalized CPP (top), normalized H1*-A3* (middle), and normalized OQ (bottom) during target vowel interval pooled across speakers.

On the other hand, PC2 (Figure 8, right panel) represents variation in CPP and OQ. An increase in s ₂ indicates a decrease in CPP and OQ, and vice versa. We can interpret PC2 as capturing variation in how voice quality is realized across different measurements. For example, in some tokens, breathy voice quality might be indicated by low CPP, while OQ does not necessarily show a high value. PC2 does not seem to account for the variation in H1*-A3*, as changes in s ₂ do not result in large variations of H1*-A3*.

Figure 9 shows the s ₁ and s ₂ scores for high and low registers. The figure indicates that the high register has lower s ₁ values compared to the low register. Although s ₂ is also lower for the high register, the difference is less pronounced. This observation is supported by the results of the linear mixed-effect regressions.

Figure 9:

Violin plots of the s ₁ and s ₂ PC-scores of voice quality measures by registers. Red dots represent the mean of each category.

Table 6 presents the predicted s ₁ and s ₂ scores based on register. The effect of register on s ₁ is significant: s ₁ is larger for the low register, as indicated by the positive estimated effect size (β ₁). This finding suggests that the low register has a breathier voice quality than the high register.

The effect of register on s ₂ is also significant, but the marginal R ² is very low, with the fixed effect of register explaining only 5 % of the variance in s ₂. The conditional R ² is slightly higher, indicating that the entire model, including both fixed and random effects, explains 14 % of the variance. This indicates that s ₂ represents variation across speakers rather than registers. Suggesting that voice quality is realized differently by different speakers. However, the effect is rather small, since the model still explains only 14 % of the variance.

The reconstructed voice quality measures from the estimated marginal means of s ₁ are shown in Figure 10. High register tokens consistently exhibit higher CPP, and lower H1*-A3* and OQ than low register tokens. In other words, the high register shows more modal voice quality, while the low register shows breathy voice quality.

Figure 10:

Reconstructed normalized CPP (top), normalized H1*-A3* (middle), and normalized OQ (bottom) trajectories from the estimated marginal means of s ₁ for different registers.

2.2.4 Vowels: formant frequencies

Figure 11 illustrates the shape variation of vowel quality measures, F1 and F2, captured by PC1 and PC2. The first two principal components explained 71.3 % and 17.4 % of the variance in F1 and F2 trajectories, respectively. For PC1 (Figure 11, left panel), an increase or decrease in s ₁ corresponds to a decrease and increase in F1 and F2 together. An increase in s ₁ indicates the lowering of both F1 and F2 (blue trajectories). In other words, s ₁ captures vowel height and frontness: a high s ₁, indicating low F1 and low F2, signifies a higher and more back vowel.

Figure 11:

First two PCs of normalized F2 (top) and normalized F1 (bottom) during target vowel interval pooled across speakers.

On the other hand, for PC2 (Figure 11, right panel), the increase and decrease cause F1 and F2 to come closer together (blue trajectories) or shift further apart (red trajectories). Therefore, the modulation of s ₂ also captures vowel height and frontness again: a high s ₂, indicating high F1 and low F2, signifies a higher and more front vowel.

Figure 12 shows the s ₁ and s ₂ scores for high and low registers. The figure indicates that only a few vowels exhibit prominent differences in PC-scores. Furthermore, the vowels that do show differences do not consistently exhibit the same direction of the effect. This observation is supported by the results of the linear mixed-effect regressions.

Figure 12:

Violin plots of the s ₁ (top) and s ₂ PC-scores (bottom) of vowel quality measures by registers separated by vowels. Red dots represent the mean of each category. o- and ɔ- represents /o/ and /ɔ/ in open syllables respectively.

The full model results are summarized in Appendix C. All models have the vowel /a/ with high register as the baseline. The model predicting s ₁ explains 88.1 % of the variance with only fixed effects and 88.4 % with both fixed and random effects. The effect of register on s ₁ is not significant (t(33.0) = −0.55, p = 0.59), suggesting that there is no difference in vowel formants across registers for the vowel /a/. We observed significant effects of the interaction between register and the vowels /o/ in closed syllables (t(1,888.3) = −4.91, p < 0.001), /ɔ/ in closed syllables (t(1,888.3) = 7.38, p < 0.001), and /ɑ/ (t(1,888.4) = 2.72, p = 0.007). This indicates that only the vowels /o/ in closed syllables, /ɔ/ in closed syllables, and /ɑ/ show differences in s ₁ across registers. However, the direction of the effect varies for different vowels. Specifically, /o/ in closed syllables shows a negative effect of register, indicating that the low register has lower s ₁, i.e., higher formant frequencies, which correspond to lower and more front vowels. In contrast, /ɑ/ and /ɔ/ in closed syllables show a positive effect of register, indicating that the low register has higher s ₁, i.e., lower formant frequencies, which correspond to higher and more back vowels.

On the other hand, the model predicting s ₂ explains 19.9 % of the variance with only fixed effects and 20.6 % with both fixed and random effects. The effect of register on s ₂ is significant (t(60.6) = 4.16, p < 0.001), suggesting that there is a difference in vowel formants across registers for the vowel /a/. We also observed significant interactions between register and the vowels /ɔ/ in closed syllables (t(1,890.3) = −5.63, p < 0.001) and also in open syllables (t(1,893.4) = −2.23, p = 0.026), and /ɑ/ (t(1,890.3) = −2.17, p = 0.03). To confirm whether the effect of register is significant for /ɔ/ and /ɑ/, we extracted the estimated marginal means using the emmeans package (Lenth 2022) in R. The marginal means of s ₂ with 95 % confidence intervals for each register are as follows: /ɔ/ in closed syllables High register: 4.4 [CI _95 % 3.6 5.1]; Low register: 2.7 [2.4 3.1]; /ɑ/ High register: −0.3 [−0.5 −0.04]; Low register: 0.7 [0.3 1.1]. The confidence intervals for high and low registers do not overlap for these vowels and the vowel /a/, while they do for all other vowels. This indicates that only /a/, /ɔ/, and /ɑ/ show significant differences in s ₂ across registers. However, the effect is rather small, since the model still explains only 19 % of the variance.

The reconstructed formant frequency trajectories from the estimated marginal means of s ₁ are shown in Figure 13. Although s ₂ also shows differences for certain vowels, the effect size is small, and the model explains only a small proportion of variance (R ²), so we do not consider it further in our interpretation. As described in the analyses above, only three vowels exhibit differences in formant frequency across registers when considering s ₁. However, the direction of the effect is inconsistent among these vowels. Based on the typology of phonetic properties of register, only /ɔ/ in closed syllables and /ɑ/ show the expected direction of the effect (higher vowels in the low register). This suggests that vowel quality is likely not a distinguishing property of registers. The difference in vowel quality across register that we observed might be due to independent factors.

Figure 13:

Reconstructed normalized F2 (top) and normalized F1 (bottom) trajectories from the estimated marginal means of s ₁ for different registers. o- and ɔ- represents /o/ and /ɔ/ in open syllables respectively.

2.2.5 Relation between f0 and voice quality measures

In the previous sections, we observed consistent differences in f0 and voice quality across registers. An interesting question that arises is whether f0 and voice quality are related or if they vary independently. In this section, we apply multidimensional FPCA to both f0 and voice quality measures to investigate whether variations in one correlate with variations in the other.

Figure 14 illustrates the shape variation in f0 and voice quality measures as captured by PC1 and PC2. The first two principal components explained 47.3 % and 20.5 % of the variance in the combined f0 and voice quality trajectories, respectively. For PC1 (Figure 14, left panel), an increase in s ₁ corresponds to a decrease in f0 and CPP, along with an increase in H1*-A3* and OQ. Conversely, a decrease in s ₁ is associated with an increase in f0 and CPP, and a decrease in H1*-A3* and OQ. These directional variations align with the expected register contrasts: high register shows high f0, high CPP, low H1*-A3*, and low OQ, while low register shows the opposite pattern. Additionally, the standard deviations of s ₁ are comparable across all measurements, suggesting that variations in s ₁ induce changes in f0 and voice quality measures of similar magnitude.

Figure 14:

First two PCs of normalized f0, normalized CPP, normalized H1*-A3*, and normalized OQ during target vowel interval pooled across speakers.

In contrast, PC2 (Figure 14, right panel) primarily captures the simultaneous increase and decrease in CPP and OQ, similar to what was observed in the PC2 analysis of voice quality measures alone. An increase in s ₂ corresponds to a decrease in CPP and OQ, and vice versa. However, PC2 does not seem to account for variations in f0 and H1*-A3*, as changes in s ₂ do not lead to significant variations in these measures.

Table 7 presents the predicted s ₁ and s ₂ scores based on register. The effect of register on s ₁ is significant, with s ₁ being larger for the low register, as reflected by the positive estimated effect size (β ₁). This result suggests that the low register is associated with both a lower f0 and a breathier voice quality compared to the high register.

In contrast, the effect of register on s ₂ is not significant, and the marginal R ² is quite low, indicating that the fixed effect of register accounts for only 2 % of the variance in s ₂. The conditional R ² is slightly higher, suggesting that the full model, which includes both fixed and random effects, explains 13 % of the variance.

The results indicate a significant and comparable variation in both f0 and voice quality measures, as captured by s ₁, which supports the notion that these two dimensions are correlated. Changes in f0 are consistently associated with corresponding changes in voice quality across tokens. The significant effect of register on s ₁ further supports this finding, indicating that the relationship between f0 and voice quality is not only present but also meaningful in distinguishing between high and low registers. In other words, f0 and voice quality vary together in the production of register distinction.

2.2.6 Inter-speaker variation in production

We also explored potential individual differences in the acoustic realization of register. Figure 15 shows the predicted s ₁ values for f0, voice quality, and vowel quality across high and low registers for each speaker from the linear mixed effect regression models discussed in Section 2.2.2, Section 2.2.3, and Section 2.2.4 above.

Figure 15:

Predicted s ₁ values for f0, voice quality, and vowel quality across high and low registers for each speaker.

For f0, the variation in s ₁ across registers, which reflects overall f0 height, is pronounced and consistent across all speakers, with similar magnitudes. All speakers exhibit lower s ₁ values for the high register and higher s ₁ values for the low register, indicating that f0 is consistently higher for the high register and lower for the low register for all speakers.

In contrast, the difference in voice quality across speakers is less uniform. While most speakers display a clear difference in s ₁, with lower s ₁ values for the high register and higher s ₁ values for the low register – suggesting that the high register is more modal and the low register is breathier – there are exceptions. For example, speaker F01 does not follow the expected pattern, and speaker F03 exhibits only a slight difference between registers. This variation suggests that voice quality may not serve as a reliable acoustic correlate of register for these speakers.

Regarding vowel quality, Figure 15 shows the predicted s ₁ values for vowel /a/ (n = 761), which has the largest number of tokens, and for vowel /ɔ/ in closed syllables (n = 336), which was observed to be higher for the low register, as expected based on the typology outlined in Table 1. The figure indicates no systematic difference in vowel quality for vowel /a/. However, for vowel /ɔ/, all speakers show lower s ₁ values for the high register and higher s ₁ values for the low register, indicating that the vowel is consistently lower and more fronted for the high register, and higher and more backed for the low register, across all speakers.

One question regarding inter-speaker variation in register production that has been raised in previous literature is the relative weight of each register property. To address this, we also explore the relative contributions of f0, voice quality, and vowel quality to distinguish high and low registers. It is important to note that comparing the relative weight between f0 and voice quality may be problematic, given the strong correlation between these two properties, as discussed in Section 2.2.5. A more detailed discussion of this issue can be found in Section 4.3.

To compare the relative weight of each register property, we estimated the magnitude of the difference between high and low registers for each speaker using Cohen’s d (Cohen 1988). Cohen’s d was chosen because it normalizes effect sizes across different units of measurement, facilitating direct comparison. We first extracted the PC-score s ₁ for the f0, voice quality, and vowel quality trajectories. PC-score s ₁ was chosen over a raw data point from the trajectories because s ₁ represents the entire contour, arguably providing a more accurate description of the entire trajectory rather than focusing on an arbitrary part of the contour. Following the approach of Brunelle et al. (2020), we then calculated the vowel-weighted mean difference between the two registers for s ₁ of each trajectory, divided by the pooled mean standard deviation weighted by vowel and by register. While this measure is straightforward to compute, it is crucial to interpret the results cautiously, as potential correlations, particularly between f0 and voice quality, are not accounted for in this approach.

Figure 16 displays the Cohen’s d scores for each production property, with all scores multiplied by −1 so that positive values represent the effect size in the expected direction of register differences. Scores below zero indicate that the direction of the effect for a given measurement is opposite to what is expected for register properties. Following Cohen (1988), scores lower than 0.8 are considered not meaningful.

Figure 16:

Individual variation in the use of acoustic properties distinguishing high and low registers, presented as Cohen’s d scores calculated from PC-score s ₁. Cohen’s d values have been multiplied by −1. Shaded area represents the area where Cohen’s d is less than 0.8. Participants are arranged based on their gender and year of birth.

The effect size of f0 is large for all speakers, indicating that f0 is one of the main properties for distinguishing register. Among the 17 speakers, 14 exhibit f0 as the most reliable acoustic correlate of register. However, three speakers (M01, M09, and M11) show the largest effect size for voice quality measures, with M10 also displaying a similar effect size between f0 and voice quality. Interestingly, M09, M10, and M11 – the speakers with large effect sizes for voice quality – are the oldest participants, whereas M01, who also exhibits a strong effect for voice quality, is among the youngest. This variation makes it difficult to determine whether age influences the production of voice quality. Several other speakers also show voice quality as a secondary property, as indicated by the second-largest effect sizes. On the other hand, some speakers, such as F01, F02, and F03, display relatively small effect sizes for voice quality measures, with Cohen’s d values below 0.8, suggesting that voice quality may not be a significant distinguishing feature of registers for these individuals. Notably, all male participants exhibit meaningful effect sizes for voice quality, whereas only three female participants show similar patterns. We did not observe a systematic trade-off observed between f0 and voice quality measures across speakers.

Vowel quality did not emerge as a robust correlate of register for most speakers, with none except M10 and M04 exhibiting Cohen’s d values greater than 0.8. M10 and M04 alone shows a relatively small difference in vowel quality between registers. No clear patterns related to gender or geographical origin were observed in the results.

3.1.1 Participants, stimuli, and data collection

The perception experiment involved 29 listeners (10 females and 19 males). Among these participants, twelve had also taken part in the production experiment. The birth years of the participants ranged from 1971 to 2002 (21–52 years old at the time of the experiment). They were from various townships in Mon state: 10 participants were from Paung township, 7 from Thanbyuzayat township, 3 from Mudon township, 3 from Ye township, 2 from Kyaikmaraw township, 2 from Mawlamyine township, 1 from Chaungzon township, and 1 from Thaton township (see Figure 1). All participants were native speakers of Mon, and reported Mon to be their primary language. They also demonstrated fluency in Burmese and possessed basic proficiency in Thai.

Following the methodology of Brunelle et al. (2020), participants completed a forced-choice identification task, which involved listening to synthesized stimuli varying in fundamental frequency (f0), first formant (F1), and phonation type. They were then required to identify the stimuli by pressing one of two keyboard buttons corresponding to images displayed on a laptop screen. The stimuli comprised two sets. In one set, participants had to choose between /tɑk/ ‘building’ and /tɑ̤k/ ‘poor’ (henceforth /ɑ/-stimuli), while in the other set, the choices were /tɔ/ ‘shaft, rod, handle’ and /tɔ̤/ ‘close; cover, hide’ (henceforth /ɔ/-stimuli). Responses from 11 participants to the second set of stimuli (/tɔ/ vs. /tɔ̤/) were omitted from analysis due to ambiguity regarding the image corresponding to /tɔ/. Subsequently, responses from 29 participants were analyzed for /ɑ/-stimuli, while responses from 18 participants were analyzed for /ɔ/-stimuli.

Stimuli were synthesized using KlattGrid synthesis in Praat version 6.3.10. The script to synthesize the stimuli was based on Brunelle et al. (2020). As pitch, voice quality, and vowel quality have all been previously implicated in acoustic descriptions of the Mon register contrast, we manipulated three associated Klatt parameters: fundamental frequency (f0), open quotient (OQ), and the first formant frequency (F1). To establish the range for f0 and OQ, maximum and minimum values were derived from measurements obtained from target tokens produced in isolation by a male Mon speaker during a pilot study conducted in 2019. Since F1 did not covary with register in the speech of the model speaker, F1 ranges were based on the values reported in Abramson et al. (2015) and Behr (2018). These parameters were then combined to create all possible combinations of acoustic properties. The total number of stimuli was 45 tokens per word pair. The three parameters were manipulated as follows:

1. f0: Pitch targets were set at three points within the vowel intervals to replicate the falling pitch pattern observed in words in isolation: the onset of the vowel (f0 onset), 50 ms into the vowel (f0 at steady state), and 90 ms before the end of the vowel (f0 pre-offset). Five sets of pitch targets were defined to generate five distinct pitch contours, with each set consisting of three values representing f0 onset, f0 at steady state, and f0 pre-offset, respectively: (i) 200, 170, 130 Hz; (ii) 180, 160, 125 Hz; (iii) 160, 150, 120 Hz; (iv) 140, 140, 115 Hz; and (v) 120, 130, 110 Hz. A fixed target of 100 Hz was set at the end of the vowel.

2. Open Quotient: OQ was set to be stable throughout the vowel intervals, with three possible values: 0.3, 0.4, 0.5.

3. F1: The two vowel stimuli were configured with distinct F1 targets. For the /ɑ/-stimuli, three possible values were set during the steady state period, from 20 ms into the vowel to 90 ms before the end of the vowel: 900, 800, and 700 Hz. Fixed targets of 600 Hz and 350 Hz were set at the onset and end of the vowel, respectively, to model the F1 transition from consonant onset and from coda. For the /ɔ/-stimuli, as native speakers tend to exhibit a slight off-glide in the vowel (see Figure 26 in Appendix B), F1 targets were adjusted differently at 20 ms into the vowel and 90 ms before the end of the vowel to replicate F1 contours. Three sets of potential values were established: (i) 800, 600 Hz, (ii) 650, 525 Hz, and (iii) 500, 425 Hz. A fixed target of 600 Hz was set at the beginning of the vowel to represent the F1 transition from consonant onset.

The resulting stimuli were perceived as natural by the participants, with none reporting awareness that the stimuli had not been recorded from real speakers. Anecdotally, some participants reported that certain tokens sounded as if they were produced by a non-native speaker of Mon. Figure 17 illustrates the acoustic measurements of the resulting stimuli extracted using PraatSauce. Note that due to interactions among the parameters, the acoustic properties of the stimuli were not consistently precisely on target. In particular, the manipulation of OQ did not influence the CPP of the stimuli. However, given that the manipulation produced H1*-A3* values that were in line with the production data, combined with the fact that acoustic differences in H1*-A3* were much greater than those of CPP in the productions of the pilot speaker as well as in the production experiment, we judged the manipulations adequate.

Figure 17:

Means of the key acoustic parameters of the stimuli in the identification experiment. Top panel: /ɑ/-stimuli. Bottom panel: /ɔ/-stimuli. The ribbons correspond to one standard error above and below the mean.

The experiment was conducted using OpenSesame (Mathôt et al. 2012). Participants received instructions through visual aids and verbal explanations provided by a native Mon speaker. Before the testing phase, participants first completed a training phase using the two tokens with the most extreme parameter values, representing the expected natural production of the two registers (high register: highest f0, lowest OQ, highest F1; low register: lowest f0, highest OQ, lowest F1). Each stimulus was repeated three times. During training, participants listened to each stimulus while the corresponding image was highlighted as the audio played. Next, participants completed a first practice phase using the same two tokens as in the training phase, where they were asked to identify the stimuli by pressing the appropriate response buttons (5 repetitions per stimulus). During this phase, eight participants (F08, F09, F12, M03, M06, M09, M15, M16) had accuracy rates below 60 %. Finally, in a second practice phase, participants were presented with 10 randomly selected stimuli. These 10 random stimuli were drawn from the same pool as those used in the testing phase. Participants were encouraged to take short breaks between blocks.

In the testing phase, stimuli were presented in six blocks, with three blocks for each word pair alternating between /ɑ/- and /ɔ/-stimuli. Each block contained all 45 randomized stimuli for the relevant word pair, resulting in a total of 240 tokens per listener. The entire experiment took 15–20 minutes per participant. Responses with reaction times exceeding 2 seconds were not recorded.

3.1.2 Statistical analyses

Responses were analyzed using mixed-effects logistic regressions implemented in the lmerTest package (Kuznetsova et al. 2017) in R. Separate models were fitted for each word pair. The dependent variable was participant response, indicating whether they chose the high or low register stimulus. Fixed effects included the f0, OQ, and F1 values of the synthesized stimuli, which were centered and treated as discrete numerical variables. For example, the five levels of f0 were coded as [−2, −1, 0, 1, 2], ordered from lowest to highest f0 of the stimuli. Two-way interactions between the fixed-effect variables were also included in the maximal models. Subjects were included as random intercepts with random slopes for all non-interaction fixed effects. The maximal model can be expressed with R syntax as follows:

We began by fitting the maximal model, which included all possible interactions and fixed effects. The model was then simplified through a step-wise process, sequentially removing the interaction or fixed effect with the lowest F-value. Interactions were removed first, followed by fixed effects. The final, simplified models were those that did not result in a significant increase in the Akaike Information Criterion (AIC) score when compared to the more complex models. Throughout the process, the random effect structure was kept unchanged.

3.2.1 Identification experiment

The proportion of high register responses obtained from listeners is illustrated in Figures 18 and 19. Summaries of the final logistic regression models are provided in Table 8 for /ɑ/-stimuli and Table 9 for /ɔ/-stimuli.

Figure 18:

Proportion of high register responses for each type of stimulus, by F1, f0, and open quotient (OQ), averaged over all listeners. Top: /ɑ/-stimuli. Bottom: /ɔ/-stimuli.

Figure 19:

Heat plots of high register responses across each combination of f0, F1, and OQ. Each cell represents the proportion of high register responses for each stimulus. Darker shades represent higher proportion of high register responses. Top: /ɑ/-stimuli, bottom: /ɔ/-stimuli. Left: OQ by f0, middle: F1 by f0, right: F1 by OQ.

For /ɑ/-stimuli, the final model includes fixed effects of f0, OQ, and F1, but no interaction terms. Among these factors, f0 demonstrates the largest effect size, indicating its strongest weight on the perceptual responses. The effect sizes of OQ and F1 are comparable, suggesting equally robust effects. Specifically, positive correlations are observed between f0 and F1 and high register responses, indicating that higher f0 and F1 values correspond to a higher proportion of high register responses, as expected based on the typology and the production results. Conversely, OQ exhibits a negative correlation with high register responses, implying that higher OQ values (indicating breathier phonation) correspond to a lower proportion of high register responses.

For /ɔ/-stimuli, the final model includes fixed effects of f0 and OQ only, but not an effect of F1 or any interaction terms. Similar to /ɑ/-stimuli, f0 demonstrates the largest effect size and is positively correlated with high register responses. Similarly, OQ exhibits a negative correlation with high register responses, consistent with the patterns observed in /ɑ/-stimuli.The lack of a significant F1 effect in the perception of /ɔ/-stimuli is consistent with the observed production behavior of /ɔ/ in open syllables.

The analysis of conflicting cues in Figures 18 and 19 demonstrates how listeners navigate ambiguity when acoustic cues suggest different register categories. For the /ɑ/-stimuli, when f0 is 170 Hz (indicating high register) or 130 Hz (indicating low register), listeners tended to respond accordingly, regardless of whether the F1 or OQ values conflict. When f0 is ambiguous, however, small effects of both F1 and OQ can be observed. For the /ɔ/-stimuli, responses to stimuli with ambiguous f0 values (140–150 Hz) tended to be around chance, although here again the likelihood of a high response decreased when OQ was indicative of low register.

These findings suggest that when faced with conflicting cues, listeners tend to make preferential use of the f0 cue. However, there is also evidence that they attend to voice and vowel quality cues, especially if f0 is indeterminate.

3.2.2 Inter-speaker variation in perception

In line with the production experiments, we also investigated potential individual differences in perception. Initially, logistic regressions were fitted for each listener, with fixed effects of f0, OQ, and F1 on the two sets of perception stimuli. Due to the limited number of tokens tested with each participant, no interactions or random effects were included. We interpret the coefficients of the logistic regression as the perceptual weights of each perception parameter, following previous works (e.g., Brunelle et al. 2020; Schertz et al. 2015).

The coefficients for each participant are illustrated in Figure 20 for each perception parameter. Coefficients for OQ have been multiplied by −1, so that positive values indicate effect sizes in the expected direction, and values below zero suggest that the direction of the effect is unexpected. A perceptual weight greater than zero indicates that the participant used the parameter to distinguish between the two registers.

Figure 20:

Individual variation in the use of perceptual properties to distinguish between high and low registers, presented through the coefficients from logistic regression models. Coefficients associated with OQ have been multiplied by −1 for clarity. Participants are arranged based on their gender and year of birth.

As shown in Figure 20, some listeners were unable to differentiate between the two register categories using the manipulated parameters (F09, F11, F12, M01, M06, M18, M19). This is evident from the clustering of perceptual weights around zero for all parameters. This could mean that their perceptual strategies for distinguishing registers may not rely on the manipulated parameters, but could also indicate that our manipulation was in some respects unsuccessful. It is worth noting that, among these seven participants, only three participants (F09, F12, M06) had accuracy rates below 60 % in the practice phase.

For the majority of participants (F03, F04, F08, M02, M03, M05, M09, M12, M13, M15, M22), f0 was the cue with the highest perceptual weight. Notably, one listener (M04) consistently relied on F1 differences across both sets of stimuli to distinguish the two registers.

Voice quality, at least as implemented in terms of OQ differences, was not found to be the most strongly weighted parameter for any of our listeners. While both F1 and OQ typically have non-zero weights, the precise weighting varies considerably across speakers. Some participants exhibited a stronger perceptual weight for F1 over OQ, while others demonstrated the opposite pattern. We did not find any correlation between perceptual weights and participants’ gender or year of birth.

Heat plots for each listener can be found in Appendix D.