1.1. Korean oral stops: The role of voice quality

The laryngeal contrast in Korean has been extensively studied in the past few decades (for a review, see Lee, Holliday, & Kong, 2020, and references therein). The three oral stop series—commonly labeled as ‘aspirated,’ ‘fortis,’ ‘lenis’—shed light on typological patterns on both synchronic and diachronic levels. First, these stops are all voiceless in phrase-initial position, which is typologically uncommon among the languages with a three-way laryngeal contrast. Second, in Seoul Korean, phrase-initially, the VOT (Voice Onset Time) difference between two of the three stop series has decreased in the course of the last century, while the following f₀ (fundamental frequency) difference has increased (Bang, Sonderegger, Kang, Clayards, & Yoon, 2018; Kirby, 2013; Silva, 2006, to cite just a few). This phonologization of f₀ has been argued to be incorporated into the Korean intonational phonology (Jun, 1993, 1998) (see discussions in Choi, Kim, & Cho, 2020).

To a lesser extent, the voice quality differences have also been found to contribute to the laryngeal contrast in Korean. Cho, Jun, and Ladefoged (2002) concluded that the three stop series can be “differentiated from each other by the voice quality of the following vowel.” However, results of spectral tilt measures from previous literature are variable (for a review, see Lee & Jongman, 2012, Table 7). Restricted to Seoul Korean, it has been consistently found that lenis stops are followed by breathy vowels, and fortis stops by modal or laryngealized vowels (Cho et al., 2002; K.-H. Kang & Guion, 2008; Yu, 2018). On the other hand, aspirated stops do not show clear-cut patterns. H. Ahn (1999) (cited in Cho et al., 2002) and Lee and Jongman (2012) found that H1–H2 (i.e., amplitude difference between the first and second harmonics) was higher when following aspirated than lenis stops, suggesting breathier voice with the former. This is different from Cho et al’s. (2002) findings, showing that H1–H2 values following aspirated stops were higher than fortis stops but lower than lenis ones, which according to the authors suggested modal voice with aspirated stops, breathy voice with lenis ones, and laryngealized voice with fortis ones. K.-H. Kang and Guion (2008) found yet another pattern: H1–H2 differed little between aspirated and lenis stops. Gender and age differences have also been reported: In M.-R. Kim (2014)’s study on 13 speakers in their 20–30s, H1–H2 was higher at vowel midpoint after lenis than aspirated stops only for females. Similarly, in Yu’s (2018) study on 11 older and 17 younger speakers (gender-balanced), H1–H2 and H1-A2 (i.e., amplitude difference between the first harmonic and the second formant) were higher when following lenis than aspirated stops only for young female speakers. Note, however, that this pattern (i.e., lenis breathier than aspirated) was found with older male speakers in Cho et al. (2002).

The aforementioned studies reported spectral tilt measures at fixed time points (generally at the vowel onset or midpoint) or averaged over the entire vowel. While f₀ has been reported to be relatively stable over the entire vowel in Korean, we do not fully understand how voice quality evolves in the course of the vowel. This could potentially explain the contradictory results: It might be the case that the vowel following aspirated stops is breathier due to the aspiration, thus near the onset, but becomes less breathy during the latter part of the vowel. In the current study, we collected electroglottographic (EGG) data for the purpose of illustrating more accurately the time course of both f₀ and voice quality thanks to the high temporal resolution of glottal variation reflected in the EGG signals (see the principle of the EGG method in sections 2.3 and 2.4). Finally, it should also be noted that a broad voice quality category, such as breathy voice, may correspond to different glottal mechanisms in terms of control of intrinsic laryngeal muscles as well as timing, speed, and magnitude of glottal opening/closing, airflow, and air pressure, as demonstrated in electromyographic, articulatory, and aerodynamic studies on Korean oral stops (e.g., Hirose, Lee, & Ushijima, 1974; H. Kim, Maeda, Honda, & Crevier-Buchman, 2018; Lee & Jongman, 2012).

Given the inconclusive results with respect to voice quality, the first question of this study is: How does voice quality on the following vowel differ between aspirated and lenis stops? Is the discrepancy in previous findings due to the time course of the voice quality over the entire vowel?

1.2. Denasalized and oral stops: A four-way contrast

Independently, some attention has been drawn to another interesting sound pattern in several Korean varieties including Seoul and Gyeonggi dialects: denasalization. In these dialects, domain-initial positions have a stronger tendency for denasalization than domain-medial positions: For example, phrase-initial nasal onsets are often produced with weak nasality, and sometimes with no nasality, or even become devoiced (M.-J. Ahn, 2013; M. Chen & Clumeck, 1975; Y. S. Kim, 2011; Yoo, 2015; Yoshida, 2008), and they are likely to be perceived as oral voiced stops by nonnatives (M. Chen & Clumeck, 1975; Y. S. Kim, 2011). In gestural terms, a real-time MRI study has shown that nasals in Korean are subject to greater variability of intergestural timing in syllable-onset than syllable-coda position (Oh, Byrd, Louis, & Narayanan, 2020). Finally, based on a recent apparent-time corpus study, denasalization is very likely to be an ongoing sound change in the process of stabilization (Yoo & Nolan, 2020).

Phrase-initial denasalization has rarely been analyzed together with the three oral stop series (but see a perception study from M.-J. Ahn, 2019, summarized in Section 1.3). In phrase-initial position, the weakening or absence of nasality makes the prototypical form of a nasal stop similar—although not identical to—a prevoiced oral stop (Y. S. Kim, 2011). As a consequence, together with the three voiceless oral stops in this position, speakers of these Korean varieties are most often producing four oral stop series on the surface level. The realizations of the nasal and oral stops are, however, conditioned by prosodic positions (Cho & Keating, 2001) and will be discussed in Section 4.3. The current study is mainly restricted to the phrase-initial position. The four-way contrast makes it necessary to examine the interplay of multiple phonetic properties in the production of the four stop series as a whole. How are the four stop series distinguished one from another? What are concomitant and conflicting properties of each stop series?

First, for the three voiceless oral stops, the positive range of VOT is divided into two spaces, with fortis (short-lag) on one side, lenis and aspirated (long-lag) on the other side. Now, the inclusion of nasal stops means that the negative VOT range is occupied as well. Second, f₀ is lower at the vowel midpoint after lenis and nasal stops than aspirated and fortis stops (Y. Kang, 2014). In Korean intonation, aspirated and fortis stops and affricates trigger a high boundary tone while all the other onsets (including zero onsets, i.e., vowel-only syllables, and sonorant onsets) trigger a low boundary tone (Jun, 1993, 1998). Putting the four stop series altogether, the f₀ space is divided into two, with fortis and aspirated in the high space, and lenis and nasal in the low space. Third, as summarized in Section 1.1, the vowel is breathier following lenis stops than fortis stops, while the voice quality pattern following aspirated stops is inconclusive. The inclusion of nasal stops would potentially add another complexity with its use of voice quality. Nasality and breathy voice are often tightly linked both synchronically and diachronically (Matisoff, 1975), partly motivated by their similar acoustic characteristics, such as the presence of anti-resonances (Ohala, 1975). Garellek, Ritchart, and Kuang (2016) reviewed the interactions between nasality and breathy voice, attributable to the misperception between the two properties (Arai, 2006; Ohala, 1975; Ohala & Busà, 1995) and/or mutual articulatory enhancement (e.g., Stevens & Keyser, 1989). In their study, they also reported that vowels were breathier when following nasal than oral consonants in three Yi (Loloish) languages. A possible scenario in Korean could be that the weakening of nasality is compensated by an articulatory enhancement of breathy voice on the following vowel. As Garellek et al. have explained, in cases where nasality may be involved, given the acoustic similarities between breathy and nasal voice, acoustic signals are less reliable for spectral tilt measures. Therefore, EGG signals have another advantage of not being interfered by supralaryngeal settings including nasality.

We attempt to use EGG data to look into the second question of this study: How would nasality/denasalization interact with breathiness and how is this interaction affected by the breathiness of lenis and aspirated stops?

1.3. Previous perception data

The output of this phonetic investigation would be crucial to understand the mapping of these cues into listeners’ perceptual system. What are the acoustic properties present in the production of the four stop series that could be used as perceptual cues?

In a recent perception study on the four stop series, M.-J. Ahn (2019) examined how incongruent consonantal and vocalic cues affected 16 Seoul listeners’ stop identification. The author concluded a critical role of VOT in the identification of fortis stops, and a major role of f₀ in the distinction between nasal and oral stops as well as between aspirated and lenis stops. A detailed inspection of her data uncovers some asymmetries that are worth noting. First, the role of f₀ is asymmetric in categorizing nasal and fortis stops: When f₀ of the vowel following a nasal stop was artificially raised, the consonant led to 100% identification of a fortis stop, but when f₀ of the vowel following a fortis stop was artificially lowered, the consonant led to 43.2% identification of a nasal stop, but also 37.5% and 19.3% identification of a fortis and a lenis stop, respectively (M.-J. Ahn, 2019, Table 3). This suggests that a high f₀ on the following vowel is sufficient for cueing a fortis stop regardless of the nasal portion (although the nasal properties in the stimuli were not reported) or at least of the prevoiced portion, whereas a low f₀ on the following vowel is necessary but insufficient for cueing a nasal stop. Second, the role of VOT is asymmetric in categorizing the three oral stops: A post-fortis vowel mismatched with a long VOT led to a nearly unanimous identification of an aspirated stop, while a post-aspirated vowel mismatched with a short VOT still led to 51% identification of an aspirated stop, followed by 40.1% identification of a fortis stop. That is, when the following vowel has a high f₀, a long VOT is sufficient but unnecessary for cueing an aspirated stop, while a short VOT is necessary but insufficient for cueing a fortis stop. Furthermore, a post-lenis vowel mismatched with a short VOT led to a predominant identification of a lenis stop (M.-J. Ahn, 2019, Table 4), suggesting that VOT alone is insufficient for cueing a lenis stop.

Other perception studies restricted to oral stops have also shown such perceptual asymmetries. Lee, Politzer-Ahles, and Jongman (2013) varied VOT orthogonally with f₀ of the following vowel (12 VOT steps × 12 f₀ steps). Results were first presented along the two continua, and then broken down by the three most representative steps (min, mid, max) for the two continua (Figures 7, 9, 11), showing that stimuli with the highest f₀ were perceived as a fortis stop in the short-mid VOT range and an aspirated stop in the mid-long VOT range but never as a lenis stop, whereas those with the lowest f₀ were perceived as a lenis stop in the mid VOT range but more likely as an aspirated stop when VOT increased. These results replicated in a more refined way those from M. Kim (2004) with a similar experimental design. Therefore, the asymmetric roles of f₀ and that of VOT were also exhibited with aspirated versus lenis stops: The bias from a high f₀ or a long VOT yielding the identification of an aspirated stop was stronger than from a low f₀ or a short VOT yielding the identification of a lenis stop.

M. Kim and Lee et al.’s studies have shown another interesting result: Stimuli with very short VOTs were unambiguously perceived as fortis stops, regardless of f₀. However, an important detail is that these stimuli were constructed with a non-breathy vowel (i.e., extracted from a fortis-initial syllable). In contrast, M.-J. Ahn (2019) has shown that a short VOT was most often perceived as a lenis stop when it was cross-spliced with a breathy vowel (i.e., extracted from a lenis-initial syllable). These apparently contradictory results suggest that the combination of VOT and f₀ might not be sufficient for categorizing the three oral stops and voice quality is likely to be involved. Indeed, M.-R. Kim, Beddor, and Horrocks (2002) concluded that while f₀ was the most salient cue to lenis stops, a combination of VOT, f₀, and voice quality contributed to the distinction between fortis and aspirated stops. The effect of voice quality has also been studied and confirmed in Francis and Nusbaum (2002), where the perception of five Korean-speaking listeners was affected by multiple acoustic parameters including clarity of formant structure (a voice quality parameter) in natural stimuli. More recently, Schertz, Kang, and Han (2019) conducted a perception study which manipulated orthogonally VOT, f₀, and voice quality in two other Korean dialects than Seoul Korean, and found that voice quality played a role in the fortis–aspirated and more substantially in the fortis–lenis distinction.

To summarize, previous perception data revealed perceptual asymmetries. The biases from both VOT and f₀ are asymmetrical in terms of the direction: (a) a high f₀ is sufficient for the identification of a fortis or aspirated stop, but a low f₀ is necessary but insufficient for the identification of a nasal or lenis stop; (b) a long VOT is sufficient for the identification of an aspirated stop, but a short VOT does not exclude the possibility of identifying any of the four stop series. VOT and f₀ seem insufficient for stop identification, thus other dimensions such as voice quality might come into play.

We thus raise the third question of this study: Can the production pattern explain the asymmetries in perception? First, do we observe asymmetrical distributions of VOT and f₀ in production, mirroring the directional biases in perception? Second, given a short VOT and a vowel with low f₀, the consonant can be identified as a nasal, fortis, or lenis stop, calling for the need of another cue. Do we thus observe the use of voice quality in this VOT–f₀ range in particular? Finally, how about the nasal series? Few studies have investigated the perception of the nasal stop series alone. They will be summarized in Section 4.2 to further discuss the link between production and perception.

1.4. The current study

The main purpose of this study is to re-explore the laryngeal properties by linking together the denasalized stop series with the three oral stop series in Korean, as illustrated by the quadruplet 불 /pul/ (lenis) ‘fire,’ 풀 /p^hul/ (aspirated) ‘grass,’ 뿔 /p’ul/ (fortis) ‘horn,’ and 물 /mul/ (nasal) ‘water.’ We aim to provide data on VOT distributions of the four stops, f₀, and voice quality evolution over the vowel following the four stops. In addition, we will briefly examine the acoustic realizations of nasal onsets. We will focus on phrase-initial position, but discussions will also be made on prosodic and positional variations. Throughout the paper, we still refer to the four stop series as ‘aspirated,’ ‘fortis,’ ‘lenis,’ and ‘nasal’ for simplicity purposes, even though some terms do not accurately describe the phonetic nature of the targeted sounds. (Fortunately enough, nasals are also stops, thus we do not go through further terminological choices.)

3.1. Nasal stops: Variation in nasal weakening

Y.S. Kim (2011) has reported that phrase-initial nasals are often similar, but not identical to word-medial lenis stops that are prevoiced in Korean. While her use of spontaneous speech may have led to frequent medial voicing, closure voicing in medial position is uncommon in our controlled speech (in line with Jun, 1994, as explained in Section 2.2), making it difficult to compare denasalized stops with oral prevoiced stops. For this reason, we will compare phrase-initial with phrase-medial nasal stops, the rationale being that phrase-initial nasals will show lower energy than their phrase-medial counterparts, especially in the high-frequency range, which could be indicated by lower intensity (normalized to vowel), lower Center of Gravity (CoG), greater skewness, and/or greater kurtosis.

Figure 1a shows the intensity differences between nasal onsets and their following vowel by position. (Three outliers of less than –30 dB are not shown for aesthetic purposes.) As expected, the relative intensity of nasals to vowels is overall lower in phrase-initial than medial position. The distribution of the intensity is more dispersed for initial position, due to a large speaker variability: Some speakers have distinct patterns between the two positions, while others show a nearly complete overlap. Figure 1b shows the acoustic duration of nasal onsets by position, including all phonetic variants. The duration is again overall shorter and more variable in phrase-initial than medial position. And again, speaker variability is observed: At least M10 and M11 do not show a clear duration difference between the two positions. Individual plots for intensity and duration are shown in Appendix A.

Figure 1

Violin plot with embedded boxplot by position for nasals (a) C-V intensity difference in dB; (b) Nasal duration.

The CoG data are less clear-cut when pooled across speakers. However, as shown by the individual plot in Figure 2, most speakers show lower CoG for phrase-initial than medial nasals, which could suggest that phrase-initial nasals had less energy in the high-frequency range. The plots for the other three spectral moments are shown in Appendix A. The spectral distribution for phrase-initial nasals compared with medial ones has the following patterns for most speakers: overall less dispersed (i.e., with lower standard deviation), similarly skewed but less flat (i.e., with higher kurtosis), suggesting that low-frequency energy is more prominent.

Figure 2

Violin plot with embedded boxplot by position and speaker for Center of Gravity of nasal onsets.

In addition to speaker variability, variable realizations of phrase-initial nasals may also occur within one speaker. They can be a devoiced stop, a prevoiced stop (sometimes with very short voice lead) with no audible nasality, a prevoiced stop with audible nasality (stronger or weaker), or a plain nasal stop. A prevoiced stop with nasality in our data, which is the most typical realization, is characterized by a strong voice bar and a release burst and relatively low energy in higher-frequency regions, as shown in Figure 3a (speaker F01). The spectral structure looks similar to an illustration of Spanish prevoiced stop with nasal leak (Figure 3d in Solé, 2018). (Note, in passing, that nasal leak is often observed with prevoiced oral stops across languages.) The intensity of the murmur is relatively low but constant, which is different from a typical prenasalized stop with its drop in intensity during the closure signaling the transition from nasal to oral murmur (e.g., Burton, Blumstein, & Stevens, 1992). As a comparison, phrase-medial nasals produced by the same speaker have higher intensity and richer formant structures indicating stronger nasality (Figure 3b). Figure 4 illustrates other realizations of phrase-initial nasals: devoiced, prevoiced with short voice lead, prevoiced with no audible nasality, and plain nasal stop.

Figure 3

From top to bottom: FFT spectrum at nasal midpoint using a 30 ms Gaussian window, waveform, EGG signal, spectrogram of the acoustic waveform. (Amplitude scale: [–1, 1]; dynamic range: 60 dB). (a) Phrase-initial /na/; (b) Phrase-medial /na/.

Figure 4

From top to bottom: FFT spectrum at nasal midpoint in phrase-initial position using a 30 ms Gaussian window, waveform, EGG signal, spectrogram of the acoustic waveform. (Amplitude scale: [–1, 1]; dynamic range: 60 dB). (a) Devoiced (speaker F01); (b) Prevoiced with short voice lead (speaker M06); (c) Prevoiced with no audible nasality (speaker F01); (d) Plain nasal (speaker F02). (Intensity and spectral analyses were excluded for (a) and (b)).

3.2. VOT distributions

Figure 5 shows the distributions of VOT of the four stop onsets. (All distribution plots in this article are stacked rather than overlaid, so as to make overlapped bars more visible. For example, the bars of lenis stops are piled up on top of aspirated stops on the rightmost side and on top of the other three stops in the VOT range between 10 and 50 ms. For aesthetic purposes, one outlier with a VOT at 251 ms is not shown in the VOT plots.) Nasal stops are clearly separated from the other three stop series by their negative VOTs, but 13% (48 out of 361 occurrences) of them have positive VOTs. The distributions are approximately trimodal, with strong overlap between lenis and aspirated stops, and some overlap between fortis and the two other oral stop series.

Figure 5

VOT distributions (stacked) by onset manner.

An LME model was fitted to the raw VOT data, with sum-coded predictors MANNER, PLACE, and VOWEL, and random intercepts for speakers and repetitions, which all improved the fitness of the model. SEX was not included as it did not improve the model. Estimated marginal means for VOTs of each stop series are shown in Table 2. Pairwise comparisons confirm that the contrast between each pair of stops is at p < .0001 level. The summary of the full model and pairwise comparisons is given in Appendix B (Tables B.1 and B.2).

3.3. f₀ curves

Figure 6 shows the averaged f₀ curves in Hz over the vowel following each stop series and zero onset (i.e., vowel-only) by sex group. The separation between fortis/aspirated and lenis/nasal/zero onsets is clear and extended over the entire vowel. Table 3 shows the f₀ values in Hz averaged over all time points after each stop and zero onset. The f₀ difference is about 50 Hz higher after aspirated than lenis stops for females, and about 30 Hz higher for males.

Figure 6

Aggregated f₀ curves in hertz (with standard errors) by onset manner and sex group.

To visualize the individual variability within the f₀ range of each speaker, Figure 7 shows the individual plots of normalized f₀ curves scaled and centered by speaker (mean at 0 and SD at 1 for all normalizations in this study), using the scale_by function of the standardize package (Eager, 2017) in R. The pattern is relatively homogeneous across female speakers, with a clear boundary between the high f₀ zone and the low f₀ zone. While two male speakers follow this pattern, the other two males show a more continuous f₀ (aspirated > fortis > lenis > nasal stops)—their lengths of stay in Japan happened to be the longest (M04: 7 years; M06: 1.5 years).

Figure 7

Normalized f₀ curves (with standard errors) by onset manner and speaker.

To visualize the distinctness of f₀ among stop and zero onsets, Figure 8 shows the distributions of normalized f₀ by sex group, averaged over all time points, as f₀ height is found to be the crucial property, rather than f₀ contour over time. Bimodal distributions are observed both for female and male speakers, with more across-categories overlap for males than females.

An LME model was fitted to the normalized f₀ data averaged over all time points. We excluded the zero onset and focused on the stop series. The predictors, sum-coded, included the maximal factors which improved the model: MANNER, PLACE, and VOWEL. The interactions between MANNER and PLACE and between MANNER and SEX were also added and improved the model. Random intercepts were included for speakers and for repetitions. MANNER|SPEAKER random slope led to singular fit and thus was excluded. Pairwise comparisons (Table 4) showed a large difference of f₀ between aspirated/fortis stops and lenis/nasal stops and a relatively minor difference between other pairs: aspirated > fortis > lenis ≥ nasal (no lenis-nasal difference for female speakers). Summary of the full model is given in Appendix B (Table B.3).

Figure 8

Distributions (stacked) of normalized f₀ averaged over all timepoints by onset manner and sex group.

3.4. Oq curves

Figure 9 shows the averaged Oq curves over the vowel following each stop series and zero onset by sex group. Oq is higher after aspirated than after lenis stops, extending over almost the entire vowel, and it is higher after lenis stops than the other stops and zero onsets, the difference diminishing from vowel onset to offset. Oq is on average higher for female than male speakers. The pattern between fortis, nasal, and zero onsets differs between males and females. However, this difference is more likely due to individual than sex-based variability, as shown in Figure 10. Three speakers out of nine clearly show higher Oq after nasal than fortis stops, contrary to one speaker who clearly shows the reverse pattern, while the remaining five speakers show similar Oq curves after fortis and nasal stops. Oq is more variable after zero onset, but patterns with nasal stops most frequently.

Figure 9

Aggregated Oq curves (with standard errors) by onset manner and sex group.

Figure 10

Normalized Oq curves (with standard errors) by onset manner and speaker.

As pointed out in Section 1.1, previous studies reported inconsistent results with respect to the voice quality difference between lenis and aspirated stops. Our individual data of normalized Oq (scaled and centered by speaker) (Figure 10) show higher Oq for aspirated than lenis stops for most speakers (7 out of 9) mostly over the entire vowel. One speaker (M11) shows the reverse pattern, and one speaker (M06) shows similar Oq curves after the two stop series.

To visualize the distinctness of Oq among stop and zero onsets, Figure 11 shows the distributions of normalized Oq by sex group. Since the Oq difference across all onset types is shown notably at the first part of the vowel, only the first three time points are taken into account, which correspond to the first third of the entire vowel. Unlike the trimodal VOT distributions or the bimodal f₀ distributions, the Oq distributions are strongly overlapped between the onset types.

Figure 11

Distributions (stacked) of normalized Oq averaged over the first three timepoints by onset manner and sex group.

Restricted to the first three time points, an LME model was fitted to the normalized Oq data (time point ≤ 3), excluding the zero onset. The predictors, sum-coded, included the maximal factors which improved the model: MANNER and VOWEL. The interaction between MANNER and VOWEL was also added and improved the model. Random intercepts were included for speakers, and by-speaker random slopes were included for MANNER, which accounted for variable effects of MANNER across speakers. (Random intercepts for repetitions were excluded since the model was overfitted.) Summary of the full model is given in Appendix B (Table B.4). Pairwise comparisons (Table 5) show no difference for the lenis–aspirated pair and the fortis–nasal pair, but a difference between all other pairs. However, based on Figure 10, the same analysis (but including random intercepts for repetitions) without speakers M06 and M11 shows that Oq is higher for aspirated than lenis stops (Table 6).

3.5. Classification tree analysis

The above analyses show that all three properties contribute to some degree to the distinction between certain members of the four stop series. What is, then, the relative importance of each property in classifying each stop series?

A classification and regression tree (CART) analysis was used to determine the predictors of each category, that is, the produced stop series, using the rpart package version 4.1-15 (Therneau, Atkinson, & Ripley, 2019) in R. The analysis explores the predictor used in each binary partition and the cut-off point which maximizes the homogeneity of two split groups, recursively until the maximal homogeneity within each group. This analysis has been used in van Alphen and Smits (2004) for assessing the ranking of acoustic and perceptual cues in the classification of Dutch initial plosives, and in Brunelle (2009) for assessing the ranking of perceptual cues in the classification of Vietnamese tones. The resulting tree gives a visual representation of the relative importance of each predictor in classifying a homogeneous group. However, as warned in Brunelle (2009), it should not be taken that such recursive decision-making process is used by speakers in their production or perception of a phonological category.

In our analysis, the response variable was stop category. The three properties, VOT (in ms), f₀ (in semitone (st), averaged over the vowel), Oq (normalized, of time points 1–3), were included as numerical predictors, and MANNER, PLACE, VOWEL, and SEX as categorical predictors. Complexity parameter (CP, i.e., the minimum improvement of the model at each node) was set to 0.01. No post-pruning was needed as the analysis already had the lowest prediction error rate (7.11%) in cross-validation (CP table and confusion matrix in Appendix C).

Figure 12 shows the classification tree plotted by the rpart.plot package version 3.0.8 (Milborrow, 2019) in R. Only VOT and f₀ are used in the tree construction. VOT first divides the four stop series into two: When VOT is longer than or equal to 30 ms, it is most probably a lenis or an aspirated stop; otherwise, it is most probably a fortis or a nasal stop. Within the VOT ≥ 30 ms group, f₀ further divides the stops into two groups: It is most probably a lenis stop when f₀ < 0.15 st (mean f₀ is at 0.09 st) and an aspirated stop, otherwise. Within the VOT < 30 ms group, when VOT is less than 3 ms, it is certainly a nasal stop; otherwise, f₀ < –0.75 st predicts very well a fortis stop, and f₀ < –0.75 st predicts reasonably well a nasal stop.

Figure 12

Classification tree of the stop categories in production. For each split, the left branch (‘yes’) satisfies the rule based on which the split is made and the right branch (‘no’) does not. Below each node, the predicted probability of each response category in the order ‘lenis,’ ‘aspirated,’ ‘fortis,’ and ‘nasal’ is printed, followed by the percentage of their observations out of the entire dataset.

4.1. Phonetic properties of the four stop series

This study examined the phonetic properties of the three oral stop series and the nasal stop series in phrase-initial position in Seoul and Gyeonggi Korean. We first briefly reported the nasal or denasalized property of nasal stops based on our acoustic data. We have shown an overall weaker nasality and shorter duration of nasal stops in phrase-initial compared to phrase-medial position, but also important variability, both between- and within-speaker, in line with previous studies. It is also possible that denasalization occurs less often in our read speech than in spontaneous speech as reported in Y. S. Kim (2011), but we will not make any comparisons due to the methodological differences between the two studies.

More crucially, we examined the laryngeal properties of the four stop series in phrase-initial position, namely VOT, f₀, and voice quality as estimated by glottal open quotient. Our VOT results of the (voiceless) oral stops are highly consistent with previous reports: shortest for fortis, longest for aspirated, and intermediate for lenis stops; lenis and aspirated stops overlap with each other and they both remain well distinguished from fortis stops. Of the nasal stops, 13% are devoiced, but in the majority of the cases, they can be distinguished from the other three stop series based on VOT alone. Our f₀ results are also in line with previous findings, with a clear bimodal distribution: Aspirated and fortis stops are followed by high f₀, while lenis, nasal stops, and zero onsets are followed by low f₀, and this difference is observed over the entire vowel. For two male speakers who have stayed in Japan longer than the others, f₀ following fortis and lenis stops is in an intermediate zone between aspirated and nasal stops, but no causal link can be established by our limited data.

Voice quality, on the other hand, is less straightforwardly conditioned by the stop onset. Overall, lenis and aspirated stops are followed by breathier vowels than fortis and nasal stops, especially at the beginning of the vowel. We raised two more specific questions in Sections 1.1 and 1.2:

1. How does voice quality on the following vowel differ between aspirated and lenis stops? Is the discrepancy in previous findings due to the time course of the voice quality over the entire vowel?

2. Is the weakening of consonantal nasality compensated by an enhancement of breathier voice on the following vowel (due to its acoustic similarity with nasality)?

For question (1), our results did not show any compelling patterns regarding the time course after aspirated and lenis stops, but they demonstrated speaker variability which could also be a reason why previous studies had different results. Nevertheless, for seven out of nine speakers, the vowel was breathier following aspirated than lenis stops. For question (2), no difference of breathiness was found between nasal and fortis stops. But this result needs to be interpreted with caution due to the low statistical power. Note, however, that Figure 10 shows that for three out of nine speakers, Oq was higher after nasal than fortis stops, which could suggest some articulatory enhancement with a breathier voice after nasal stops, or a more laryngealized voice with their production of fortis stops.

More globally, the contribution of voice quality to the overall distinction between the four stop series is less robust than VOT or f₀, as indicated by the strong overlap of Oq after the four stop onsets and important individual variations, and further confirmed by our CART analysis.

4.2. Mismatch between production and perception

Another issue we raised was whether and how the production data could explain the two intriguing issues in perception with respect to VOT and f₀. VOT and f₀ from our production data are displayed in a scatterplot (Figure 13). Let’s examine them in relation to the observations on previous perception data (see Section 1.3).

Figure 13

Scatterplot of f₀ (st) against VOT (ms). (Ellipse type: multivariate t-distribution).

1. Can the distributions of f₀ and VOT in production explain the asymmetrical perceptual biases in terms of the direction, that is, the perceptual bias from high f₀ or long VOT is stronger than from low f₀ or short VOT?

2. Given a short VOT and a vowel with low f₀, the consonant can be identified as a nasal, fortis, or lenis stop. Are there any other cues that might be involved to distinguish between them in this VOT–f₀ range?

The perceptual bias from high f₀. Concerning the production of f₀, we could expect the distributions of aspirated and fortis stops to be skewed into the low f₀ range (i.e., negatively skewed), but to a lesser degree in the other direction for the other two stop series, so that listeners could be less used to the co-occurrence between high f₀ and nasal/lenis stops. However, this is not exactly the case. As shown in Figure 13, the observed numbers of intrusions are comparable between the high and low f₀ range approximately divided by the 0 semitone line. Median skewness was measured on the f₀ distributions after the four stop series, showing a symmetric distribution for the aspirated series (–0.01)—contrary to the expectation—and skewed distributions of a similar degree for the other three stop series (fortis: –0.27, lenis: 0.26, nasal: 0.22). Thus, the directional bias of f₀ does not find a clear answer in the production pattern. We then propose a tentative explanation which is perception-based: Korean-speaking listeners are more sensitive to high f₀ than to low f₀, which could have a language-independent and/or a language-specific basis. First, the human perceptual system, notably the attentional system, is found to be more sensitive to raised than lowered f₀ in speech (Hsu, Evans, & Lee, 2015, and references therein). Second, the increase of the f₀ difference in Korean might be more attributable to the f₀ raising after aspirated and fortis stops than the f₀ lowering after the other onset series, as suggested by the f₀ enhancement in clear speech of young speakers by raising the high f₀ rather than lowering the low f₀ (K.-H. Kang & Guion, 2008, Figure 1c).

The role of breathiness in production and perception. Concerning the production of VOT, aspirated stops are much more likely to occur within the very long VOT range than lenis stops (which also reflects the historical change of VOT shortening of aspirated stops). This probably leads to a strong bias towards the identification of an aspirated stop when the stimulus has a very long VOT even combined with a low f₀. In the other direction, aspirated stops are rarely produced with a very short VOT, thus we might also expect that short VOTs in the stimuli would exclude the identification of an aspirated stop in perception. However, previous studies have shown that, when followed by a breathy vowel, a short VOT may still be perceived as an aspirated stop. Therefore, we conjecture that aspiration is not a necessary cue in perception when breathiness is involved. It is possibly a perceptual bias that a breathy vowel may be re-interpreted as aspiration in perception. This perceptual bias has probably served as a phonetic basis for a sound change from breathy voice to aspirated onsets, which has occurred in a number of languages, including Central Thai (e.g., Abramson & Erickson, 1992), Tamang (Tibeto-Burman) (Mazaudon, 1978) (see also a Tai dialect in Pittayaporn & Kirby, 2016, and their review of literature). On the other hand, in producing an aspirated stop, aspiration appears as an active gesture, suggested by a long VOT, which corroborates previous physiological studies showing higher glottal opening peak and airflow peak for aspirated than lenis stops (H. Kim et al., 2018).

In the low f₀ range, our production data indicate that short VOT + low f₀ can be mapped into three stop categories: nasal, fortis, and lenis, although neither nasal stops nor fortis stops are prototypically produced in this zone. This mirrors exactly the perceptual confusion in this VOT–f₀ zone. We will come to nasal stops in the following. How, then, do speakers and listeners distinguish lenis from fortis stops? Again, voice quality may come into play. In our production data, an overall difference has been found between lenis and fortis stops, in that lenis stops are produced by a breathier voice than fortis stops. If it is indeed the case that voice quality is a rescue option when the other two cues are ambiguous, we could expect speakers to actively control voice quality, for example, by enhancing the voice quality difference in this ambiguous VOT–f₀ zone. Closer scrutiny into the VOT–Oq relation does suggest this trend. Figure 14 plots normalized Oq against normalized VOT (scaled and centered by speaker to minimize inter-speaker variability), showing a negative correlation, though weak, only for lenis stops (R = –0.22, p < .0001). This suggests that, when VOT of a lenis stop decreases, the following vowel is breathier. In contrast, our production data do not show such a correlation for aspirated stops, suggesting that breathiness on the vowel might be a consequence of aspiration rather than an intentional laryngeal maneuver to produce a breathy vowel.

Figure 14

Scatterplot of normalized Oq (first 3 timepoints) against normalized VOT, with fitted regression lines and coefficients.

To sum up, breathiness seems to be adopted in production as a strategy for the disambiguation of a lenis stop in perception. It is interesting to note that breathiness has been found to trade with multiple parameters perceptually, including nasality (see Section 1.2), aspiration (see this section), low f₀ (Gao, Hallé, & Draxler, 2020; Kuang & Liberman, 2018), and consonantal voicing (Kingston & Diehl, 1994). In Korean lenis stops, breathiness possibly trades with both aspiration and low f₀. On the other hand, breathiness might not be actively produced in the production of an aspirated stop, but it plays a role in perception when it is artificially created in the stimuli.

The perceptual use of nasality. Although nasality is weak and inconsistent in production, two of our previous studies on Korean nasal stops show an interesting paradox in its perception: Nasality contributes to the identification of a nasal stop, but not necessarily to its goodness rating. In one study (Yun et al., 2020), Korean-speaking listeners were presented with isolated /Ca/ stimuli along the oral-nasal continuum created by a mechanical vocal-tract model and were asked to choose between a nasal-onset and an oral-onset syllable. Stimuli with stronger nasal coupling categorically increased the nasal response rate. In the other study with isolated Klatt-synthesized /Ca/ stimuli (Yun & Arai, 2020), longer nasal portion increased the nasal judgement by Korean-speaking listeners, however, two-thirds of the listeners judged oral stimuli as better /na/ syllables than nasal stimuli in a /na/-goodness rating test. These results suggest that nasality plays an important role in the nasal stop identification, and strong nasality is very likely to be a sufficient cue to a nasal stop. However, when categorized in advance as a nasal stop, nasal stimuli were mostly judged as less prototypical of a nasal stop than oral stimuli. In the following section 4.3, we will discuss why nasality, still useful in perception, is weakened in production.

4.3. Syntagmatic variation and paradigmatic contrast

Several proposals have related the synchronic pattern of nasal ‘weakening’—despite this term—to the domain-initial ‘strengthening’ effect which accounts for variations in segmental properties according to their prosodic hierarchy (Cho & Keating, 2001; Yoshida, 2008). Crosslinguistically, a weaker nasal energy in domain-initial than medial nasal stops is also found in French, as evidenced by their results from nasal flow measures averaged across four participants (Fougeron, 2001), although probably to a much lesser degree than in Korean and never fully denasalized. In terms of the duration differences, Fougeron reported that both articulatory and acoustic durations of nasals were longer in domain-initial than medial positions, representing a form of domain-initial strengthening. In English, on the other hand, while the articulatory duration of nasals is longer in utterance-initial than other positions, their acoustic duration is shorter (Fougeron & Keating, 1997). Cho and Keating (2001) also reported such a pattern for three speakers of Korean (although they excluded quite a few utterance-initial nasal stops with very short articulatory duration due to sampling limitations with electropalatography). Their acoustic duration results are corroborated by ours. Therefore, if we assume that denasalization is a form of domain-initial strengthening, it is clear that its phonetic manifestation is different across languages: In Korean, the acoustic duration and the nasal articulation are in fact both reduced. Prosodically driven enhancement is often found to strengthen language-specific segmental features (Cho & McQueen, 2005) and suprasegmental/tonal features (Y. Chen & Gussenhoven, 2008), maximizing paradigmatic contrasts (Georgeton & Fougeron, 2014). However, if we assume that nasals in Korean have the [+nasal] feature, following the analysis of regressive nasal assimilation rule in Yoo (2016), this feature is in fact weakened in prosodically stronger positions. For the present, our reasoning is based on several assumptions, hence we must leave this debate to future research on the interface between segmental and prosodic features (see relevant discussions in Yoo & Nolan, 2020, Section 4.6). On the perceptual side of prosodic units, denasalization may serve as a cue in prosodic parsing. This possibility has been hinted at in Cho and Keating (2001) and Y. S. Kim (2011, p. 142), then supported by empirical evidence from a perceptual study (Yoo, 2018). Plausibly, Y. S. Kim has also pointed out that this perceptual cue is probably made more important by the regressive nasal assimilation process in Korean which may give rise to sequences of nasals at syllable boundaries, calling for a need to demarcate initial boundaries from non-initial nasals.

Another possibility, which might be hidden in plain sight, is that rather than strengthened, nasal stops undergo nasal weakening as a form of hypoarticulation. The motivation of this hypoarticulation could include speech style, lexical frequency, and predictability (Lindblom, Guion, Hura, Moon, & Willerman, 1995), which needs to be addressed in further research. In the meantime, the syntagmatic variation related to prosody/position should not compromise the maintenance of the paradigmatic contrast (when the functional load is high). Thus, nasal stops might resist hypoarticulation in domain-medial position, which would cause perceptual confusion with a lenis stop, while their hypoarticulation would not have this effect in domain-initial position. Although in our read speech, phrase-medial lenis stops are rarely fully voiced, they are more often so in spontaneous speech (e.g., Y. S. Kim, 2011). Even devoiced lenis stops are almost twice shorter in closure duration than fortis stops (e.g., H. Kim et al., 2018), which could lead to a percept of a voiced stop. Therefore, denasalization is avoided to maintain the paradigmatic contrast between nasal and lenis stops. In our previous study on the position-dependent devoicing and f₀ enhancement in Tokyo Japanese, we have explained this different phenomenon using the same motivation: positional variation and the maintenance of paradigmatic contrast (Gao & Arai, 2019; Gao, Yun, & Arai, 2019).

The realization of nasal stops is also highly variable even in phrase-initial position. This is in line with what is observed in corpus data (Yoo & Nolan, 2020) which, based on their apparent-time analysis, further suggests that denasalization is a sound change in progress. This turns our attention to the diachronic change of the nasal series. Denasalization has been reported in a few languages as an ongoing or established sound change, such as the ones reviewed in Y. S. Kim (2011, p. 44). In Karitiana spoken in Brazil, nasal stops are produced as nasal stops by older speakers, but as oral stops by young speakers in word-initial position followed by an oral vowel, and some form of denasalization is also realized in non-initial position (Storto, 1999; Storto & Demolin, 2012). In a number of Chinese dialects, various degrees of denasalization are observed (Hu, 2007), notably in Southern Min dialects (Norman, 1988, p. 235f). What is common to these languages as well as Korean is that they all lack(ed) prevoiced stops in their phonemic inventory. Finally, in some descriptions of several Pacific Northwest languages in North America, nasal stops are also impressionistically transcribed as oral stops or intermediate sounds between oral and nasal stops (Kinkade, 1985; Thompson & Thompson, 1972). One note is of particular interest to us: The phenomenon of denasalization in Pacific Northwest languages is not spread to the varieties that already have an oral voiced stop series (Kinkade, 1985, p. 480). In summary, it seems highly plausible that denasalization of the nasal stops fits better in a phonological system which lacks prevoiced stops, thus not threatening any paradigmatic contrast.

It is also likely that denasalization is more common than generally described, but remains underdocumented crosslinguistically. One of the reasons is that native linguists are ‘deaf’ to it due to their top-down biases, as argued in Y. S. Kim (2011), and acoustic evidence is usually not sufficient to account for nasal properties. Nonetheless, by observing how nasal/denasalized stops interact with other stops in different contexts and positions, we may gain insight into how a language may reach equilibrium between prosodic/positional variations and paradigmatic contrast.

In conclusion, the four stop series in Korean illustrate dynamic clashes between articulatory constraints, perceptual biases, and language-specific syntagmatic and paradigmatic structures, resulting in an ongoing reorganization.