Voice onset time in English and Korean stops with respect to a sound change^*

In English, there are voiced and voiceless stops that voiced stops /b, d, g/ are realized as either voiced or devoiced and voiceless stops are realized as either aspirated or unaspirated. When voiced stops are fully voiced, they are realized with voicing lead (negative VOT). When they are devoiced, they are realized with short-lag VOT (Lisker & Abramson, 1964). When voiceless stops are aspirated, they are realized with long-lag VOT. When they are unaspirated (e.g., /ptk/ after /s/), they are realized with short-lag VOT. These phonetic realizations are well presented in early findings below.

Table 1 shows that VOT changes are taking place. Previous studies have shown that some English speakers produce some or all of their initial voiced stops with prevoicing rather than short-lag VOT. For example, Lisker & Abramson (1964) provide two sets of VOT values for English voiced stops, one with a positive short lag, and the other with a negative voicing lead (Flege, 1982). In contrast, Docherty (1992) reports positive values only for both voiced /b, d, g/ and voiceless /p, t, k/ stops (see also Klatt, 1975). Keating (1984) discusses that English voiced stops are sometimes realized with some lead values but mainly realized with short lag and long lag values, corresponding to Chodroff & Wilson’s (2017). In Table 1, mean VOTs for voiceless stops are getting longer. In Chodroff & Wilson (2017), their grand means for the voiceless stops were somewhat higher than previous findings. They speculate that it reflects an overall slow speaking rate in their experiment. Taking the results of previous studies, it is worth investigating the change in VOT of English stops.

With respect to the VOT boundary between voiced and voiceless stops in English, some studies report that the VOT of /b/ is less than 15 ms, whereas that of /p/ is more than 30 ms (Lieberman & Blumstein, 1988; Nakai & Scobbie, 2016). Other studies report that 25 ms was said to be the boundary between /p/ and /b/ (Abramson & Lisker, 1970). Based on previous reports, it is not easy to determine the ranges of VOT for English voiced and voiceless stops. Keating (1984) points out that, since English stops have a great deal of positional variation, it is hard to simply refer to VOT boundary with a specific number. In addition, there is some overlap in the VOT range of /b/ and /p/. This issue will be examined in this study.

In English, voiceless stops are unaspirated when they follow /s/. They are expected to have short-lag VOT (Keating, 1984). However, less attention was given to their VOT values on how voiceless unaspirated stops differ from devoiced stops as well as Korean tense stops. These questions are answered in the current study by examining them separately from voiceless aspirated stops in English.

In Korean, instead of a voiced-voiceless contrast, there is a three-stop voiceless category, called as ‘tense’, ‘lax’, and ‘aspirated’ stops and transcribed as [p*, t*, k*] (or [p’, t’, k’]), [p, t, k], and [p^h, t^h, k^h], respectively (Han & Weitzman, 1970; Kim, 1965). Because of the lack of a voicing contrast, they are sometimes defined as ‘unaspirated, slightly aspirated, and heavily aspirated’ in terms of the amount of aspiration. The greater aspiration, the longer the VOT. Table 2 presents mean VOTs of Korean stops from previous findings where it is clear that VOT changes are taking place. In Lisker & Abramson (1964), on average, VOTs are shortest (12 ms) for Korean tense stops, longer (30 ms) for lax, and longest (103 ms) for aspirated. There was a VOT overlap between tense and lax stops (Han & Weitzman, 1970; Kim, 1965). However, in the studies of Kim (1994) and Kim (2020), mean VOTs for lax stops are getting longer to 51 ms and 68 ms, while those for aspirated stops are getting shorter to 78 ms and 81 ms. That is, VOT merger between lax and aspirated stops have occurred, replicating previous findings such as Kim (2008), Silva (2006), Kang (2014) and among others.

It is clear that Korean stops are hard to be differentiated in terms of short-lag and long-lag VOT. Many recent studies have already reported that VOT alone is not sufficient to contrast three voiceless stops in Korean and Korean stops are undergoing a sound change (Silva, 2006 and among many others). Since Kim (2000), fundamental frequency (f0 or tone) has arisen as a primary acoustic parameter in distinguishing lax from aspirated stops in Korean (e.g., ‘달’ [t^hàl] and ‘탈’ [t^hál]) (Kim et al., 2002 and among many others).

Unlike English stops, Korean stops have had a long-standing problem not to be explained well under the three VOT dimensions (Lisker & Abramson, 1964 and among many others). Much attention was given to each language, but not much attention was given to the cross-linguistic comparison between English and Korean stops. In addition, very few studies have examined English unaspirated stops. Including English voiceless unaspirated stops after /s/. separately, the research questions in this study are as follows:

The current study is designed to answer these questions.

Twenty four students (11 American and 13 Korean) at different universities participated in the experiment. The Korean data were replicated from Kim (2014). For the English data, eleven native speakers of American English (7 female and 4 male) participated. All were undergraduates or graduates at a university in the Western area of America. The mean age was 24.6 and the individuals ranged from 20 to 29. All but two spoke a western dialect. For the Korean data, thirteen native speakers of Seoul Korean (7 females, 6 males; mean age 25 years) participated. All were undergraduates at a university in Seoul, Korea. They grew up in Seoul and spoke the standard dialect of Seoul. All participants reported that they had no experience in living in an English speaking country.

All participants resided in their countries at the time of the recording. For participants employed in the current study, socio- indexical factors such as age, gender, and dialect were controlled to minimize any plausible effects. In addition, speaking rate was also controlled (see section 2.3). No speakers had any history of speaking pathology nor any phonetic training.

For the English data, eighteen monosyllabic words were constructed from a balanced list of the three stop types (voiced, voiceless aspirated, voiceless unaspirated after /s/) and the three places of articulations (labial, alveolar, and velar) followed by a vowel /a/ context. The syllable type was either CV(C) or sCV(C) where the final coda was an unreleased stop [t]. The words consisted of real and nonsense words (e.g., bot, pot, dot, tot, got, cot, ba, pa, da, ta, ga, ka, spa, sta, ska, spot, stot, scot). Voiceless unaspirated stops (s/ptk/) were intentionally added to see how they differ from English devoiced stops and Korean tense stops.

The same Korean data as in Kim (2014) were employed. Eighteen monosyllabic words (3 stops×3 places×2 syllable types) were constructed from a balanced list of the three stop types (lax, aspirated, and tense) and the three places of articulations (labial, alveolar, and velar) followed by a vowel /a/ context. The syllable type was either CV or CVC where the final coda was an unreleased stop [t] or [k]. The words consisted of real and nonsense words in Korean (e.g., [pat] ‘farm’, [tat] ‘anchor’, [gat] ‘hat’, [pa] ‘method’, [ta] ‘all’, [ka] ‘the edge’, [p^hat] ‘red bean’, [t^hat] ‘blame’, [k^hat] ‘not applicable’, NA’, [p^ha] ‘green onion’, [t^ha] ‘Get in!’, [k^ha] ‘NA’, [p*ak] ‘head’(slang), [t*ak] ‘precisely’, [k*ak] ‘NA’, [p*a] ‘to grind’, [t*a] ‘to pick’, and [k*a] ‘to peel’).

Prior to recording, participants completed a brief language background questionnaire and has a practice session. They read target words in “Say ——— again” in English or [igə ————— hasεjo] “Say this ———” in Korean where the words were located in phrase-initial position. In the position, the target words were expected to be fully emphasized. Sentences were presented either in English or Hangul, the writing system of Korean.

Recordings were done in either Korea or America. In Korea, recordings were done in a quiet office directly into a Samsung SENS NT900XC4C-A78 laptop computer using a Safa voice recorder (model SR-M190N). In the US, recordings were made in a sound-attenuated booth in the Phonetics Lab using a Shure (model SM 10A) head-mounted microphone. Recordings were saved on a flash card using a Marantz digital recorder (PMD 670) at a sampling rate of 44.1 kHz. Each speaker was asked to read the words on the monitor in a natural intonation. The monitor connected to the computer was inside the lab but the computer itself was outside the lab to minimize background noise. The words were automatically popped up at a 3-second interval. This can control speech rate by using the frame sentence and limiting the speakers’ production time of sentences to a fixed 3 sec. There were 792 English tokens (18 words×11 speakers×4 repetitions) and 936 Korean tokens (18 words×13 speakers×4 repetitions), respectively. A total of 1,728 tokens excluding filler words were obtained for analysis. All utterances were analysed using Praat 6.0.37, a speech analysis program (Boersma & Weenink, 2018).

VOT was measured from the release burst of a stop to the onset of periodicity in the waveform (Lisker & Abramson, 1964). The onset of the vowel in the waveform was determined by the onset of the first full glottal pulse of the vowel as well as the pitch of the spectrogram. The onset of the voicing energy in the second formant shown in a time-locked spectrogram was used to help determine voicing onset in conjunction with the waveform. The onset of voicing (=vowel onset) was defined as the first and periodic pulse of a vocalic waveform that show features typical of a vowel. The measurements were semi-automatically done in the following steps: 1) all target words were labeled into consonants (C) and vowels (V) using waveforms and spectrograms, 2) the duration of C and V were automatically obtained using a textgrid.

The VOT data obtained were statistically tested using repeated measures analysis of variance (RM ANOVA) in the context of a univariate context of a general linear model (SPSS/PASW, 2017). The main goal of the present study was to examine whether there is any main effect of stop (i.e., voiceless, voiced 1 & 2, s/p, t, k/ for English and aspirated, lax, and tense for Korean), place (i.e., labial, alveolar, and velar), gender (i.e., male vs. female), and individual speakers on VOT and whether there was any interaction effect between factors. Repeated measures ANOVA includes both “between” subjects effects (language and gender) and “within” subject effects (i.e., stop and place). Their main and interaction effects were statistically analyzed at a 0.05 significant level. Post hoc Tukey HSD (Honest Significant Difference) multiple tests were also run to answer the following questions: (i) for each factor, whether any differences in pairs among types were significant and (ii) for any factor, whether any differences in pairs among the stop types in the individual data were significant. For statistical analysis, the VOT differences among stop types in the two languages were mainly focused because of most central interest in the current paper.

For the group-normative data averaged across eleven native speakers of English, mean VOTs of English stops according to four phonetic types are illustrated in Figure 2. For the voiced category, fully voiced stops with negative VOT values were labeled as “voiced 1” and devoiced stops with positive VOT values were statistically labeled as “voiced 2.” For the voiceless category, voiceless unaspirated stops were labeled as “s/ptk/” and voiceless aspirated stops were labeled as “voiceless” for short. Results of univariate RM ANOVA showed a main effect of stop on VOT [F(3, 691)=4,073.064, p<.0001], in that VOT was positively long for voiceless (104.4 ms), short for both s/ptk/ (20.9 ms) and voiced 2 (20.1 ms), and negatively long for voiced 1 (–115.2 ms). Post hoc Tukey HSD multiple comparisons revealed that differences in any pairs except for the voiced 2-s/ptk/ pair were statistically significant (p<0.5). When voiced stops were devoiced, their VOTs were statistically the same as those of s/ptk/, as clearly seen in Figure 2. This finding is the first to reveal that the VOT difference between devoiced and s/ptk/ stops were not statistically significant. However, this does not mean that they are phonetically the same because of other phonetic correlates.

Figure 2. Mean VOTs (ms) of English stops according to voiced 1, voiced 2, s/ptk/, and voiceless (aspirated) stops aggregated from 11 American speakers (a total of 792 tokens). Error bars indicate ±1 standard deviation.

Download Original Figure

Table 3 presents the numeric VOT value of each phonetic type and their VOT dimensions along with means, ranges, and the number of tokens.

The results in Table 3 tells that, for the voiced category, almost all speakers produced voiced stops with a short-lag VOT, especially in phrase-initial position. This correspond to Chodroff & Wilson (2017), Docherty (1992), Keating (1984), and Klatt (1975), but does not analogous to Flege (1982) and Lisker & Abramson (1964). Taken the findings together, it is safe to say that initial voiced stops in English became fully devoiced. For the voiceless category, mean VOTs for voiceless (aspirated) stops were even longer than those in Chodroff & Wilson (2017), suggesting that English voiceless stops became heavily aspirated. In this study, the devoiced and lengthening process can be considered a sound change.

Concerning the VOT boundary between voiced and voiceless category, as pointed out by Keating (1984), it is not easy to determine the ranges of VOT for English voiced and voiceless stops. The results in Table 3 tells that there is a remarkable overlap in the VOT range of /b/ and /p/. Thus, it is hard to simply refer to VOT boundary with a specific number as previous studies did (Abramson & Lisker, 1970; Lieberman & Blumstein, 1988; Nakai & Scobbie, 2016). However, since there was a huge VOT difference (84 ms) between voiced and voiceless stops, the three VOT dimensions seem to work well for the English data.

Next, there was a main effect of place of articulation on VOT [F(2, 691)=116.671, p<.0001] in that mean VOTs were shortest (40.8 ms) for labials, longer (50.1 ms) for alveolars, and longest (53.1 ms) for velars (labials<alveolars<velars). Post hoc Tukey HSD multiple comparisons revealed that differences among the three places were statistically significant (p<.05). Table 4 shows that, although there is some overlap in the range, the effects of place of articulation on VOT hold well for across stop types. The place effect have been extensively documented in the literature for a variety of languages (Docherty, 1992; Klatt, 1975; Lisker & Abramson, 1964).

Next, consider the effects of gender on VOT. Figure 3 represents mean VOTs of stops according to gender and Table 5 presents their values. There was a main effect of gender on VOT [F(1, 691)= 105.231, p<.0001] in that mean VOTs were statistically shorter (45.7 ms) for female speakers than those (52.1 ms) for male speakers (female<male).

Figure 3. Mean VOTs (ms) of English stops according to voiced 1, voiced 2, s/ptk/, and voiceless stops by male and female speakers. Error bars indicate ±1 standard deviation.

Download Original Figure

Taken Figure 3 and Table 5 together, it is clear that the effects of gender on VOT were very small and linguistically not important. We can speculate that the difference may be mainly due to only one female speaker who produced both voiced and voiceless stops slightly different from other speakers, as discussed below.

Next, consider how individual speakers produced four phonetic types. Figure 4 illustrates individual speaker’s VOT production and Table 7 presents their numeric values. Statistical analysis confirmed significant differences among individual speakers [F(9, 691)=25.667, p<.0001] and a weak interaction between stop and individual [F(18, 691)=20.983, p<.0001] in that individual speakers produced stops differently. However, the differences were not very big, as clearly seen in Figure 4. Out of eleven speakers, one female speaker AF4’s production alone was remarkably different from others. Her data was intentionally included to capture individual speaker variations.

Figure 4. Mean VOTs (ms) of English stops according to voiced 1, voiced 2, s/ptk/, and voiceless aggregated from each individual speaker. Error bars indicate ±1 standard deviation.

Download Original Figure

Taking Figure 4 and Table 6 into a consideration together, we can conclude that there are little speaker variations on VOT for the production of the English stop types. This can be comparable to those of the Korean stops next section.

In summary, the results of English stops show the main effects of stop, place, gender, and speaker on VOT. Voiced stops were rarely produced with voicing lead but mostly with short voicing lag. Devoiced stops were statistically the same as voiceless unaspirated stops after /s/. Mean VOTs for voiceless aspirated stops were relatively longer than previous findings. The results in the current study suggest that English stops undergo a sound change in VOT.

Figure 5 shows mean VOT of Korean stops. For the pool data, results of univariate RM ANOVA showed a main effect of stop on VOT [F(2, 933)=1,188.657, p<0.001] in that mean VOT values were significantly longer (83 ms) for the aspirated and lax stop (71 ms) than for the tense stop (15 ms). Although mean VOTs for aspirated stops were higher than those for lax stops, the VOT differences between them were not statistically significant.

Figure 5. Mean VOTs (ms) of Korean stops according to aspirated, lax, and tense aggregated from 13 Seoul Korean speakers (a total of 936 tokens). Error bars indicate ±1 standard deviation.

Download Original Figure

Post hoc Tukey HSD multiple comparisons revealed that there were two significant subsets: aspirated/lax and tense (tense< lax=aspirated). The present results correspond to previous findings on the VOT merger of lax and aspirated stops (Kang, 2014; Kim, 2020; Silva, 2006). The results suggest that Korean stops truly undergo a sound change.

There was a main effect of place of articulation on VOT [F(2, 933)=11.500, p<.0001] in that mean VOTs were shortest (51 ms) for labials, longer (54 ms) for alveolars, and longest (64 ms) for velars (labials<alveolars<velars). The effect of place of articulation on VOT held well across the three types of stops, as seen in Figure 6(a). Pairwise comparisons revealed that differences in each pair among the three places were statistically significant (p<.005).

Figure 6. Mean VOTs (ms) of (a) Phonation*Place and (b) Phonation *Gender according to aspirated, lax, and tense targets in initial position. Error bars indicate ±1 standard deviation.

Download Original Figure

There was also a main effect of gender on VOT [F(1, 930)= 40.634, p<.0001] and an interaction between phonation and gender F(2, 930)=28.429, p<.0001] in that VOT differences between the aspirated and lax stop were significantly smaller for female speakers than male speakers. Mean VOTs for both aspirated and lax stops were significantly longer for female speakers than male speakers, as can be seen in Figure 6(b).

Table 7 summarizes individual mean VOT values for Korean stops. Compared to the English individual data, inter- and intra- speaker variations on VOT in Korean were much greater.

As already discussed, differences are apparent between male and female speakers in Table 7. There are also intra-speaker differences even in the same gender. For example, all female speakers but SF1, aspirated and lax stops were undergoing a merger as evidence of a sound change. In contrast, all male speakers but SM6 showed a three-way contrast of stops in terms of VOT. Thus, there were inter- and intra-speaker variations on VOT. And the VOT ranges among individuals were largely overlapped.

Figure 7 shows mean VOTs of English and Korean stops together.

Figure 7. Mean VOTs (ms) of English and Korean stops averaged across a total of 1,728 tokens. Error bars indicate +1 standard deviation.

Download Original Figure

There was a main effect of language on VOT (p<.05) in that mean VOTs were significantly shorter in English than in Korean. Post hoc Tukey HSD multiple comparisons revealed that there were four significant subsets: Group 1 for voiced 1, group 2 for the voiced 2, s/ptk/ and tense categories, group 3 for lax and aspirated, and group 4 for voiceless category, as presented in Table 8.

English has a voiced-voiceless contrast, whereas Korean doesn’t. However, the results in Figure 7 and Table 8 show that the two languages are very similar in that their stops were mainly produced with either short-lag or long-lag VOT. Under the long-lag VOT dimension, there are two statistical groups: lax and aspirated stops in Korean form one group and voiceless aspirated stops in English form the other group in Table 8. One questions arises: are they belong to the same dimension or not? Either way is problematic. Table 9 summarizes cross-linguistic differences on VOT.

In Table 9, it is common that both English and Korean stops undergo a VOT shift and the stops in the two languages were mainly produced with either short-lag or long-lag VOT. Except for these, English stops are slightly different from Korean stops in several aspects, as summarized in Table 9.