1.1. Influence of listener and talker characteristics on speech perception

Under a traditional prototype-based view of phonetic representations, dialect-based variability in perception stems from the same source as the variability in production: dialect-specific phonetic prototypes. On a broad level, group differences in perception of phonetic contrasts have been found to reflect those in production. For example, Lee, Politzer-Ahles, and Jongman (2013) found that Kyungsang Korean speakers, who use f0 less than Seoul speakers to distinguish lenis and aspirated stops in production, also rely less on f0 than Seoul listeners in perception. Similarly, work on dialectal differences in vowel perception (e.g., Willis, 1972; Fridland & Kendall, 2012) has shown support for a relationship between overall production and perception patterns (see Thomas, 2002 for a review). However, close examination of these studies reveals that almost none have found a straightforward link between perception and production on all of the predicted contrasts: For example, out of the four predicted differences in Willis (1972), two were supported and two showed “problematic results” (Willis 1972, p. 246).

At the same time, a large body of recent work has explored how social knowledge and expectations influence speech perception. One of the first studies demonstrating the importance of expectations on perception was by Strand and Johnson (1996), who asked listeners to categorize identical /s/-/ʃ/ continua paired with videos of either a male or a female talker. Listeners showed different category boundaries depending on the talker shown in the video, requiring a higher frequency threshold for /s/ classification when the face shown in the video was female, reflecting the fact that the /s/-/ʃ/ boundary tends to be higher in frequency for females than for males. Subsequent work has provided evidence that expectations based on other social factors, including race (Staum Casasanto, 2010), social class (Hay, Nolan & Drager, 2006a; Dufour, Kriegel, Alleesaib, & Nguyen, 2014), persona (e.g. ‘Valley Girl,’ D’Onofrio, 2015), and talker age (Hay et al., 2006b; Koops, Gentry, & Pantos, 2008) also influence speech perception (see Drager, 2010 for a review). However, effects found in these studies are often very small and/or inconsistent, and null results have been found in replications (e.g., Lawrence, 2015; Pharao, Lundholm Appel, Wolter, & Thogersen, 2015; Chang, 2017). This suggests that although there is robust evidence that social information can have an effect on perception, there is much work to be done on understanding how, when, and why this influence occurs.

Few studies have looked at the interaction between listener- and talker-level influences. One notable exception, Evans and Iverson (2004), examined perception of the vowel in English words like ‘bud’ by listeners of two accents of British English that differ in their pronunciation of this vowel, which is produced with a lower F1 in Northern (Sheffield) English than in Southern Standard British English. Listeners heard target words differing in the formant frequencies of the vowel embedded in carrier phrases spoken in the two different accents. Results showed that both listeners’ regional affiliation and the talker’s accent affected listeners’ ratings. Northern listeners’ highest-rated tokens had lower F1 than those of Southern listeners. However, the accent of the carrier phrase also influenced responses in the expected direction, with both listeners preferring tokens with lower F1 when preceded by a Northern accent. These results indicate that both listener and talker information can jointly influence perception, such that neither plays a deterministic role.

Communities in which there is a sound change in progress provide a particularly rich context for exploring the influence of expectations on speech perception. Hay et al. (2006b) found support for their prediction that listeners would use their knowledge of an ongoing merger between the diphthongs in ‘near’ /iə/ and ‘square’ /eə/ in New Zealand English to inform phonetic categorization: Listeners showed less accurate identification of the diphthongs in younger speakers. Similarly, in the context of an ongoing ‘pin’-‘pen’ ‘unmerger’ in Houston, Texas, Koops et al. (2008) found that listeners were more likely to assume a merged system when listening to an older (versus middle-aged) talker. These results indicate that listeners are sensitive to the social distribution of pronunciation variants and that they use this knowledge in speech perception.

While most such studies have focused solely on the age of the talker in predicting how social information will be used, Drager (2011) also examined how the influence of talker age might interact with the age of the listener. Drager found that older, but not younger, participants showed different vowel category boundaries depending on the apparent age of the talker. In other words, older listeners appeared to use age-related variability to inform their speech perception, while younger listeners did not. Drager hypothesized that older listeners, having witnessed the full time course of the sound change, might be more likely to be aware of age-based social patterning. We explore whether there are similar age-based differences in sensitivity to talker-level information in the current work.

1.2. Linguistic background

1.2.2. Contextualizing Chinese Korean

The majority of ethnic Koreans currently living in China are descendants of immigrants who crossed the border between the mid 19th century and the end of the Second World War (Jin, 2008). The Korean communities in China have traditionally maintained strong ties to their Korean culture and language, including Korean-language education. However, what we refer to in this work as ‘Chinese Korean’ is far from a homogenous dialect. Inhabitants of different regions and cities are descendants of speakers of different Korean dialects. There is also influence from both standard North and South Korean: The North Korean standard (Pyongyang) was used as the model for Chinese Korean standardization in the mid-20th century (Tai, 2004), and since the establishment of diplomatic ties between China and South Korea in 1992, there has been an increase in exposure to South Korean culture and media, complemented by an increase in travel to South Korea by ethnic Koreans living in China. Furthermore, the influence of Mandarin has increased in recent years: The majority of speakers are bilingual (Jin, 2008), and there is a shift in dominant language use from Korean to Mandarin in some communities (e.g., Tai, 2004; Han, 2011; Han, 2014).

The current work focuses on two Chinese Korean communities in the border cities of Hunchun (Jilin province) and Dandong (Liaoning province) (Figure 1). Hunchun is smaller (~200,000) than Dandong (~800,000), but has a larger Korean population: It is located within the Yanbian Korean Autonomous Prefecture, and 36% of the residents are ethnic Koreans (China Data Center, 2006). On the other hand, Dandong has a smaller Korean community of around 20,000 (Cuī, 2011). There is high exposure to Seoul Korean in both cities, both via South Korean media and visits to South Korea (details given below). Specific information about the language background of our participants will be reported in the Methods section; in the following section, we discuss some of the linguistic characteristics of the two dialects.

Figure 1

Korean communities in Hunchun and Dandong, China, are the target populations reported in this work. Map modified from an image from Google Maps.

1.2.3. Dialectal differences in production of the stop contrast

As discussed above, there is considerable dialectal variability in production of the stop contrast, both in terms of relative use of specific cues and in whether or not there is an ongoing change in cue use. The varieties examined in the current study, dialects spoken in the cities of Hunchun and Dandong, stem from two distinct North Korean dialects (Hamkyeong dialect for Hunchun speakers and Phyeongan dialect for Dandong). One difference between these two dialects is that Hamkyeong (and in turn, Hunchun) maintains the lexical pitch accent distinction of Middle Korean that was lost in the Western dialects, including Phyeongan/Dandong, as well as Seoul (Ito & Kenstowicz, 2017). Below we discuss previous work reporting production patterns in Hunchun, Dandong, and cognate dialects (Hamkyeong and Phyeongan). We also show graphs of production patterns from the same Hunchun and Dandong participants of the current work, along with a matched group of Seoul participants, in Figure 2. This data is analyzed in Kang, Schertz, and Han (forthcoming), and full statistical results are available in the supplementary materials corresponding to the current article.

Figure 2

Data from 159 speakers from Hunchun, Dandong, and Seoul born in the 1930s-1990s (from Kang et al., forthcoming). Each data point represents a by-participant mean for a given laryngeal category, and the lines show the best-fit regression line. (a) VOT, (b) f0, (c) H1-H2.

Hunchun/Hamkyeong: Previous studies on various varieties of Hamkyeong Korean have shown large and consistent differences from present-day Seoul patterns, with fortis and lenis stops both showing very short VOT values, in contrast to very long VOTs for aspirated stops (Ito & Kenstowicz, 2008; Chung, 2011; Kang & Han, 2012; Kang et al. forthcoming). In terms of f0, lenis stops are to be systematically lowest, and aspirated stops highest, but the difference is quite small, and can vary based on pitch-accent status (Kang et al., forthcoming). In terms of voice quality, all work on Hamkyeong Korean shows substantially lower H1-H2 for fortis than for lenis and aspirated stops, similar to Seoul Korean, and Kang and Han (2012) further showed that speakers from Qingdao, China (who speak a variety of Hamkyeong Korean) produced lenis stops with higher H1-H2 than aspirated stops. This small but systematic effect was mirrored in the data of Kang et al. (forthcoming) in Hunchun speakers, but a direct cross-dialectal statistical comparison showed no significant differences between use of H1-H2 in Hunchun versus Seoul.

Dandong/Phyeongan: Previous studies on Phyeongan Korean have been mixed. Jin (2008) and Jin and Silva (2017) found VOT patterns similar to those seen in Seoul, including an ongoing merger between lenis and aspirated stops, based on a production study of thirty-five 18–74-year-old speakers in Shenyang, the capital of Liaoning province. However, in Dandong (a different city in the same province), while there is evidence of an ongoing VOT merger, it is not as far advanced as the one in Seoul, such that Dandong speakers’ VOT values for lenis and aspirated stops being better separated, and f0 values being less well-separated, than their age-matched counterparts in Seoul (Kang & Han, 2012; Kang et al., forthcoming). In terms of voice quality, Kang and Han (2012) and Kang et al. (forthcoming) found, as expected, the lowest H1-H2 values for fortis stops, as well as systematically lower H1-H2 for aspirated than for lenis stops.

In order to formulate predictions for the current work, we summarize the use of cues in terms of each two-way contrast separately, and how this differs across dialects. These generalizations are based on previous production studies, as well as production data from the current participants, as reported in Kang et al. (forthcoming) and available in the supplementary materials.

Lenis-fortis contrast: The fortis and lenis categories are highly overlapping in both VOT and f0 for Hunchun speakers, as in other Hamkyeong varieties (Kang et al., forthcoming). Dandong speakers have greater separation between the two categories on both dimensions, and the categories are even further separated in Seoul speakers. The relative overlap of the two categories in Hunchun suggests that Hunchun listeners may rely more heavily on other cues, such as voice quality of the following vowel, to make the distinction. A greater use of H1-H2 in Hamkyeong versus Seoul speakers was found by Kang and Han (2012), and this idea was also supported by patterns of a single speaker reported in Ito and Kenstowicz (2008); however, in production data from Hunchun, Dandong, and Seoul participants reported in Kang et al. (forthcoming), there were no apparent dialectal differences in use of H1-H2 for the lenis-fortis contrast.

Lenis-aspirated contrast: This contrast is the focus of most of the work done on dialectal and diachronic variability in the three-way contrast. VOT values for the two categories are well-separated in Hunchun speakers, almost totally overlapping in Seoul, and intermediately separated in Dandong. The use of f0 goes in the opposite direction: There is substantial overlap in Hunchun, in contrast to complete separation in Seoul, with an intermediate level of separation in Dandong. In terms of H1-H2, there is evidence of higher values corresponding to lenis versus aspirated stops in Hunchun and Dandong (Kang & Han, 2012; Kang et al., forthcoming), but this difference is very small.

In terms of within-dialect variability, no age-related differences have been found in Hunchun speakers except a small decrease in VOT for aspirated stops (Kang et al., forthcoming), but in both Dandong and Seoul, there is more overlap in VOT, and less in f0, in younger than in older speakers; nevertheless, the younger Dandong speakers do not show as much VOT overlap as older Seoul speakers (Kang & Han, 2012; Kang et al., forthcoming). Finally, in Seoul, females showed more merged VOT than their male counterparts (Oh, 2011; Kang, 2014; Kang et al., forthcoming). No consistent gender differences were found in Hunchun or Dandong.

Fortis-aspirated contrast: There is no clear dialectal variation in the fortis-aspirated contrast between Hunchun and Dandong listeners. In Seoul, fortis stops have, on average, a slightly lower f0 than aspirated stops, but the two categories remain well-separated in terms of VOT, so we expect consistently primary use of this cue in all dialects. We also expect voice quality to play a role, as this separates the cues in all dialects.

1.3. The current study

This study presents a systematic investigation of perceptual cue weighting in two dialects of Korean which have been shown to differ both in use of cues in production and in the existence of a sound change in progress (where Dandong shows a clear sound change and Hunchun does not). We explore whether these dialectal and age-related differences are mirrored in perception, and whether listeners modify their perceptual strategies depending on the (apparent) dialectal affiliation of the talker.

We test listeners’ use of VOT, f0, and voice quality in perception of the stop contrast via a series of identification experiments, where stimuli have been manipulated to vary orthogonally on these three dimensions. Using a novel analytical method in an effort to avoid problematic aspects associated with analysis of the three-way contrast in previous work, we first examine how listeners’ phonetic category space along these dimensions differs by dialect and age. We then test whether listeners’ knowledge of dialectal variation affects perception by comparing participants’ responses across two conditions, one in which they are told that the talker is a member of their local dialect community, and the other in which the talker is said to be from Seoul.

Following previous work, we assume that the primary perceptual cues to the stop contrast are VOT and f0. However, our design also includes a manipulation of the baseline vowel (taken from natural productions of fortis and nonfortis vowels), in order to systematically assess the role that vocalic information, most notably voice quality, may play in listeners’ perception of the contrast. Voice quality has been shown to serve as a cue in production (e.g., Cho et al., 2002) and perception (Kim et al., 2002; Kong, Beckman, & Edwards, 2011). However, in previous perception studies, voice quality was not manipulated separately from f0, making it difficult to assess its contribution independently of f0.

3.1. Perception of the three-way contrast in the VOT-f0 space

Figure 5 shows overall perception patterns across the acoustic space used in our experiments: Stimuli varied in VOT, f0, and baseline vowel (fortis and nonfortis). These heat plots allow for visualization of the overall patterns in the same dimensionality as the production data. However, it is difficult to see subtle differences, so we show here only the factors that had the largest influence on our results: Listener Dialect and baseline Vowel. Responses are broken down by all other factors in detail using categorization curves in the subsequent sections.

Figure 5

Overall perception results, with darkness showing the percentage of listeners’ responses for (a) lenis, (b) fortis, and (c) aspirated stops in the VOT-f0 space. In each plot, separate sub-plots are shown for Hunchun (left panel) and Dandong (right panel) listeners, and for nonfortis (top panel) and fortis (bottom panel) baseline vowels separately.

We focus first on the two primary dialect-related differences seen in the production data in Figure 2. First, we saw that Hunchun speakers showed more overlap in the VOT-f0 space than Dandong speakers in production of the fortis-lenis contrast. We can see a similar pattern in these perception graphs: The low-VOT region (left side of the plots) shows more of a mixture of lenis and fortis responses for Hunchun listeners, as compared to Dandong, for whom the category boundary is somewhat more clear-cut. Comparing the perception of stimuli across the fortis and nonfortis vowel conditions, we see that the baseline vowel does influence categorization in perception: Focusing on the middle graph (% Fortis choice), we see that there are overall slightly more fortis than nonfortis responses when the vowel is fortis (bottom panel) in both dialects. However, this difference is small and clearly not a categorical cue in either dialect: If it were, we would find no fortis responses in the nonfortis vowel condition.

The second main dialectal difference in the production data was the relative placement of the lenis-aspirated contrast in the VOT-f0 space, with Hunchun speakers producing primarily a VOT contrast (with little to no f0 contrast), while Dandong speakers’ productions contrasted more in f0, and less in VOT. We see this pattern mirrored in the perception results most clearly by looking at the Aspirated plot: In Hunchun, Aspirated responses are most clearly defined by long VOT, while in Dandong, stimuli with long VOT are not necessarily categorized as aspirated, and this can be attributed to the fact that the lenis category extends across the low-f0 region, even when VOT is long, in Dandong speakers.

Overall, then, it appears that the main trends in production are reflected, at least to some degree, in perception. We now turn to a detailed examination of each pairwise contrast, and how perception differs based on all of our factors of interest.

3.1.1. Perception of the fortis-aspirated contrast

Results of listeners’ perception of the fortis-aspirated continuum (i.e., those stimuli with the highest values of f0) are shown in Figure 6 and Table 3. The graphs show listeners’ responses across the fortis-aspirated continuum (spanning the range of VOT values along the highest step of f0), with line types showing how responses compare across each level of our factors of interest. Our only prediction regarding this contrast was that we expected fortis vowels to elicit more fortis responses; we did not expect any differences in perception of this contrast based on listener or talker characteristics.

Figure 6

Categorization curves along the fortis-aspirated continuum, broken down by each predictor variable: (a) Listener Dialect, (b) Age, (c) Vowel, (d) Talker, (e) Dialect Label.

Table 3

Output of statistical model predicting aspirated response along the fortis-aspirated continuum. The reference level for each factor is in italics.

	Estimate	SE	z value	p value
Intercept	4.676	0.292	15.999	<.001***
Continuum Step	2.287	0.124	18.490	<.001***
Listener Dialect (Hunchun versus Dandong)	0.061	0.176	0.348	0.728
Age (older versus younger)	–0.342	0.176	–1.946	0.052.
Vowel (nonfortis versus fortis)	–0.197	0.088	–2.240	0.025*
Talker (female versus male)	–1.019	0.143	–7.148	<.001***
Dialect Label (Local versus Seoul)	–0.018	0.125	–0.144	0.886

• ***p < .001; **p < .01; *p < .05; .p < 0.1.

We first observe general patterns which hold across all factors, then examine each factor separately. As expected, aspirated responses range from 0% at the lower endpoint to 100% at the high endpoint of the continuum, and listeners primarily chose fortis when not choosing aspirated. However, note that at the low end of the continuum there was a small but substantial percentage of lenis responses, which we had not expected. In other words, our assumption that listeners’ decisions along this continuum are constrained to a binary choice does not always hold at the lower end of the continuum. Nevertheless, our response variable as it is formulated (‘Aspirated’ versus ‘Other’) does behave as expected across all conditions.

Our logistic regression model predicts aspirated response (the purple lines in Figure 6) as a function of Continuum Step (in this case, VOT step), Listener Dialect, Age, Vowel, Talker, and Dialect Label. Based on the model selection procedures laid out above, the optimal model included random slopes for Continuum Step and Talker, and the model reported below with no interactions in the fixed effects did not differ significantly from a model with all interactions. As expected, there was a significant effect of continuum step, indicating an increase in Aspirated response across the continuum, and the significant intercept indicates an overall higher probability of Aspirated response.

Listener-level effects: There was no significant main effect for Listener Dialect, shown in the corresponding graph by the similar shape of the Aspirated response curves across the two dialects. Overall, listeners of the different dialects and age groups do not differ in their perception of the Fortis-Aspirated contrast.

Vowel effects: There was a main effect of Vowel, with fortis vowels having a negative coefficient, corresponding to fewer aspirated responses and a slightly higher category boundary; in other words, the fortis category extended further on the continuum when the baseline vowel was fortis. However, this effect is very small, with an average crossover difference of only 0.10 steps.

Talker effects: There was a main effect of Talker, with a lower coefficient for the male talker representing a 0.51 step higher boundary for the male than the female talker.

Dialect Label effects: The effect of Dialect Label was not significant.

3.1.2. Perception of the lenis-fortis contrast

Results of listeners’ perception of the lenis-fortis continuum (i.e., those stimuli on the left edge of the VOT-f0 space) are shown in Figure 7 and Table 4. The graphs show listeners’ responses across the lenis-fortis continuum, with line types showing how responses compare across each level of our factors of interest. As expected, we see increasing fortis and decreasing lenis responses across the continuum, and there are practically no aspirated responses. However, the curves are flatter than the idealized categorization curves we would expect to see, and they do not reach floor or ceiling at the endpoints. Furthermore, the slopes of the curves differ by condition: For example, Dandong listeners’ responses appear to be more categorical (i.e., showing a steeper slope) than those of Hunchun listeners. We discuss implications of this, and do a follow-up analysis, after our primary analysis (see Section 3.2).

Figure 7

Categorization curves along the lenis-fortis continuum, broken down by each predictor variable: (a) Listener Dialect, (b) Age, (c) Vowel, (d) Talker, (e) Dialect Label.

Table 4

Output of statistical model predicting fortis response along the lenis-fortis continuum. The reference level for each factor is in italics.

	Estimate	SE	z value	p value
Intercept	–0.074	0.130	–0.568	0.570
Continuum Step	1.244	0.070	17.885	<.001***
Listener Dialect (Hunchun vs. Dandong)	0.258	0.364	0.709	0.478
Age (older vs. younger)	–0.452	0.249	–1.817	0.069.
Vowel (nonfortis vs. fortis)	1.367	0.112	12.182	<.001***
Talker (female vs. male)	–0.510	0.136	–3.748	<.001***
Dialect Label (Local vs. Seoul)	0.296	0.132	2.243	0.025
Listener Dialect * Age	0.593	0.491	1.207	0.227
Listener Dialect * Vowel	–0.549	0.253	–2.172	0.030*
Listener Dialect * Talker	–0.607	0.374	–1.622	0.105
Listener Dialect * Dialect Label	–0.680	0.265	–2.560	0.010*
Age * Vowel	0.098	0.186	0.528	0.597
Age * Talker	–0.058	0.267	–0.216	0.829
Age * Dialect Label	0.152	0.260	0.586	0.558
Vowel * Talker	–1.203	0.165	–7.280	<.001***
Vowel * Dialect Label	–0.356	0.164	–2.172	0.030*
Talker * Dialect Label	–0.484	0.491	–0.986	0.324
Listener Dialect * Age * Vowel	0.162	0.370	0.438	0.662
Listener Dialect * Age * Talker	1.471	0.528	2.787	0.005**
Listener Dialect * Age * Dialect Label	–0.607	0.516	–1.177	0.239
Listener Dialect * Vowel * Talker	–0.086	0.330	–0.260	0.795
Listener Dialect * Vowel * Dialect Label	–0.259	0.329	–0.786	0.432
Listener Dialect * Talker * Dialect Label	–0.569	1.012	–0.563	0.574
Age * Vowel * Talker	0.115	0.329	0.351	0.726
Age * Vowel * Dialect Label	0.840	0.329	2.552	0.011*
Age * Talker * Dialect Label	2.786	0.984	2.832	0.005**
Vowel * Talker * Dialect Label	0.358	0.372	0.964	0.335

Our logistic regression model predicts fortis response (the yellow lines in Figure 7) as a function of Continuum Step, Listener Dialect, Age, Vowel, Talker Voice, and Dialect Label. Based on the model selection procedures laid out above, the optimal model included random slopes for Continuum Step, Vowel, and Talker Voice, and three-way interactions in fixed effects. As expected, there was a significant effect of continuum step, indicating an increase in fortis response across the continuum.

Listener-level effects: There were no significant main effects of Listener Dialect or Age, which we take to indicate a lack of any major effects primarily determined by these listener-level factors. However, both Listener Dialect and Age showed significant interactions with other factors, as discussed below.

Vowel effects: There was a significant main effect of Vowel, with a fortis vowel eliciting more fortis responses, corresponding to a 1.37 step decrease in category boundary when compared to the nonfortis vowel condition. There were significant two-way interactions of Vowel with Listener Dialect, Talker, and Dialect Label, as well as a significant three-way interaction between Vowel, Age, and Dialect Label. The magnitude of the effect of Vowel broken down by these interacting factors is shown in Figure 8. The graph shows the difference in predicted ‘fortis’ response (in terms of log odds) between baseline fortis and nonfortis vowels for each subgroup. The direction of the effect was the same across all subgroups (with fortis vowels eliciting more fortis responses), but groups differed in the extent of this effect. Hunchun listeners made (slightly) more use of the baseline vowel than Dandong listeners, and Vowel influenced listeners’ responses for the female more than for the male talker. In terms of the three-way interaction, there was a significant difference in the use of Vowel by Dialect Label for older, but not for younger, listeners, with older listeners using Vowel more in the Seoul than in the Local condition.

Figure 8

Interaction effects for the lenis-fortis contrast. In each plot, the point shows the predicted difference in log odds for ‘fortis’ choice for each level of the relevant factor (Vowel in (a), Talker in (b), and Dialect Label in (c)), broken down by subgroups that interacted significantly with the factor. For example, both Dandong and Hunchun listeners were more likely to respond fortis for fortis than nonfortis baseline vowels, as represented by the positive effect size, but this vowel-based difference is of greater magnitude in Hunchun than in Dandong. Error bars show one standard error based on model predictions (testInteractions function in the phia package in R, De Rosario-Martinez, 2015). Black lines indicate a significant effect of the relevant factor (e.g., Vowel in the first panel) for each subgroup.

There were main effects of both Talker and Dialect Label, as well as significant interactions of these two factors with others, as shown in Figure 8. However, as can be seen in the interaction plots, there was not a clear or consistent effect of either of these factors.

Talker effects: The main effect for Talker shows that the female talker elicited more fortis responses than the male talker, and follow-up tests suggest that this effect is driven by fortis vowels: When the vowel was fortis, there were more fortis responses for females than males. Put together with the findings from Vowel above, this corresponds to an overall greater apparent use of voice quality for the female talker⁶. However, in terms of the interaction with Listener Dialect, Age, and Dialect Label, the effects do not appear consistent: Only one subgroup showed significant effects (the female talker elicited more fortis responses for older Dandong listeners in the Seoul condition). Given the inconsistency of the effects, we do not interpret this further.

Dialect Label effects: We did not have any predictions regarding Dialect Label but found a main effect of Dialect Label, which indicates that overall, the Seoul condition elicits more Fortis vowels than the Local condition. The effect appears to be driven by Hunchun listeners, while there is no effect for Dandong listeners. Broken down by Vowel, Age, and Talker, no consistent Dialect Label effects emerge. We therefore do not interpret these further.

3.1.3. Perception of the lenis-aspirated contrast

Results of listeners’ perception of the lenis-aspirated continuum are shown in Figure 9 and Table 5. The graphs show listeners’ responses across the hybrid lenis-aspirated continuum (ranging across low-f0 and continuing up the high-VOT dimension), with line types showing how responses compare across each level of our factors of interest. As expected, we see an increase in Aspirated responses across the continuum, and a corresponding decrease in Lenis responses. At the beginning of the continuum, there are some (unexpected) Fortis responses, more so for Hunchun listeners. This reflects what we have seen earlier: Some listeners have categorically fortis responses across the whole low VOT range. As this has already been accounted for in the lenis-fortis analysis, we focus on aspirated responses, where we see the full range of responses (0% to 100%).

Figure 9

Categorization curves along the hybrid lenis-aspirated continuum, broken down by each predictor variable (for this contrast, the effect of age is shown separately for Hunchun and Dandong listeners because the effect of age is substantially different in the two dialects).

Table 5

Output of statistical model predicting Aspirated response along the lenis-aspirated continuum. The reference level for each factor is in italics.

	Estimate	SE	z value	p value
Intercept	1.448	0.195	7.434	<.001***
Continuum Step	1.219	0.043	28.529	<.001***
Listener Dialect (Hunchun versus Dandong)	–2.866	0.300	–9.549	<.001***
Age (older versus younger)	–1.863	0.299	–6.223	<.001***
Vowel (nonfortis versus fortis)	0.089	0.072	1.233	0.218
Talker (female versus male)	0.708	0.165	4.289	<.001***
Dialect Label (Local versus Seoul)	–0.728	0.168	–4.344	<.001***
Listener Dialect * Age	–1.437	0.519	–2.771	0.006**
Listener Dialect * Vowel	0.048	0.142	0.335	0.738
Listener Dialect * Talker	0.478	0.307	1.558	0.119
Listener Dialect * Dialect Label	0.075	0.319	0.233	0.815
Age * Vowel	0.146	0.142	1.029	0.304
Age * Talker	0.296	0.306	0.966	0.334
Age * Dialect Label	–0.669	0.319	–2.097	0.036*
Vowel * Talker	–0.244	0.135	–1.798	0.072.
Vowel * Dialect Label	0.089	0.135	0.659	0.510
Talker * Dialect Label	0.513	0.578	0.889	0.374

Our logistic regression model predicts aspirated response (the purple lines in Figure 9) as a function of Continuum Step (on the hybrid VOT/f0 continuum), Listener Dialect, Age, Vowel, Talker, and Dialect Label. Based on the model selection procedures laid out above, the optimal model included random slopes for all within-subjects factors (Continuum Step, Vowel, Talker, and Dialect Label), and two-way interactions in fixed effects. As expected, there was a significant effect of continuum step, indicating an increase in aspirated response across the continuum. The positive intercept indicates more aspirated responses than chance, as indicated by the crossover point below the midpoint of the continuum.

Listener-level effects: Listener Dialect, Age, and their interaction were significant, and Age also interacted with Dialect Label. Dandong listeners had a higher crossover point (i.e., fewer aspirated responses, corresponding to more use of f0 relative to VOT) than Hunchun listeners, and within each group, younger listeners had a higher crossover point than older listeners. The significant interaction indicates that the age effect is larger in Dandong than in Hunchun (though follow-up tests show it is significant in both dialects). In terms of the interaction with Dialect Label, the age effect is significant in both Dialect Label conditions, though slightly greater in the Seoul condition; this interaction will be accounted for in the Talker-level effects section below.

Vowel effects: The main effect of Vowel was not significant, nor were its interactions with other factors, indicating that the voice quality of the vowel does not influence the lenis-aspirated category boundary.

Talker effects: There was a main effect of Talker, with the female talker eliciting fewer aspirated responses (i.e., a higher crossover point, corresponding to more use of f0) than the male talker. Talker was not involved in any significant interactions.

Dialect label: There was also a main effect of Dialect Label, with the Seoul talker eliciting fewer Aspirated responses (i.e., higher crossover point/more use of f0) than the Local talker. The aforementioned interaction of Dialect Label with Age indicates that the effect is larger for younger than older listeners. Follow-up tests showed that the Dialect Label effect was significant (p < .001) for younger listeners, and not significant (p = .086), but numerically in the same direction for older listeners.

3.1.4. Interim summary and interpretation

In order to compare the magnitude of overall effects, Figure 10 shows the absolute values of the difference in crossover points between the two levels of each factor. The color of the bars indicates the significance of the factor: Black bars indicate overall significance of the main effect, either without interactions, or, if there are interactions with other factors, significance holds at each level of the other factor. Grey bars indicate that the factor interacted with one or more other factors, and that while the direction of effect was consistent, it was not always significant when broken down by other factors. White bars indicate that the factor showed inconsistent or no significant effects.

Figure 10

Differences in 50% crossover points between the two levels of each predictor variable. Units are continuum steps. Bar colors represent significance (see details in-text).

Fortis-Aspirated contrast: The only factor predicted to affect the fortis-aspirated category boundary along the VOT dimension was Vowel, and it did, in the expected direction. However, the effect with the greatest magnitude for the fortis-aspirated contrast was one that was not predicted, Talker. The fortis category extended further along the VOT dimension for the males. Perception of the fortis-aspirated contrast did not appear to differ based on listener-level characteristics or Dialect Label.

Lenis-Fortis contrast: As expected, listeners used information in the following vowel to inform their categorization across the lenis-fortis continuum, and Hunchun listeners used this information more than Dandong listeners. There was also more use of vowel for the female than for the male talker, which appeared to be driven by the fact that the female’s fortis vowels elicited more fortis responses, as opposed to a symmetrically larger contrast. Hunchun, but not Dandong, listeners showed a small Dialect Label difference, with the Seoul condition eliciting more Fortis responses than the Local condition. There were not consistent overall differences based on Listener Dialect or Age.

Lenis-Aspirated contrast: As predicted, there were strong and consistent dialect and age-related differences in perception of the lenis-aspirated contrast, with Dandong listeners showing higher category boundaries (corresponding to heavier reliance on f0 than VOT) than Hunchun listeners. The same pattern holding for younger versus older listeners, and an interaction indicated a greater effect of age for Dandong listeners. There was a consistent influence of Talker on responses, with more use of f0 for the younger, female talker. Finally, Dialect Label showed an overall effect in which the Seoul label elicited a higher category boundary than the Local label, but this was only significant for younger listeners.

3.2. Follow-up: Dialectal differences in predictiveness of f0 versus voice quality in the lenis-fortis contrast

Our choice of continua assumes that the primary dimensions influencing the category boundary are VOT (for fortis-aspirated), f0 (for lenis-fortis), and combined VOT-f0 (for lenis-aspirated). That this assumption holds can be seen clearly from most of the graphs: We see floor and ceiling categorization effects at the ends of the relevant continuum, with only a small difference in category boundary for other factors. However, as discussed above, this is not the case for the lenis-fortis contrast. We found a strong effect of vowel, and the relatively shallow slope of the curves implies that although f0 was an important predictor of Fortis versus Lenis response, this expected primary cue was not as ‘primary’ as the dimensions covered by the continua for the other two contrasts.⁷

We therefore performed a direct test of the predictiveness of f0 as compared to baseline vowel in categorization of the lenis-fortis contrast. We built three models predicting Fortis response from 1) f0 Continuum Step, 2) Vowel, and 3) both f0 Step and Vowel. Listener Dialect was also included as a fixed factor in both models since we found differences in use of the dimensions across the two dialects, and random by-subjects slopes and intercepts were included. It is not possible to directly compare the effect size of the two predictors given that they are on different scales; therefore, we quantified predictiveness by comparing how well each model predicted listeners’ responses. Table 6 gives confusion matrices showing how model predictions compared with listeners’ responses, as well as d’, a measure of accuracy.

We focus on the d’ scores as a metric of the predictability of the models. As expected from the previous results, the vowel-only model was more successful in predicting responses of Hunchun listeners than Dandong listeners (70% versus 63%), while the f0 continuum step model was more successful for Dandong than Hunchun listeners (82% versus 74%). The fact that the accuracy of the vowel model was lower for both dialects suggests that f0 is more predictive than vowel in the lenis-fortis distinction, even for Hunchun listeners. The combined models for each dialect were more accurate than either of the individual models, as expected; however, the model was still substantially more successful for Dandong than Hunchun listeners (84% versus 77%). Therefore, it appears that Hunchun listeners’ responses are generally not well-explained by the parameters manipulated in this work, nor by age or talker information examined here, as evidenced by the fact that they did not show consistent significant effects in the main analysis above.

4.1. Summary of main findings

We approached this study with hypotheses stemming from two assumptions: 1) the listener’s perception is guided by dialectal and age-based production norms, and 2) listeners adjust their perceptions based on the apparent dialect of the talker. Overall, the results of the experiments show that both listener- and talker-related characteristics do affect perception, in the direction predicted by our hypotheses; however, in both cases, we also found unpredicted effects or asymmetries that raise further questions.

For the first, ‘listener-level,’ premise, we found that the two major dialectal differences seen in Hunchun and Dandong production patterns, reported in Kang et al. (forthcoming), had analogs in perception. First, f0 was indeed less predictive of Hunchun listeners’ than Dandong listeners’ responses, reflecting the fact that the lenis-fortis contrast showed more f0 overlap in Hunchun than in Dandong productions. Second, dialectal differences in production of the lenis-aspirated contrast (greater f0 and smaller VOT differences in Dandong versus Hunchun talkers), were also reflected in perception, as evidenced by a higher category boundary for Dandong listeners on a ‘hybrid continuum’ that was designed to quantify the relative use of the two dimensions. Finally, an age-based difference in perception of the lenis-aspirated contrast was predicted for Dandong listeners, where production data suggests a change in progress. This prediction was supported in that there was an age effect in Dandong, with younger listeners showing increased reliance on f0 relative to VOT. However, an age-related difference in the same direction, albeit of a smaller magnitude, was found in Hunchun as well, which was not predicted given that there is no apparent sound change in Hunchun production. This unexpected finding is discussed further below.

For the Dialect Label manipulation, we expected that listeners’ responses would be affected by their expectations, based on their experience with dialectal differences, namely that Seoul speakers make more use of f0 in production of the lenis-aspirated contrast than Dandong or Hunchun speakers. In support of this, listeners’ category boundaries indicated greater use of f0 when listening to the Seoul talker, although this was only significant for younger listeners. We also considered the possibility of the influence of Dialect Label on the lenis-fortis contrast, given dialectal differences in production of this contrast. We found that Hunchun listeners showed more fortis responses, i.e., a lower category boundary along the f0 continuum, in the Seoul than the Local talker condition. We give one possible explanation for this result below.

Finally, we found two Talker-related differences. First, in the fortis-aspirated contrast, the older, male talker elicited a higher category boundary (i.e., more fortis responses) along the VOT continuum. Second, in the lenis-fortis contrast, the younger, female talker elicited more fortis responses than the male, but only when the vowel was a baseline fortis vowel.

4.3. Use of voice quality in perception of the Korean laryngeal contrast

While most previous work looking at dialectal differences in perception of the laryngeal contrast has focused on VOT and f0, in this work we showed that listeners from different dialects also vary in their use of voice quality, or other spectral information present in the vowel. The baseline vowel manipulation used in this study provides the first systematic investigation into the independent role of voice quality in perception of the Korean stop contrast, supporting findings by Kim et al. (2002) (Chang, 2013 showed a similar secondary use of voice quality in perception of the Korean fricative laryngeal contrast). Our results showed that, as expected, voice quality of the following vowel played influenced listeners’ perception, with fortis vowels eliciting both a small increase in fortis responses along the fortis-aspirated continuum (0.2 steps), and a more substantial increase along the lenis-fortis continuum, with a fortis vowel lowering the threshold for a fortis response by almost a whole continuum step (0.9 steps). Voice quality did not affect perception of the lenis-aspirated contrast, which was expected given the similar breathier (high H1-H2) quality of vowels following lenis and aspirated stops, which may also underlie inconsistent findings in previous production studies (e.g., Cho et al., 2002; Ahn, 1999). The two base vowels in this study represented the two ends of the voice quality continuum and were therefore not expected to capture the smaller voice quality difference between lenis and aspirated stops. A systematic examination of the acoustic and perceptual correlates of voice quality as it is used in the laryngeal contrast, going beyond the most widely-used metric of H1-H2, is a topic for future work.

One of our expectations was that Hunchun listeners would make less use of f0, and rely more on voice quality, as predicted by a claim about Yanbian Korean, a related dialect, by Ito and Kenstowicz (2008), and by their production patterns, which show a high degree of overlap on the f0 dimension. We found that Hunchun listeners do in fact make less use of f0 and more use of voice quality than Dandong listeners. However, this dialectal difference is small, even smaller than the difference in voice quality use based on the talker. Furthermore, if Hunchun listeners showed a primary reliance on H1-H2, as predicted by Ito and Kenstowicz (2008), then a model with Vowel as a predictor should better predict listeners’ responses than a model with f0 as a predictor. However, our follow-up analysis found that Hunchun listeners still rely more on f0 than on Vowel, and that the overall predictability of the two dimensions together is less for Hunchun than Dandong. In other words, while we found the expected effect of less use of f0 by Hunchun listeners, we did not find a compensatory cue that they rely on instead. Whether listeners use other cues that were not considered in this study, whether the lack of results were a function of the specific task used in this study, or whether the contrast is simply less robust, therefore remains an open question.

4.4. Perceptual cue-weighting in the context of sound change

Production patterns show evidence of change in Dandong speakers’ use of acoustic cues to the lenis versus aspirated stop contrast, in the form of an ongoing VOT merger, similar to the well-known sound change in Seoul. On the other hand, production of the contrast appears stable across age groups in Hunchun. Interestingly, while evidence of a sound change in production was only clearly present in Dandong, we found age-related differences in perception of the lenis-aspirated contrast in Hunchun as well, with younger listeners in both dialects showing decreased reliance on VOT relative to f0, in comparison with their older counterparts. Here we discuss several potential reasons for this finding.

One possibility is that there is an ongoing sound change in Hunchun. There are two age-related differences reported in Kang et al. (forthcoming) that could be interpreted as a change in progress: First, younger speakers’ aspirated stops are shorter in VOT. This is unlikely to be driving a sound change, since lenis and aspirated stops are not approaching a merger. Another finding is that speakers are using less f0 in pitch accent, which Lee and Jongman (2018) argue allows for greater use of f0 in the laryngeal contrast. However, our Hunchun listeners if anything use less f0 to signal the laryngeal contrast. Alternatively, there could be a sound change in Hunchun that is not captured by the type of production data (reading words in isolation) reported here. Bang, Sonderegger, Kang, Clayards, and Yoon (2015) found that both VOT merger and f0 enhancement in the Seoul lenis-aspirated contrast disproportionately affect high-frequency words (which tend to be hypo-articulated or reduced). Similarly, Harrington, Kleber, Reubold, and Siddins (2015) found that the German tense-lax vowel distinction is more merged when de-accented (also a context for hypoarticulation), and thus more likely to induce misperception. Examination of different speech styles more likely to show hypoarticulation might therefore reveal an incipient change in Hunchun production. In either of these cases, given that we do see change in other dialects with this methodology, the change is clearly at a much earlier stage in Hunchun.

Another possibility is that there is no sound change in production yet, but that there will be, under the idea that sound change in perception precedes that of production (Ohala, 1993; Hyman, 1976); i.e., a change in production comes about as a result of listeners misparsing intrinsic phonetic effects of f0 as extrinsic, eventually phonologizing them. This account predicts that a sound change in Hunchun should occur in the future.

However, it is also possible that there is, in fact, no current, incipient, or impending change in Hunchun production norms. Rather, the age-related difference in Hunchun perception could be driven by exposure to change in other communities (including Seoul). Even in the absence of a community-level sound change in production, perception norms could shift because of changes in other communities. Younger listeners might show a greater effect because they are disproportionately exposed to speech from younger members of the other community. However, even assuming that the exposure is qualitatively equivalent for younger and older listeners, an asymmetry might be expected because the older listeners’ representations are more influenced by their long-term experience with the ‘older’ version of the other community’s speech norms. If it is indeed the case that there are differences in perception due to exposure to dialectal variation, then a more general implication is that people who are exposed to many different accents might have different perceptual patterns than those who have not, without necessarily showing any difference in production patterns.

The current results do not allow us to distinguish between these possibilities: 1) there is an incipient or forthcoming sound change in production, or 2) production is stable, but perception is shifting. Either of these would have interesting implications. The first case would provide strong evidence that perception is leading, and perhaps predictive of, this sort of sound change. This could be tested by examining the Hunchun population across a longer period of time (or examining younger listeners), but the directionality of perception versus production in sound change could also be tested more generally via comparison of perception and production data from other dialects of Korean which are known to be undergoing sound change. In the second case, if the ‘change’ in perception is being driven by Seoul, it is worth noting that it also applies even in the Local condition, in which case exposure to dialectal variability is also affecting perception of listeners’ ‘own’ dialect.

Our current results, therefore, bring up more questions than answers about the time-course of sound change in perception. However, we hope that the asymmetry between perception and production shown by Hunchun listeners in this work points to the need for the provenance behind this mismatch to be tested empirically.

4.5. Use of social expectations in speech perception

The effect of the apparent dialect of the talker on listeners’ responses, with the Seoul talker eliciting more use of f0 in the lenis-aspirated contrast than the talker said to be from the local community, is consistent with previous findings demonstrating an influence of top-down information on perception, reinforcing previous findings of talker effects on speech perception (e.g., Niedzielski, 1999; Strand & Johnson, 1996; Hay et al., 2006b).

Most work looking at the effect of perceived social information has focused on vowels, presumably because vowels are a primary source of dialectal variation in English. There has also been some work on fricatives (Dufour et al., 2014; cf. Chang, 2017). Moreover, previous work has primarily focused on domains where there is meta-linguistic awareness of the variable in question. Many phonological and phonetic factors are familiar to Korean speakers as dialectal markers, including vowel quality, sibilant place of articulation, intonation, and speech rate (e.g., Lee, 2009; Jeon, 2011; Yang, 2013); however, the phonetic realization of stop contrasts has not been mentioned as a perceptually salient dialectal difference. Therefore, the use of talker information found in this work shows that listeners can make use of their exposure to dialectal variation even when the variation is below the level of conscious awareness.

The Dialect Label effect was only significant for younger listeners (although it was numerically in the same direction for older listeners). This age-based discrepancy cannot be attributed to less exposure to the dialectal variation by older speakers, since they have as much (self-reported) exposure to Seoul Korean as younger listeners. However, it is possible that the exposure is qualitatively different: Older speakers may be exposed to different types of speakers in the present, or their input may be influenced by longer experience with ‘older’ versions of Seoul Korean. Another possibility is that younger listeners are more sensitive to dialectal variability. In either case, this finding stands in contrast to the one other study of which we are aware that examines the interaction of listener age and use of social information: Drager (2011) found that older, not younger, listeners were more sensitive to a community-level change in progress. Together, these findings suggest that age-based differences in use of social information in speech perception may be an interesting avenue for future work to explore.

Although any talker-level differences need to be interpreted with caution, we discuss one here: an overall greater use of f0 for the female than the male talker. We chose age-gender pairings that we hoped would represent extremes of a sound change in progress: a younger female (a group that might be expected to lead change and represent the most innovative production patterns) and an older male (who might be expected to represent a more conservative speech style). Listeners’ behavior was in accordance with our prediction of more use of f0 for speakers who would be expected to be at more advanced stages of the sound change. Therefore, we also interpret this as suggestive of listeners’ use of their expectations about the speech of talkers based on their perceived age and/or gender. However, given the confounds inherent in using two different talkers, there are alternative possible explanations for this finding. Differences could also result from other unpredicted social expectations, and acoustic properties of the two voices could also lead to perceptual differences without having to reference social information at all. A study with balanced gender/age and multiple talkers from each age group is necessary to tease apart these factors more definitively.

4.6. Conclusion

This work explored the joint role of listener-level factors (listeners’ own demographic information) and talker-level factors (listeners’ expectations given social characteristics of the talker) on speech perception. We documented dialectal and age-based variation in how Korean listeners from two dialects used VOT, f0, and vocalic information to classify their native three-way stop contrast, the realization of which varies substantially across dialects and ages. While most group-level perception differences corresponded with dialectal differences seen in production, there was an interesting age-related asymmetry, an apparent change in perception in one of the dialects (Hunchun) where there is no evidence of a corresponding change in production. We also showed that listeners’ perception differs across talkers, consistent with the hypothesis that younger listeners used top-down information about dialectal variation to inform their perception of stops. This highlights the importance of considering listener expectations, as well as both the social and acoustic characteristics of the talker, when interpreting differences in performance on experimental tasks. We hope this will spur more specific investigations of how listener-level norms, talker-level expectations, and their interaction with listeners’ experience shape perception.