A first factor is the age at which a person acquires the L2. If the acquisition of the L2 starts later, it is often more difficult to hide a foreign accent, that is, the L1 shapes L2 speech more than if the L2 acquisition begins at an earlier age (see Flege, 1992, 1995; Munro et al., 1996; for discussion, see also Birdsong, 2006). For instance, German-speaking children (L1) are more sensitive to the differences between the English /æ/ and a near German vowel (/ɛ/) than German-speaking adults (L1; see Weiher, 1975; Butcher, 1976 both cited in Flege, 1995).

A second factor is the amount of experience a person has in the L2. Usually, individuals with more experience in the L2 sound more native-like (see Best and Strange, 1992; Flege, 1995; Baker and Trofimovich, 2005). For example, Bohn and Flege (1992) show that German speakers of English who had spent on average about 7 to 8 months in an English-speaking environment barely distinguished between /ɛ/ and /æ/ in English, while German speakers of English who had spent on average 7.5 years in an English-speaking environment clearly separated between the two vowels (for related results, see Kautzsch, 2010; Simon and D’Hulster, 2012). The German speakers with longer exposure to English were more similar to English native speakers in that they produced a greater height difference between /ɛ/ and /æ/ in English than the German speakers with less experience in English. Also, the /æ/ of the German speakers with a low amount of experience in English was articulated higher than the /æ/ of German speakers with more experience in English and English native speakers.

Although the two factors age and amount of experience play a crucial role in the acquisition of an L2 in general and with regard to how much the L1 shapes the L2, we see that even highly advanced language users still differ from native speakers. For instance, Bohn and Flege (1992) show that even the German group with a long experience in L2 English differs in some ways from the group of English native speakers (for more discussion, see also Hickey, 2019). That is, first, in terms of spectral properties, the English /ɛ/ produced by the German speakers with a lot of experience in English was not different from their German /ɛ/ and /ɛː/, although at least some research suggests subtle spectral differences between the English and the German version of the vowel (see Bohn and Flege, 1990). Second, the German speakers with much experience in English differed from the English native speakers with regard to vowel duration, too. Both German groups, the inexperienced and the experienced one, pronounced English /æ/ clearly shorter than native speakers of English, and the German speakers with a high level of experience in English realized English /ɛ/ shorter than the English native speakers. These results illustrate the complexity of the issue: on the one hand, the durations of the two English vowels were not native-like even in the German group with a high level of experience in English; on the other hand, this group expressed the expected duration difference between the two English vowels and was, so to speak, native-like with respect to the duration contrast. A further finding from Bohn and Flege (1992) indicated that higher experience in an L2 does not necessarily lead to sounding more native-like. In an intelligibility test, they found that both words with /ɛ/ and words with /æ/ were less intelligible if pronounced by any of the two German groups (inexperienced or experienced in English) in comparison to the group of English native speakers (see also Flege et al., 1997).

A third factor is the degree of similarity between specific phenomena in the L1 and L2. It has been stated that a higher degree of similarity between phenomena in the L1 and L2 typically results in greater L1 influence on the L2 and more difficulties in correctly acquiring the L2 sound (see Flege, 1987, 1995; for discussion, see also Baker and Trofimovich, 2005). Bohn and Flege (1990) keep the adjective “similar” apart from “new” and “identical”. For instance, they consider /m/ to be identical in German and English, /u/ to be similar, and /æ/ to be a new sound for German learners of English. It goes without saying, however, that even a new L2 sound, as English /æ/, can be articulatorily and acoustically close to an L1 category. Since both English /ɛ/ and /æ/ as well as German /ɛ/ and /ɛː/ are front, open-mid, and unrounded vowels, they are acoustically and articulatorily quite similar. In an L2, individuals might fail to detect these fine acoustic differences between two sounds of the L2 or between sounds in the L1 and those in the L2, which can cause the assimilation of actually distinct sounds into a single category in the L2 (see Flege, 1995; but see also discussion above). Hence, the creation of a category for the L2 sound arises only if the respective differences and fine details are processed, which is, in turn, more likely if L1 and L2 sounds differ more clearly (see Flege, 1995).

To reach an intermediate summary, we can say that whether and to what extent the L1 affects the acquisition of the L2 has been widely examined in the literature, and several factors have been taken into consideration. However, we feel that another factor has been mostly neglected, namely the role of sociolinguistic aspects of an individual’s L1 in the context of L2 speech acquisition. It is the aim of the present paper to investigate this issue. Precisely, we aim at investigating whether the degree to which German native speakers distinguish between the two front, unrounded vowels /eː/ (e.g., dehnen, [ˈdeːnən], “to stretch”) and /ɛː/ (e.g., Dänen, [ˈdɛːnən], “the Danish”) in their L1 has an impact on how well these individuals identify /æ/ and discriminate between the two English front, unrounded vowels /ɛ/ (e.g., pen, [pɛn]) and /æ/ (e.g., pan, [pæn]). We therefore intend to expand previous research on how a high degree of similarity between a phenomenon in the L1 and one in the L2 affects L2 speech acquisition in detail, by considering sociophonetic variation across different native speakers of German.

Fifty-one native speakers of German from northern Germany participated in the experiment [33 female, 18 male; mean age: 24.6 years (SD: 3.4 years; range: 19–35 years)].³ They had an academic background and English was their second language. They did not declare a speech disorder or hearing impairment, did not use a hearing aid, and could see clearly.

All participants were well-educated and had taken English as one of the central school subjects for at least 8 years. That is, all of them had a solid knowledge and command of English. Despite this similarity between our participants, they varied to some extent with regard to their level of English. We grouped our participants into three proficiency groups (“low,” “mid,” and “high” competence in English). Categorizing participants into the three groups was done in the following way. The first author of the paper and a native speaker of English listened to sound files of items, which were recorded in a short production experiment prior to the perception experiments. They examined the pronunciation in terms of nativelikeness, and assigned the participants to the groups “low,” “mid,” and “high.” If the first author and the native speaker of English did not agree, the case was discussed until agreement was reached. Our evaluation of the participants’ level of English is thus based on their pronunciation only. The participants’ level of English was later entered as a control variable into the statistical models.

A key element of our work is whether and to what extent our participants distinguish between the German vowels /eː/ and /ɛː/. In order to obtain an objective reflection of this, we relied on the data from the production experiment reported in Frank (in preparation). In that study, our 51 native speakers of German read a list of 174 words that included 19 items containing /eː/ (e.g., dehnen, [ˈdeːnən], “to stretch”) and 29 items containing /ɛː/ (e.g., Dänen, [ˈdɛːnən], “the Danish”) in Standard German. Frank (in preparation) analyzed F1 and F2, which, in turn, were then used to calculate a Pillai score (see Nycz and Hall-Lew, 2013). For each speaker, a Pillai score was calculated over all tokens. A score of “0” means that a speaker does not distinguish between the two vowels at all (= /eː/-/ɛː/ merger). A Pillai score of “1,” in turn, means that a person fully distinguishes the two vowels /eː/ and /ɛː/ (= /eː/-/ɛː/ maintainer). The German Pillai score entered our statistical models as a continuous fixed effect (see below). Remember from section 1.2 that, even though merging the two German vowels is quite common in northern Germany (where our participants are from), we still observe variation even within a region.⁴

Two minimal pairs were selected for this experiment, namely pan/pen and paddle/pedal. The four words were spoken and recorded by a 24-year-old male US American from California in a silent room. The four sound files were normalized (70 dB) and subsequently used as the basis for the following manipulations. We created a spectral continuum with 11 steps from an extreme /æ/ (Step 1) to an extreme /ɛ/ (Step 11) based on a Praat script (Winn, 2019; Boersma and Weenink, 2021). Steps 3 and 9 represented the spectral information of the original /æ/ and /ɛ/, respectively. From Steps 3 and 9, we created two steps below (Steps 1 and 2 and Steps 7 and 8) and two steps above (Steps 4 and 5 and Steps 10 and 11). Step 6 was the middle of the continuum. The detailed spectral information is summarized in Table 2.

In addition, three vowel duration categories, short, middle, and long, were used. We relied on the original duration of the /æ/ words to define the long duration and on the original duration of /ɛ/ to specify the short duration. The middle duration was then positioned between the two. The vowel durations are given in Table 3.

The experiment was conducted in a silent room. Participants sat about 60 cm in front of a computer screen and wore KOSS R-80 over-ear headphones. We used E-Prime 3 (Psychology Software Tools, 2016) for stimulus presentation. In each trial, the participants saw two pictures on the screen, either the image of a pan and the image of a pen or the picture of a paddle and the picture of a pedal. Their task was to indicate (via button press on a regular keyboard) whether the word they had heard represented the first or second image. The position (left side vs. right side of the screen) of the pictures (pan vs. pen and paddle vs. pedal) was counterbalanced across participants. Participants were requested to press the left and right arrow on the keyboard to indicate that they believed they had heard the item on the left or right, respectively.

The two minimal pairs pan/pen and paddle/pedal were tested in separate blocks. Further, we distinguished between three blocks for each minimal pair, one for each vowel duration type (short, middle, long). Hence, the identification task consisted of six distinct blocks (2 minimal pairs × 3 vowel duration types per minimal pair). The three blocks relating to one minimal pair always appeared one after the other in the same order (middle, short, long). All the pan/pen blocks appeared first, all the paddle/pedal blocks afterwards. Each person was tested on 198 trials (2 minimal pairs × 3 vowel duration types × 11 steps of the continuum × 3 times each sound file). Hence, each of the six blocks contained 33 cases (11 steps of the continuum × 3 times each sound file).

The two response variables were the percentage of the selection of the picture representing the /æ/ word (Selection of /æ/) and ReactionTime. We used a binomial logistic regression to examine the first and linear mixed effects models to analyze the second one in R (R Core Team, 2021), using the lme4 (Bates et al., 2015) and lmerTest packages (Kuznetsova et al., 2017).⁵ In the analysis of ReactionTime, we first log transformed (to the base 10) the values (see Winter, 2020). We then checked the dataset for extreme values (statistical outliers), defined as the overall mean plus/minus 2.5 standard deviations (see Loewen and Plonsky, 2016). No extreme values were detected.

Model fitting proceeded in the following way (fit with ML or REML; see Field et al., 2012). On the level of random effects, we started with the structure containing the intercepts for Subject and Item as well as the slope of our central fixed effect, GermanPillai, by Item (see Winter, 2020). Complex random effects structures can lead to a statistically unstable analysis, that is, convergence issues can arise (see Barr et al., 2013; Matuschek et al., 2017; Cohen and Kang, 2018). If this was the case, we first changed the optimizer to “bobyqa” (see Winter, 2020). If this was not enough, this initial structure was reduced manually and checked again, first without and, if necessary, with the different optimizer (bobyqa). The initial model was reduced step by step until convergence was reached.

On the level of the fixed effects structure, we included the German Pillai score (GermanPillai), Step (1–11, factor-transformed, see Winter, 2020), Duration (short, middle, long), LevelOfEnglish (low, mid, high), and the three interactions GermanPillai × Step, GermanPillai × Duration, and GermanPillai × LevelOfEnglish in the initial model. If a fixed effect was not significant, it was removed from the model. The new model, without the excluded fixed effect, was then checked again. If significance was detected, the respective case was further investigated using three criteria mentioned in Plag et al. (2017). Only if all of the three criteria went in the expected direction, we considered the case to be significant and retained the fixed effect in the model. The first criterion implied that the t statistics of the respective fixed effect had to be lower than −2 or higher than 2.⁶ Second, the model with the fixed effect had to have a better fit than the model without it; this would be reflected in a significant difference between the two models, as detected in an ANOVA. Third, the Akaike Information Criterion (AIC) had to be smaller for the model with, in contrast to the model without, the respective fixed effect (see also Pinheiro and Bates, 2000; Wu, 2010).

Finally, for the categorical fixed effects, planned pairwise comparisons (Tukey tests) were conducted using the emmeans package (Lenth, 2020) in R to shed light on the comparisons that we cannot see in the model, that is, in Tables 4, 5. This method has been used in the context of different kinds of regression analyses, such as logistic / ordinal regression (see Montrul et al., 2019; Kim and Yoon, 2020).

We start presenting the results for Selection of /æ/. The most important results are the two significant interactions GermanPillai × Step and GermanPillai × Duration (see Figures 1, 2). The statistical details are presented in Tables 4, 6.

We see two significant interactions, GermanPillai × Step and GermanPillai × Duration. Let us start with the first one. In Figure 1, we see that the vertical sequence of the profiles matches the sequence of steps, hence the materials and manipulations worked out as intended. On the horizontal level, we observe that the manipulations of the acoustic signal had a considerably smaller effect on the response behavior of participants at the left than at the right margin. At the left margin, where we find /eː/-/ɛː/ mergers, differences in vowel quality bring about a change in identification of /æ/ between roughly 30% and 60%–70%. That is, /eː/-/ɛː/ mergers identified extreme renditions of /æ/ as /æ/ in only 60–70%. Likewise, /eː/-/ɛː/ mergers identified extreme renditions of /ɛ/ as /ɛ/ in only 70%. At the right margin, where we find /eː/-/ɛː/ maintainers, differences in vowel quality bring about a change in identification of /æ/ between roughly 20% and 90%. That is, /eː/-/ɛː/ maintainers were much more sensitive to the manipulations, they identified extreme renditions of /æ/ as /æ/ in about 90%. Likewise, /eː/-/ɛː/ maintainers identified extreme renditions of /ɛ/ as /ɛ/ in about 80%.

In Figure 2, the second interaction is plotted. We see again that the vertical sequence of the profiles matches the sequence of the three duration types (from long to middle to short), indicating that the materials and manipulations worked out as intended. On the horizontal level, we observe again that the manipulations of the signal had a considerably smaller effect on the response behavior of participants at the left than at the right margin. At the left margin, differences in duration bring about a change in identification of /æ/ between roughly 55% and 60%. That is, /eː/-/ɛː/ mergers identified long stimuli as /æ/ in about 60%. Short stimuli, in turn, were identified as /ɛ/ in about 45%. At the right margin, however, differences in duration bring about a change in identification of /æ/ between roughly 40 and 60%. Long stimuli were identified as /æ/ in about 60% while short stimuli were identified as /ɛ/ in about 60%. Overall, while for the /eː/-/ɛː/ mergers the percentage of /æ/ selection was rather similar for short and long vowels, /eː/-/ɛː/ maintainers clearly selected /æ/ more often for long than for short vowels.

We now turn to the analysis of ReactionTime, where we found a significant interaction of GermanPillai × Duration (see Figure 3; Tables 5, 7).

In the analysis of ReactionTime, there is a main effect of Step. Participants reacted more quickly to Step 1 than to Step 6. A Tukey post hoc test revealed further that participants responded faster to Step 2 than to Step 6 (estimate = −0.04, SE = 0.01, df = Inf, z ratio = −4.24, p < 0.01), to Step 2 than to Step 7 (estimate = −0.03, SE = 0.01, df = Inf, z ratio = −3.30, p < 0.05), and to Step 2 than to Step 8 (estimate = −0.03, SE = 0.01, df = Inf, z ratio = −3.57, p < 0.05).

Furthermore, we also detected a significant interaction, namely GermanPillai × Duration. In Figure 3, we see, first, that /eː/-/ɛː/ maintainers (at the right margin) were generally quicker than /eː/-/ɛː/ mergers (at the left margin). Second, in comparison to long and short stimuli, /eː/-/ɛː/ mergers needed more time to deal with the intermediate category. Although /eː/-/ɛː/ maintainers needed more time to react to the intermediate category, too, the difference between the intermediate category and the long and short stimuli was smaller for this group of speakers.

The participants were those from Experiment I.

The materials/sound files were those from Experiment I.

The experiment was conducted in a silent room. Participants sat about 60 cm in front of a computer screen and wore KOSS R-80 over-ear headphones. We used E-Prime 3 for stimulus presentation. In each trial, participants heard two sound files and indicated via button press, using the left and right arrow on a regular keyboard, whether they believed they were identical or different. We counterbalanced the association of arrow and identical/different across participants. As in Experiment I, both the two minimal pairs and the three vowel duration types appeared in different blocks. The spectral distance between the two sound files of a trial was 3, 2, 1, or 0 steps. That is, for instance, if the first sound file contained Step 3 and the second sound file Step 4, the distance between the two was 1 step. The interstimulus interval (ISI) was 1 s.⁷ Each person was tested on 390 trials (2 minimal pairs × 3 vowel duration types × 65 trials per block).

The two response variables were Accuracy and ReactionTime. We used binomial logistic regression to examine the first and linear mixed effects models to analyze the second one in R, using the lme4 and lmerTest packages. In the analysis of ReactionTime, only correct answers were considered (40% of the answers), following common practice (see Jiang, 2012, p. 68–69). The reaction time data was log transformed (to the base 10), before we additionally checked it for extreme values. No extreme value was detected (see also section 2.1.4).

On the level of the random effects structure, we proceeded in the way described in section 2.1.4. On the level of the fixed effects structure, we included GermanPillai, StepDistance (0, 1, 2, 3; factor-transformed), Duration (short, middle, long), LevelOfEnglish (low, mid, high), and the three interactions GermanPillai × StepDistance, GermanPillai × Duration, and GermanPillai × LevelOfEnglish in the initial model. The model was reduced step by step and significances were checked using the technique outlined in section 2.1.4. Planned pairwise comparisons were used as outlined in section 2.1.4.

We start looking at accuracy. Here, the interaction GermanPillai × StepDistance reached significance (see Figure 4). The statistical details are presented in Tables 8, 9.

In the analysis of Accuracy, we found a main effect of Duration, indicating that short vowels were responded to less accurately than long ones. A Tukey post hoc test further showed that middle vowels were responded to more accurately than short vowels (estimate = 0.16, SE = 0.04, df = Inf, z ratio = 3.82, p < 0.001).

Further, we detected a significant interaction GermanPillai × StepDistance. We see in Figure 4 that, if the StepDistance was 0, that is, if the sound files a person was exposed to in a trial were identical, and in contrast to if the StepDistance was 1, 2, or 3, /eː/-/ɛː/ maintainers (at the right margin) responded slightly more accurately than /eː/-/ɛː/ mergers. For StepDistances of 1 and 2, however, /eː/-/ɛː/ mergers (at the left margin) reacted slightly more accurately than /eː/-/ɛː/ maintainers.

We now turn to ReactionTime (see Tables 10, 11). The analysis of ReactionTime revealed two main effects, one of Duration and one of StepDistance, but no significant interaction. First, the response time was longer when the vowel had a middle duration in comparison to when it had a long duration. Also, reaction times were shorter for short than for long vowels. A Tukey post hoc test indicated that the response time was longer for middle than for short vowel durations (estimate = 0.03, SE = 0.00, df = Inf, z ratio = 11.98, p < 0.001). Second, participants needed more time to respond when the StepDistance between the two sound files of a trial was 1, 2, or 3, in comparison to when it was 0. A Tukey post hoc test showed that the response times were longer for a StepDistance of 1, compared to StepDistances of 2 and 3 (1 vs. 2: estimate = 0.01, SE = 0.00, df = Inf, z ratio = 2.94, p < 0.05; 1 vs. 3: estimate = 0.03, SE = 0.00, df = Inf, z ratio = 8.28, p < 0.001). Also, participants responded more slowly for StepDistance 2 than for StepDistance 3 (estimate = 0.02, SE = 0.00, df = Inf, z ratio = 5.95, p < 0.001).