More foreign-accented but more comprehensible: Attrition and amelioration of L1 speech in proficient L2 learners

2.1 Speakers

The J-AESOP corpus contains speech data of 183 L1 Japanese learners of L2 English as well as 20 L1 English learners of L2 Japanese. The current study focuses on 154 L1 Japanese speakers (94 female, 60 male) who had never lived overseas in a non-English-speaking country (e.g., Spain) in order to control for the potential influence of a third language (L3). The speakers were all undergraduate or graduate students at universities in and around Tokyo, Japan, between the ages of 18 and 38 (mean = 20.04, standard deviation = 2.10). Approximately two-thirds of them (n = 99; 50 female, 49 male) had spent their entire lives in Japan, with most of them having studied English for six years (age 13-18) as part of their compulsory education in secondary schools. They had also received some English instruction in college, the quality and quantity of which varied depending on the courses in which they were enrolled. The remaining third (n = 55; 44 female, 11 male) had some experience of living in an English-speaking country (32 in the US, 10 in the UK, nine in Australia, eight in Canada, three in New Zealand, with some having lived in more than one country), with the length of residence (LOR) ranging from one month to 11 years (mean = 27.45 months, standard deviation = 32.46 months), but none of them were considered simultaneous bilinguals. Their language learning background while in Japan was similar to that described above, except that some speakers attended international schools as part of their secondary education.

2.2 Materials

The speech materials to be rated are taken from Tasks 6_01 and 6_02 of the J-AESOP corpus, where the speakers read aloud the English and Japanese versions of “The North Wind and the Sun” (International Phonetic Association, 1999) shown in Table 1 and Table 2, respectively. The trisecting of the text in Table 1 is for rating purposes only (see Section 2.4.1), and the speakers read the passage as a whole.

2.3 Raters

2.3.1 English raters

Four phonetically trained native American English listeners were recruited to rate the audio samples of Task 6_01 as part of a previous research project (Konishi, 2022). All of them had a graduate degree in phonetics or a related field and were familiar with the L2 English pronunciation of L1 Japanese speakers. Such experienced raters are known to be able to consistently rate various aspects of speech, including segmental accuracy, word stress, intonation, and rhythm (Saito et al., 2017).

2.3.2 Japanese raters

Twenty-six native Japanese listeners (22 female, 4 male) were recruited to rate the audio samples of Task 6_02 for the current study (this sample includes the 10 listeners in Yazawa et al. (2023b)). They were all undergraduate or graduate students at the University of Tsukuba, Ibaraki, Japan, between the ages of 19 and 25 (mean = 21.04, standard deviation = 1.93). Twenty of them had never lived outside of Japan, while the remaining six had studied abroad in an English-speaking country (three in the UK, two in the US, and one in Australia) for one to ten months. Although the raters were not as linguistically experienced as the four native American English raters, it has been shown that such listeners can coherently evaluate global accentedness and comprehensibility (Saito et al., 2017).

2.4. Procedure

2.4.1 English rating

The native American English listeners were provided with all the trisected audio samples of Task 6_01 for self-paced rating. They were instructed to listen to each trisection and rate the speech according to four criteria (segmental accuracy, prosody, fluency, and nativelikeness) on a 10-point scale each. The current study focuses on the nativelikeness score, where a value of 1 corresponds to “strongly foreign-accented” and a value of 10 corresponds to “free of foreign accent.” Once the rating was completed, the obtained scores were averaged across the trisections per speaker, yielding a single nativelikeness score for each speaker as assessed by one rater. The trisection scheme was to improve the accuracy of the rating by having a rater evaluate the same speaker three times, which also made the resulting scores non-integer (i.e., 1.00, 1.33, 1.67, etc.). Raters received monetary compensation for their time and effort. More details of the English rating can be found in Konishi (2022).

2.4.2 Japanese rating

The native Japanese listeners rated the audio samples of Task 6_02 online via Gorilla Experiment Builder (Anwyl-Irvine et al., 2020). They signed a written consent form and completed a language background questionnaire prior to participation. After a brief tutorial and a practice session, the raters listened to a speaker’s audio sample presented in random order. The use of headphones or earphones was encouraged, but not required, since the use of headphones is unlikely to have any substantial effect on the results of most online perception experiments (Sanker, 2023). The raters then indicated the perceived impression of global accentedness and comprehensibility using two horizontal sliders. The slider for accentedness read “strong foreign accent” on one end and “no foreign accent” on the other, and the slider for comprehensibility read “difficult to understand” on one end and “easy to understand” on the other, in Japanese. The slider values were internally coded from 0 to 100 with integer increments, with a larger value corresponding to a higher level of perceived accentedness or comprehensibility. A pause screen was displayed after each trial to allow the raters to take a short break at any time during the rating. All raters completed their ratings within one week and received monetary compensation for their time and effort. The language background of the speakers was disclosed only after they completed the task.

The above procedural design, including the consent form and the questionnaire, was reviewed and approved by the Research Ethics Committee of the Institutes of Humanities and Social Sciences, University of Tsukuba on May 17, 2022 (approval number 2022-3). All procedures were conducted in accordance with the ethical standards of the Helsinki Declaration.

2.5 Results and analysis

2.5.1 English scores

Figure 1 shows the mean L2 English nativelikeness scores of all 154 L1 Japanese speakers (averaged across four native American English raters), conditioned by their presence and absence of previous residence in an English-speaking country. The intraclass correlation coefficient (ICC) using a two-way random-effects model for consistency (Koo & Li, 2016) was 0.892, indicating a very high level of inter-rater consistency. It can be seen that the scores are fairly evenly distributed, ranging from nearly 1 (“strongly foreign-accented”) to 10 (“free of foreign accent”). While speakers with overseas experience (i.e., LOR > 0) tend to have higher scores than those without (i.e., LOR = 0), there is a significant overlap in L2 nativelikeness score between the two groups (i.e., LOR > 0 vs. LOR = 0). Therefore, by-group investigation of the data would help us to disentangle the potential effect of L2 proficiency (as represented by nativelikeness score) from that of L1 disuse associated with overseas experience (as represented by LOR).

Figure 1. Mean English nativelikeness scores of all speakers, conditioned by their length of residence (LOR) abroad.

2.5.2 Japanese scores

Figure 2 shows the mean L1 Japanese accentedness and comprehensibility scores of the same 154 speakers (averaged across 26 native Japanese raters), conditioned by their previous experience of overseas residence. The ICC was 0.279 for accentedness and 0.271 for comprehensibility, indicating a weak but acceptable level of inter-rater consistency (Fleiss & Cohen, 1973). The confidence ellipses for L1 accentedness scores indicate that most speakers were not perceived as very accented, although some speakers, those with overseas experience (i.e., LOR > 0) in particular, were judged to be moderately accented. The ellipses for L1 comprehensibility scores indicate that most speakers were perceived as generally comprehensible, with a few speakers without overseas experience (i.e., LOR = 0) being apparent outliers. The two types of scores are negatively correlated (Pearson’s r = -0.62), suggesting that more accented speech tended to be perceived as less comprehensible. However, it is still possible that these score types are differentially related to L2 experience and proficiency, as will be illustrated below.

Figure 2. Mean Japanese accentedness and comprehensibility scores of all speakers, conditioned by their LOR abroad.

2.5.3 Relationship between English and Japanese scores

Figure 3 shows the relationship between the L2 English and L1 Japanese scores, again conditioned by the speakers’ overseas experience. Regarding the L1 accentedness score (left side of the figure), distinct patterns can be identified between the two LOR groups. For speakers with LOR > 0, a higher L1 accentedness score seems to be associated with a slightly higher L2 nativelikeness score, suggesting that more proficient L2 speakers tended to be perceived as more foreign-accented in their L1 speech. Given that there was a positive correlation between LOR (in months) and L2 nativelikeness score for these speakers (Pearson’s r = 0.59), this may also suggest that the longer they had stayed in an English-speaking country, the more accented their Japanese was judged to be. This explanation is further supported by the opposite pattern observed for speakers with LOR = 0, where a higher L1 accentedness score seems to be associated with a lower L2 nativelikeness score. Taken together, these results suggest that the experience of living overseas, rather than L2 proficiency per se, contributes to perceived L1 accentedness. In contrast, regarding L1 comprehensibility (right side of the figure), both LOR groups show the same tendency where a higher L1 comprehensibility score leads to a higher L2 nativelikeness score. This suggests that more proficient L2 speakers were generally perceived as more comprehensible in their L1 speech, regardless of prior experience abroad.

Figure 3. Smoothed means of L1 Japanese and L2 English scores, conditioned by speakers’ LOR abroad.

To test whether the above observations hold at statistically significant levels, a cumulative link mixed model (CLMM) was fitted to the response data per speaker group (i.e., LOR = 0 or LOR > 0), using the ordinal package (Christensen, 2023) in R (R Core Team, 2024). The structure of the model was as follows:

To further examine the effect of LOR on L1 accentedness scores, we also fitted the following linear mixed effects (LME) model, using the lme4 (Bates et al., 2015) and lmerTest (Kuznetsova et al., 2017) packages in R:

3.1 General procedure

3.1.1 Materials

The audio recordings of Tasks 6_01 and 6_02 in the J-AESOP corpus (see Section 2.2) were examined. The speech samples were annotated in Praat TextGrid format (Boersma & Weenink, 2024), first by automatic forced alignment tools (HTK (Young et al., 2006) for Task 6_01 and Julius (Lee & Kawahara, 2019) for Task 6_02) and then manually modified by trained phoneticians in the J-AESOP team. Annotators also marked segment-level phonological events such as vowel devoicing (e.g., /kitakaze/ ‘North Wind’ ➔ [ki̥takaze]) and lengthening (e.g., /taijou/ ‘Sun’ ➔ [taijoː]) in Japanese, as well as word-level speech events such as substitution (e.g., misreading cloak as coat), repetition (e.g., Then the Sun … Then the Sun shone out warmly), and insertion (e.g., wrapped around in a warm cloak) by assigning ‘tags’ to the relevant words (underlined above).

3.1.2 Data retrieval and acoustic measurement

Based on the annotations, a total of 34963 Japanese vowels (/i/ = 6504, /e/ = 3598, /a/ = 15713, /o/ = 6504, /u/ = 2444) produced by the 154 L1 Japanese speakers were retrieved for acoustic analysis. Their production of 10775 English monophthongal vowels (/i/ = 1591, /ɪ/ = 3196, /ɛ/ = 911, /æ/ = 1451, /ʌ/ = 1094, /ɑ/ = 1095, /u/ = 893, /ʊ/ = 544) was also retrieved. Devoiced and lengthened vowels, as well as vowels in tagged words, were excluded.

For each vowel interval, the mean F1 and F2 frequencies were measured using Praat. The built-in Burg algorithm was used for formant estimation, with the formant ceiling setting at 5000 Hz for male speakers and 5500 Hz for female speakers. The obtained F1 and F2 values were then Z-transformed per speaker (Lobanov, 1971), which effectively eliminates spectral variations caused by physiological differences while preserving phonological and cross-linguistic contrasts (Adank et al., 2004). The normalization was performed across Tasks 6_01 and 6_02 so that the formants could be directly compared across L1 Japanese and L2 English speech.

3.2 Acoustic correlates of accentedness

We begin our analysis by examining the potential acoustic correlates of perceived L1 foreign accent. Given the finding in Section 2 that L1 foreign accented was typical of L2 speakers with overseas experience, here we focus on the 55 speakers with LOR > 0. We first hypothesized that the formant values of L1 vowels in foreign-accented speech are dislocated from those of non- or less-accented speech, presumably due to the assimilation or dissimilation of L1 vowels to acquired L2 vowels. To test this, we used the following LME model:

To complement the above results, we also show in Figure 4 the production of L2 English vowels by the same 55 speakers as a function of their L2 nativelikeness scores (cf. Yazawa et al. (2023a)). Each circle shows the mean Z-normalized formant values of a 0.50 score range (i.e., 1.25-1.75, 1.75-2.25, etc., as in the bins of Figure 1), with darker shades representing higher scores. The arrows point from lowest through intermediate to highest score ranges based on these means, ignoring apparent outliers. The mean formant values of L1 Japanese vowels are also shown alongside as gray boxes (averaged across the L2 score ranges because the by-range differences in F1 and F2 values were too subtle to plot). Since speakers with higher L2 nativelikeness scores were perceived as having a stronger L1 foreign accent for the current sample, a comparison of Table 6 and Figure 4 can be useful to examine how acquired L2 vowels might affect L1 vowel production.

Figure 4. Japanese and English vowel formants of speakers with LOR > 0 according to L2 nativelikeness scores.

The comparison shows that only some of the results can be straightforwardly explained by L1-L2 segmental assimilation or dissimilation. For example, the results for L1 /e/ in Table 6, where lower F1 and higher F2 values predict a higher L1 accentedness score, can be interpreted as a case of dissimilation from the adjacent L2 vowel /ɛ/, which shows the opposite pattern as a function of L2 nativelikeness score in Figure 4. That is, as speakers become more proficient in the L2 (and thus more accented in the L1), their L1 /e/ production is raised and fronted while that of L2 /ɛ/ is lowered and backed, resulting in an increased distance between the two categories. In contrast, the results for L1 /o/ in Table 6, where higher F1 and lower F2 values predict a higher L1 accentedness score, can be interpreted as a case of assimilation to L2 /ɑ/, which shows the same pattern according to L2 nativelikeness scores in Figure 4. That is, as speakers become more proficient in the L2 (and thus more accented in the L1), the L1 and L2 vowels move to the same direction (similar to what Turner (2023) called “tandem drift”). The results for the other three L1 vowels, however, were more mixed. As for /i/, while the negative estimate for F1 (i.e., raising) in Table 6 may be attributed to its assimilation to L2 tense /i/ (see Figure 4), the negative estimate for F2 (i.e., backing) would contrarily indicate dissimilation from this very L2 category. The same goes for L1 /u/, where the negative F1 estimate (i.e., raising) would indicate assimilation to L2 /u/ or /ʊ/, whereas the negative F2 estimate (i.e., backing) would indicate dissimilation from these two categories. L1 /a/ is a similar case, where the positive F1 estimate (i.e., lowering) would indicate dissimilation from L2 /æ/ or /ʌ/, while the positive F2 estimate (i.e., fronting) would indicate assimilation to these categories. The supposed assimilation-dissimilation patterns are, therefore, inconsistent at best.

Another explanation becomes possible when we shift our focus from individual L1-L2 category contrasts to the whole vowel space. Considering the suggested phonetic changes in Table 6—raising and backing of /i/, raising and fronting of /e/, lowering and fronting of /a/, lowering and backing of /o/, and raising and backing of /u/—one can envision a clockwise chain shift. Judging from Figure 4, such movements of L1 vowels seem to be helpful for maintaining their distance from the developing L2 categories.

3.3 Acoustic correlates of comprehensibility

We now move on to examine the potential acoustic correlates of perceived L1 comprehensibility. Here we investigate all 154 speakers, since more proficient L2 speakers were found to be more comprehensible in their L1 regardless of overseas experience in Section 2. The first acoustic parameter to be investigated is the ‘compactness’ of vowel categories in the F1-F2 acoustic space. Kartushina and Frauenfelder (2014) found that individuals with more compact L1 vowel categories tend to show more accurate L2 vowel production, presumably because the existing L1 categories overlap less with the target L2 vowels. SLM-r (Flege & Bohn, 2021) provides a theoretical explanation of this finding through its “category precision” hypothesis, which posits that individuals with relatively precise L1 categories are better able to discern phonetic differences between an L2 sound and the closest L1 sound, thereby increasing the likelihood of new L2 category formation. If this is also true for the current sample, we can hypothesize that higher L1 comprehensibility in more proficient L2 learners is also due to their more compact or precise production of L1 vowel categories.

Following Kartushina and Frauenfelder (2014), we first calculated the compactness score (CS) per speaker and vowel type, using the following formula:

The calculated five CSs were then added to obtain the global compactness score (CS_G) per speaker:

Figure 5. Vowel ellipses of two male speakers with less (left) and more (right) compact articulation.

Another acoustic parameter we will investigate here is formant dispersion, since compact vowel categories can still overlap with one another if they are not sufficiently dispersed and non-compact vowel categories can still be clearly distinguishable from each other if they are sufficiently dispersed. Thus, we can derive an alternative hypothesis that higher L1 comprehensibility in more proficient L2 learners is due to their more dispersed L1 vowel production.

While there are several metrics for measuring formant dispersion, we used the vowel formant dispersion (VFD) metric of Karlsson and van Doorn (2012). To calculate the VFD, we first need to define the center of the F1-F2 vowel space for each speaker. The coordinate of the midpoint along the vertical (F1) axis is calculated according to:

Figure 6. Vowel vectors of two female speakers with less (left) and more (right) dispersed articulation.

Now that the acoustic parameters are defined, the question is whether individual differences in vowel compactness or dispersion predict the degree of perceived L1 comprehensibility. To test this, we used the following LME model:

To fully support the second hypothesis, however, we must also demonstrate that individuals with more dispersed L1 vowel production are more proficient in the L2. To test this, we fitted the following CLMM:

4.1 Overall summary

The primary purpose of the current study was to investigate the potential effects of L2 learning on perceived accentedness and comprehensibility of L1 speech. To this end, we examined 154 L1 Japanese learners of L2 English in the J-AESOP corpus with varied L2 learning experience and proficiency levels. The rating results revealed that more proficient L2 speakers tended to be perceived as more foreign-accented in the L1, but only if they had lived in an English-speaking country; no such relationship was found for those who had never lived abroad. Subsequent acoustic analyses of vowel production suggested that the degree of perceived foreign accent could not be straightforwardly attributed to either assimilation or dissimilation of individual L1 and L2 categories but rather to a system-wide phonetic change. In contrast, more proficient L2 speakers were consistently perceived as more comprehensible in the L1, regardless of the presence or absence of prior overseas experience. Acoustic analyses suggested that the degree of perceived comprehensibility was associated with the dispersion, rather than compactness, of vowel production, although neither dispersion nor compactness in L1 vowel production predicted L2 proficiency levels.

4.2 Implications

4.2.1 Accentedness

The rating results for accentedness suggest that L2 learners do not necessarily gain an L1 foreign accent as they become proficient in the L2. Rather, strong L1 accentedness seems to be associated with the experience of living in an L2-speaking (or non-L1-speaking) environment for an extended period of time. Therefore, to answer our first research question, a perceptible L1 foreign accent is likely due to L1 disuse rather than L2 interference. It may thus be misleading to refer to this phenomenon as “backward transfer” because the acquired L2 system is not directly affecting the existing L1 system; the term “attrition” would be more appropriate in this sense. It should be noted, however, that the degree of perceived L1 foreign accent for the current sample was less profound than that of the migrant populations in the previous studies (de Leeuw et al., 2010; Kornder, 2022). This may be because our speakers maintained their L1 oral skills through everyday conversations with their parents and siblings while abroad and/or partially reversed L1 attrition after re-immersion into the L1-speaking environment back in Japan. While the LOR information in the corpus does not reflect these possibilities, the effects of relative L1 exposure while abroad and the length of L1 re-immersion in Japanese-English bilingual returnees is extensively discussed in Laméris et al. (2024).

Our acoustic analysis of vowel production also indicated an intriguing pattern of potential L1 phonetic change in proficient L2 speakers with overseas experience: clockwise chain shift in the F1-F2 space. While no previous study has reported such a pattern, our finding would not contradict previous reports of system-wide phonetic change due to drift or attrition (Guion, 2003; Mayr et al., 2012; Turner, 2023). Thus, to answer our second research question, perceived L1 accent seems to reflect a system-wide change in all sound categories, rather than individual assimilation or dissimilation of adjacent categories between the two languages. Caution is warranted, however, because the previous studies observed a systematic lowering or raising of all vowel categories, whereas in the current study the direction of change varied from vowel to vowel (similar to Kornder (2022)). Nevertheless, the motivation for these diverging patterns may be the same, i.e., to maintain a perceptual distinction between the L1 and L2 sounds, as proposed by Guion (2003).

4.2.2 Comprehensibility

A novel finding of this study was that a higher level of L2 proficiency consistently predicted a higher level of L1 comprehensibility (consistent with our previous pilot study (Yazawa et al., 2023b)). This was somewhat surprising given the negative correlation between L1 accentedness and comprehensibility (cf. Figure 2), but perhaps not entirely unexpected since a strong foreign accent does not necessarily reduce perceived comprehensibility for L2 speech (Munro & Derwing, 1995). Thus, to answer our final research question, the influence of L2 learning on perceived L1 comprehensibility can be positive. Logically speaking, there are two possible explanations for the current finding: (i) learners whose L1 speech is comprehensible tend to acquire a high level of L2 proficiency, and (ii) as learners become more proficient in the L2, their L1 speech becomes more comprehensible. The first possibility was theoretically predicted by the “category precision” hypothesis of SLM-r (Flege & Bohn, 2021), as outlined in Section 3.3. However, the acoustic analyses did not support this prediction, since neither compactness nor dispersion of L1 vowel articulation was related to L2 nativelikeness scores (in contrast to Kartushina and Frauenfelder (2014) who found a link between L1 vowel compactness and L2 proficiency). It is then worth noting that the VFD metric did predict the perceived degree of L1 comprehensibility, because if this viable acoustic predictor of L1 comprehensibility does not predict L2 proficiency, then (i) is not very feasible. The alternative possibility is (ii), which, although no current model of L2 speech acquisition offers an explicit explanation for it, seems plausible in nature. If L2 learning involves an effort to clearly articulate speech sounds to make oneself better understood, then learners may extend this skill to their L1 production, thus increasing its comprehensibility (even though the resulting speech may differ from monolingual norms). If such an effect turns out to be true, then this L2-related ‘amelioration’ or ‘enrichment’ of L1 speech would be an important advantage of bilingualism, especially in the current era where comprehensibility is considered more important than nativelikeness.

4.3 Concluding remarks

In this study, we conducted a large-scale cross-sectional investigation of L1 phonetic change in native Japanese speakers with various L2 English learning experience and proficiency levels using the J-AESOP corpus. The overall results seem to suggest two main conclusions. First, perceived L1 foreign accent likely results from L1 disuse rather than L2 interference, where a learner’s L1 pronunciation is shifted from monolingual norms at a system-wide rather than category-specific level. Second, L2 learning can have a positive influence on perceived L1 comprehensibility, rather than individuals whose L1 speech is more comprehensible being better equipped for L2 learning. However, the current investigation is limited in that the corpus data do not inform us of how the learners’ pronunciation has actually changed over time; we are, after all, comparing different speakers from different linguistic backgrounds. What seems necessary to confirm and extend the current research, then, is a longitudinal study of intra-speaker acoustic-phonetic change and its communicative impact, similar to previous studies (e.g., Kornder (2022)) but with a larger number of speakers. To this end, a longitudinal bilingual speech corpus, which is far less common than cross-sectional ones such as J-AESOP, is expected to be useful. While building such a corpus would be labor-intensive and time-consuming, we believe that the benefits outweigh the effort. A preliminary project towards this goal has been initiated, and we look forward to sharing what it brings about in the future.