Cross-Linguistic Influences on the Production of Third Language German Vowel Length Contrasts by Cantonese

Abstract

This study examined the read speech of vowel length contrasts produced by Cantonese–English–German trilinguals, comparing their performance with that of Mandarin–English–German trilinguals, Cantonese–English bilinguals, native English speakers, and native German speakers. Acoustic and statistical analyses of vowel quality and duration across the first language (L1), second language (L2), and third language (L3) yielded several key findings. First, Cantonese-speaking trilinguals were more nativelike than the Mandarin-speaking trilinguals in both the L2 and L3, indicating that the L1 exerts a sustained influence across the multilingual system. Second, Cantonese–English–German trilinguals differed from Cantonese–English bilinguals in the L2 but not the L1, suggesting that reverse transfer from the L3 more strongly affects the L2 than the L1. Finally, individuals who produced larger quality contrasts in L2 vowel length distinctions also demonstrated greater quality contrasts in comparable L3 vowels, whereas those who produced larger duration contrasts in the L1 exhibited reduced duration contrasts in analogous L3 vowels, indicating that L1–L3 and L2–L3 bidirectional interactions emerge among phonetically similar vowels. The study highlights the dynamicity of the multilingual phonological system.

Keywords

Cross-linguistic influence second language acquisition multilingualism vowel speech production

1 Introduction

Vowel length is a phonological feature used by some languages and is realized phonetically through vowel duration and/or quality. In this study, a distinction is maintained between the terms “vowel length” and “vowel duration.” The former refers to the abstract phonological contrast (i.e., the “long/short” categories), whereas the latter refers to the phonetic, temporal measurement of a vowel in milliseconds. Languages differ in the phonological status and phonetic realization of vowel length contrasts. For instance, English features a long tense vowel /iː/ and a short lax vowel /ɪ/, distinguished by both duration and quality, whereas Mandarin Chinese does not make a long/short vowel distinction (Chao, 1968). This mismatch has inspired numerous studies exploring how learners acquire second language (L2) vowel length, particularly those whose first language (L1) lacks this contrast (Bohn, 1995; Y. Chen, 2006; McAllister et al., 2002). A common finding in these studies is that L2 learners often struggle to produce vowel length accurately. In contrast, vowel length features have not been approached from the perspective of third language (L3) acquisition. Specifically, it remains unclear whether the acquisition of vowel length in an L3 is subject to the same constraints as in L2 acquisition. Importantly, unlike L2 learners, L3 learners have two background languages to draw upon. L3 morpho-syntactic theories make different predictions regarding the sources and mechanisms of cross-linguistic influence (CLI) (Bardel & Falk, 2007; Rothman, 2011; Westergaard & Rodina, 2017). Examining the acquisition of L3 vowel length features, especially when these features differ in their realization between L1 and L2, would open a window to understanding CLIs in third language phonological acquisition (TLA). This study explored the acoustic-phonetic features of L1, L2, and L3 long and short vowels produced by trilingual speakers of L1 Cantonese, L2 English, and L3 German. In addition, it compared the vowel systems of this target trilingual group with those of L1 English speakers, L1 German speakers, and L1 Cantonese–L2 English bilinguals to uncover how the complex interplay of languages shapes the trilingual phonological system.

1.1 Vowel Length Features in German, English and Cantonese

Figure 1 lists the monophthongs of Northern Standard German, Standard Southern British English (SSBE) and Hong Kong Cantonese. Since this study involves a control group of Mandarin–English–German trilinguals, the Mandarin vowel inventory is also provided.

Figure 1.

Monophthongs in German, English, Cantonese, and Mandarin (redrawn based on Kohler, 1999; Lee & Zee, 2003; Roach, 2004; Zee, 1999).

Among these languages, German has the most systematic vowel length contrast. Phonologically, there are seven phonemic long German vowels neatly paired with the other seven phonemic short vowels (i.e., /iː/-/ɪ/, /yː/-/ʏ/, /uː/-/ʊ/, /eː/-/ɛ/, /øː/-/œ/, /oː/-/ɔ/, and /aː/-/a/). Phonetically, the primary cue for distinguishing German vowel length is temporal duration, while spectral quality serves as a secondary cue (Predeck, 2022; Tomaschek et al., 2015).

Vowel length is also present in British English, but its phonemic status and phonetic realization differ from German. Phonologically, not all English vowels are paired in length. Phonetically, English vowel length is realized mainly by quality rather than duration (Hillenbrand et al., 1995). Consequently, these vowels are more often referred to as “tense” and “lax” instead of “long” and “short.”

As for Cantonese, it has 13 vowel phones, including seven long and six short vowels. However, the distribution and the duration/quality status of vowel length are under debate. The eight-vowel system (H. N. S. Cheung, 1972; Yue-Hashimoto, 1972) considers vowel duration as a non-essential cue, so [ɪ]-[iː], [ʊ]-[uː]; [ɵ]-[œː] are regarded as allophones. The 11-vowel system (Bauer & Benedict, 1997; K. Cheung, 1986; Lee, 1983) considers vowel quality as a secondary feature, and [ɪ]-[e] and [ʊ]-[o] are considered allophones. As a result, the eight-vowel system recognizes only the /ɐ/-/aː/ pair as phonemic, whereas the 11-vowel system recognizes/e/-/ɛː/, /o/-/ɔː/, /ɵ/-/œː/, and /ɐ/-/aː/ pairs as phonemic.

Among these languages, standard Mandarin Chinese has the simplest vowel system (see Figure 1) and does not have vowel length contrasts (Chao, 1968).

These cross-linguistic differences provide a unique testing ground for trilingual vowel acquisition in this study. The durational reliance in German contrasts with the spectral reliance in English, creating an acquisition burden for L3 German acquisition. Furthermore, the varying phonemic status of vowel length in Cantonese and Mandarin allows us to isolate the extent to which trilinguals’ L1 phonology modulates the acquisition of L3 German vowel length contrasts.

1.2 The Acquisition of L3 Vowel Length Contrasts

Vowel length acquisition in an L3 can be a topic of theoretical explorations. Regarding whether the L3 vowels are influenced by L1 or L2 more, third language acquisition theories made divergent predictions. Although these theories target L3 morphosyntax, the principles can also be applied to L3 phonology. Selective transfer models, such as the Typological Primacy Model (Rothman, 2011, 2015) and the L2 Status Factor (Bardel & Falk, 2007), posit that transfers originate primarily from one prior language. Simultaneous transfer models, including the Cumulative-Enhancement Model (Flynn et al., 2004), the Linguistic Proximity Model (Westergaard & Rodina, 2017), and the Scalpel Model (Slabakova, 2017), predict influence from both the L1 and L2. Dynamic accounts, such as Dynamic Systems Theory (De Bot et al., 2007), view transfer as a fluid process involving all languages. As for the manner of L3 vowel articulation, several studies have extended the Speech Learning Model (SLM; Flege, 1995) beyond its original focus. The SLM and its revised version (SLM-r; Flege & Bohn, 2021) posit that language similarity may not necessarily lead to successful acquisition. If an L2 sound is perceived as being similar but not identical to an L1 sound, it will be assimilated into the L1 category, which will result in a foreign accent (Flege, 1995). This framework has been applied to L3 contexts, where research indicates that L3 vowels undergo category formation processes analogous to those described in the SLM (Lipińska, 2015; Sypianska, 2016).

As for empirical studies, most of them focused on L3 vowel quality, rather than vowel length. Kamiyama (2007) examined the production and perception of L3 French vowels by Japanese–English–French trilinguals. The study found instances of CLIs from both the L1 and the L2. Specifically, the trilinguals confused the L3 French /u/ with the L1 Japanese /u/, which sounded similar to French /ø/. They also mixed up the L3 French /ø/ with /y/, influenced by the L2 English spelling of “eu” pronounced /juː/. Focusing on the production of L3 German vowels, Lipińska (2015) found that Polish–English–German trilinguals merged the L3 German /œ/ with the L1 Polish /u/ and the L2 English /uː/. This occurred because the speakers perceived the sounds as similar. These studies suggest that the acquisition of L3 vowel quality is influenced by both L1 and L2. Similar cumulative effects from L1 and L2 to L3 are also found in pitch perception (Wiener & Goss, 2019).

In addition, L1 and L2 vowel qualities were found to experience reverse transfers from the L3 vowel system. Sypiańska (2016) examined the production of Polish /ɛ/, English /ɛ/, German /e, ɛ/ and Spanish /e/ by Polish–English–German, Polish–English–Spanish, and Polish–German–English trilinguals. Acoustic analyses on vowel formants revealed a reverse influence from L2 English to L1 Polish, which resulted in the raising and backing of L1 vowels. In addition, a combined reverse influence from L2 German and L3 English affected L1 Polish, leading to the lowering and backing of L1 vowels. Similarly, Kopečková et al. (2016) examined Polish, English, and German vowel qualities produced by eight 13- and 14-year-old multilinguals, revealing a marked individual variability and full interconnectedness of the multilingual subsystems. These studies on L3 vowels provide empirical support for a dynamic account of L3 acquisition. Therefore, the hypotheses of this study will be formed based on the dynamic account, assuming full interaction among the three languages.

In the domain of L3 vowel length perception, Luo et al. (2020) investigated the discrimination of three Cantonese vowel length pairs (/aː/–/ɐ/, /ɛː/–/e/, /ɔː/–/o/) by Mandarin–English–Cantonese trilinguals with varying levels of L3 Cantonese experience. Their findings demonstrated that experience with L2 English vowel length facilitated the perception of L3 Cantonese length contrasts. However, to date, no research has addressed L3 vowel length production. While perception and production are closely related, they represent distinct processes; successful perception of a phonological contrast does not guarantee accurate production, as learners must also acquire the articulatory means to realize the contrast reliably.

1.3 The Acquisition of L2 Vowel Length Contrasts

Unlike L3 studies that mainly investigated vowel quality, L2 studies have explored the vowel length feature in more depth. First, various studies found that L2 speakers tended to rely more on duration than spectral features in perceiving and producing L2 vowel length contrasts. Bohn (1995) examined the perception of English vowel length contrasts (such as beat and bit) by different groups of non-native listeners who spoke Spanish, German, and Mandarin as L1s. Results suggested that the non-native listeners all relied heavily on duration cues, whereas native listeners relied on spectral cues to distinguish English vowel length. Therefore, the Desensitization Hypothesis was proposed, stating that if the speakers’ previous language experience does not “sensitize” them to the spectral differences of certain unfamiliar vowels, they will use duration cues to distinguish these vowels instead. This perceptual pattern echoes the findings in Y. Chen (2006) on L2 vowel length production: Mandarin speakers of L2 English relied more on temporal features, whereas native English (NE) speakers relied more on spectral features to produce American English vowel length contrasts, /i/-/ɪ/, /æ/-/ɛ/, and /u/-/ʊ/.

Second, L2 vowel length acquisition is also influenced by the L1 system. McAllister et al. (2002) examined the production and perception of L2 Swedish vowel length by American English, Latin American Spanish, and Estonian speakers. The results suggested that the difficulty of acquiring L2 Swedish long and short vowels varied as a function of the importance of the role played by duration cues in learners’ L1s. The study proposed the Feature Hypothesis (Flege, 1995; McAllister et al., 2002), which states that L2 features not used phonologically in L1 will be difficult to perceive and produce. Consistent findings were obtained in Nimz (2016), which showed that Polish learners of L2 German were less able to distinguish L2 German vowel length than native speakers because the learners’ L1 vowels did not have phonemic length contrasts.

To acquire vowel length contrast in a non-native language, learners need to adopt the vowel duration and spectral configurations in the target language. The above findings show that such a process is constrained by the native language inventory as well as universal mechanisms.

1.4 The Present Study

This study supplements previous research in several important ways. To begin with, past studies on L3 phonological acquisition often compared a trilingual group with only one control group. This approach makes it challenging to accurately isolate the complex interactions among different languages. Our study addresses this issue by utilizing a more rigorous design, which compares speech samples from trilingual speakers against data collected from multiple control groups. Besides, previous empirical research has primarily concentrated on vowel quality. However, many languages distinguish vowels not only by quality but also by duration. The characteristics of vowel length are equally important and deserve exploration. This study fills this gap by extending the focus from vowel quality to vowel length, thereby providing a more comprehensive understanding of vowel features in L3 acquisition.

Moreover, this study has theoretical contributions. By examining vowel length features in this interesting language triad of Cantonese, English, and German, the study reveals the complex interplay among the three linguistic systems. The CLIs found in this study can validate existing theories on L3 morpho-syntactic acquisition, such as the Dynamic Systems Theory, from a phonological perspective. In addition, the vowel length realization patterns identified in L3 can extend the L2 acquisition theories based on vowel length data, like the Feature Hypothesis theory, to the field of L3 acquisition. The research questions (RQs), hypotheses (Hs), and data-specific predictions (Ps) are as follows:

RQ1: How do Cantonese–English–German (CEG) trilinguals implement duration and quality cues in the realization of vowel length compared with Mandarin–English–German (MEG) trilinguals?

H1: Based on the Feature Hypothesis, we expect the L1 of CEG trilinguals to benefit the acquisition of L3 vowel length, so the speakers could implement duration and quality features to produce the target vowels.

P1: The CEG trilinguals’ speech will be more similar to native German (NG) speech than the speech of the MEG trilinguals.

H2: Based on the Desensitization Hypothesis, we expect the trilinguals to rely more heavily on duration than quality to produce the L3 vowel length difference.

P2: Both the CEG and MEG trilingual cohorts will produce duration differences for L3 vowel pairs that are comparable to those of NG speakers; however, they will produce reduced quality differences in comparison with their native counterparts.

RQ2: To what extent and in what ways do L1 Cantonese, L2 English, and L3 German vowel systems influence one another?

H3: Based on the Dynamic Systems Theory and empirical findings on L3 vowel production, we assume CLIs to occur dynamically among L1, L2, and L3.

P3: CLIs will take place from L1 to both L2 and L3. MEG trilinguals and CEG trilinguals will differ in their L2 English and L3 German production.

P4: L3 will impose backward CLIs on L1 or L2. The CEG trilinguals’ L1 or L2 vowels differ from those of Cantonese-English (CE) bilinguals in quality or duration.

2 Method

Data for the study, including participants’ age, gender, language use, self-reported proficiencies, vowel production values, R codes for statistical modelling and outputs, are available online (https://osf.io/juc4k/overview).

2.1 Participants

Seventy participants were recruited, divided into five groups: 14 CEG trilinguals (7F, 7M; Mean age: 22.6 years, SD = 0.7 years), 14 MEG trilinguals (7F, 7M; Mean age: 20.3 years, SD = 0.9 years), 14 CE bilinguals (9F, 5M; Mean age: 26.1 years, SD = 2.6 years), 14 native British English speakers (NE, 6F, 8M; Mean age: 28.4 years, SD = 9.1 years), and 14 NG speakers (13F, 1M; Mean age: 25.6 years, SD = 4.1 years).

The CEG trilingual group were students taking L3 German classes at the Chinese University of Hong Kong at the time of the study. As Hong Kong is a bilingual society, these local students were native in Cantonese and quite fluent in English. The starting age of learning L2 English was 3.0 (SD = 0.7 years). Their mean score on the International English Language Testing System (IELTS; band scale 1–9) was 6.8 (SD = 0.98). As for the L3 German, the students had been taking German IV lessons weekly for 2 years through 1.5-hr lessons, but few had extensive overseas experience in German-speaking countries. Their instructor estimated their German proficiency to be around A2-B1 level in the Common European Framework. Overall, the trilinguals were considered as advanced in L2 and pre-intermediate in L3.

The MEG trilingual group consisted of students taking German classes at the Sun Yat-sen University in China. They were native speakers of Mandarin, originating from northern China. As English majors, they were also quite proficient in English. The mean starting age of learning English was 7.8 (SD = 2.1 years). All of them had passed the Test for English Majors 4 (TEM-4), and two of them took IELTS and obtained scores of 7.5 and 7.0. In terms of L3, the trilinguals had been taking German classes for 1.2 years through 3-hr weekly lessons and none of them had overseas experience. According to their instructor, their German proficiency level was at A2-B1. The above information shows that the MEG trilinguals were advanced in L2 and pre-intermediate in L3, which was similar to the CEG trilinguals.

The CE bilingual group had a similar language background as the CEG trilinguals, except that they had not learnt any German. They were also local Cantonese-speaking students recruited at the Chinese University of Hong Kong. They started to learn L2 English at the age of 3.2 (SD = 1.3 years), their English proficiency was equivalent to an IELTS score of 6.9 (SD = 0.6).

The NE group consisted of British English speakers from the United Kingdom, and the NG group consisted of northern standard German speakers from Germany. The two native groups were expats studying or working in the Netherlands. The native speakers’ accent was considered standard because they were German and English language teachers in the Netherlands. They possessed a high degree of metalinguistic awareness and deliberately spoke with a standard accent during the experiment.

Participants also self-evaluated their non-native language proficiencies (Table 1). One-way analyses of variance (ANOVAs) were performed to compare the proficiencies of different participant groups. The data were checked for the assumptions. Shapiro–Wilk tests confirmed the normality of the distributions, and Levene’s tests indicated homogeneity of variance across groups. No significant differences were found among the CEG trilinguals, MEG trilinguals, and CE bilinguals in their self-rated English proficiency across four skills and accent: speaking, F(2, 39) = 2.08, p = .139, η² = .10; listening, F(2, 39) = 1.41, p = .258, η² = .07; reading, F(2, 39) = 2.21, p = .124, η² = .10; writing, F(2, 39) = 0.46, p = .634, η² = .02; or accent, F(2, 39) = 1.78, p = .183, η² = .08. Similarly, when comparing the CEG trilinguals and MEG trilinguals on their self-rated German proficiency, no significant group differences were found in speaking, F(1, 26) = 0.00, p = 1.00, η² = .00; listening, F(1, 26) = 1.44, p = .24, η² = .05; reading, F(1, 26) = 0.59, p = .45; writing, F(1, 26) = 0.54, p = .47, η² = .02; or accent, F(1, 26) = 0.59, p = .45, η² = .02. The CEG trilinguals and the MEG trilinguals regarded themselves to be better in English (“good”) than in German (“fair to adequate”). All groups rated their English proficiency as “good.” The self-rating thus confirmed the comparability of the three groups in their L2 and L3 proficiencies.

Table 1.

Self-Reported Means (and Standard Deviations) of Proficiency (1 = Very Low, 2 = Low, 3 = Fair, 4 = Adequate, 5 = Good, 6 = Very Good, 7 = Perfect) and Accent (1 = No, 2 = Very Light, 3 = Some, 4 = Considerable, 5 = Heavy) in L2 English and L3 German.

		Proficiency rating
Group	Language	Speaking	Listening	Reading	Writing	Average	Accent rating
Cantonese–English–German trilingual	English	4.92 (0.72)	5.69 (0.63)	5.46 (0.99)	4.69 (1.04)	5.19 (0.84)	3.23 (0.99)
	German	3.00(1.41)	3.38(0.83)	3.69(0.99)	3.08(0.92)	3.29(1.03)	3.15(1.41)
Mandarin–English–German trilingual	English	4.68(1.06)	5.41(0.71)	5.45(0.50)	4.86(0.76)	5.10(0.76)	2.86(1.18)
	German	3.18(0.71)	2.77(1.13)	3.64(0.97)	4.05(0.82)	3.48(0.91)	3.18(0.72)
Cantonese–English bilingual	English	4.71(0.86)	5.28(0.70)	5.07(0.59)	4.71(0.79)	5.10(0.73)	2.50(0.62)

2.2 Materials

The reading materials were minimal and near-minimal pairs in German, English, and Cantonese, designed to contrast vowel length. When perfect minimal pairs were unavailable in the target languages, near-minimal pairs (e.g., took-tooth) were used. Front, back, and low vowels in all three languages were tested, including 14 monophthongs in German (/iː/-/ɪ/, /yː/-/ʏ/, /uː/-/ʊ/, /eː/-/ɛ/, /øː/-/œ/, /oː/-/ɔ/, and /aː/-/a/), six in British English (/iː/-/ɪ/, /ɑː/-/ʌ/, and /uː/-/ʊ/), and six in Cantonese (/aː/-/ɐ/, /ɛː/-/e/, and /ɔː/-/o/). The vowel length pairs were monosyllabic or disyllabic words (e.g., peak–pick), embedded in carrier phrases (German: Ich habe __ gesaget, jetzt sage Ich __ noch. “I said __, and now I say __”; English: I say __first, and I say __now; Cantonese: 我讀__字, 我讀__字 (ngo5 duk6 __ zi6). “I say __, and I say _”. Each pair appeared three times in random order. The long vowels always came before the short vowels to avoid mispronunciation, as in the pilot study, some trilinguals confused the L3 long with short vowels due to their unfamiliarity with German orthographical rules. The full reading material is provided in the appendix.

2.3 Procedures

Participants were individually recorded in a sound-attenuated room using a portable recorder (TASCAM DR-22WL) at a sampling rate of 44.1 kHz/16bit. The CEG trilingual and CE bilingual groups were recorded in Hong Kong, the NE and NG groups in Utrecht, and the MEG trilingual group in Zhuhai. Ethical approval has been obtained. All participants gave informed written consent to participate in the research. During the experiment, participants read the materials at a natural, comfortable speed and were allowed to make multiple attempts to correct speech errors. For bilinguals and trilinguals who read the material in several languages, the order of each language block was counterbalanced. In each language block, participants read a passage in the target language before reading the critical minimal pairs. This was to control for the language mode so that the bi-/trilinguals fully activate the target language and inhibit other interfering languages (Grosjean, 2001). After the recording, the participants completed a language background questionnaire and received payment for their participation.

2.4 Analysis

A total of 3,528 vowel productions were originally collected (42 speakers × 36 German trials + 56 × 24 English trials + 28 × 24 Cantonese trials). Of these, 83 tokens were excluded due to speech errors or background noise. The remaining 3,445 vowel productions were analysed in Praat 6.3.08 (Boersma & Weenink, 2015). Four measurements were taken to compare the long and short vowels, including vowel duration, vowel quality, duration difference, and quality difference.

Vowel duration was the total duration of the target vowel. The segmentation of vowels was based on visual inspection of spectrogram and waveform displays. A vowel onset was defined as the acoustic onset of periodicity and the offset was defined as the termination of the period of the final regular pitch pulse (S. Chen et al., 2017; Harrington et al., 2011; Oh et al., 2011; Wassink et al., 2007). For three German vowels before liquid /l/ (i.e., “Polen-Pollen,” “Höhle-Hölle,” “Mühle-Müller”), the duration included transition and the vowel steady state.

Vowel quality denoted the height and frontness of the target vowel. The F1 and F2 values were measured from the midpoint of the full vowel segment. To eliminate variations caused by physiological differences among individual speakers, the formant values were normalized with the Lobanov method (Adank et al., 2004) using the following formula:

{F_{n [V]}}^{N} = (F_{n [V]} - {Mean}_{n}) / S_{n}

where F_n[V]^N was the normalized value for F_n[V] (i.e., for formant n of vowel V). Mean_n was the mean value for formant n for the speaker in question and S_n was the standard deviation for the speaker’s formant n.

3. Duration difference was the temporal distance between each long-short vowel minimal pair, which was quantified as the Duration Ratio between long and short vowels:

D u r a t i o n R a t i o = \frac{D u r a t i o n_{l o n g}}{D u r a t i o n_{s h o r t}}

where Duration_long was the duration of the long vowel, and Duration_short was the duration of the short vowel.

4. Quality difference was the spectral distance between each long-short vowel minimal pair, which was quantified as the Euclidean distance (ED) between long and short vowels:

E D_{l, s} = \sqrt{{(F 1_{l} - F 1_{s})}^{2} + {(F 2_{l} - F 2_{s})}^{2}}

where ED_l,s was the distance between one particular long and short vowel minimal pair on an F1–F2 vowel space. F11 and F21 were the first two formants of the long vowel, and the F1_s and F2_s were the first two formants of the short vowel.

After extracting the above values, the study took three steps to analyse CLIs. First, the vowel productions were plotted on F1–F2 planes, showing each group’s vowel spaces. Possible CLIs could be inferred by a visual inspection of the vowel qualities on the planes.

Second, mixed-effects models were built on the EDs and duration ratios of each language to show group differences in their abilities to differentiate long-short vowels, using Rstudio (R Core Team, 2024), the “lme4” package (Bates et al., 2015), and the “lmerTest” package (Kuznetsova et al., 2017). To examine potential L1-to-L3 influences, the German production of CEG trilinguals was compared with that of MEG trilinguals, as these groups differ primarily in their L1 backgrounds. Furthermore, a comparison between CEG trilinguals and CE bilinguals in English production was designed to identify possible L3-to-L2 backward influences. Finally, comparing L2 and L3 productions against their respective native-speaker norms allowed for an assessment of the trilinguals’ phonetic nativelikeness.

Third, to investigate the patterns of CLI among the three languages of CEG trilinguals, we explored individual variations in the production of cross-linguistically similar vowel pairs. The vowel pairs in Table 2 were treated as “similar” based on their shared featural specifications and acoustic positioning. For instance, in Pair 1, German /iː/-/ɪ/ and English /iː/-/ɪ/ are both high-front pairs. In Pair 3, the German /aː/-/a/ and Cantonese /aː/-/ɐ/ pairs are analogues because both contrast a long, open-central vowel with a shorter counterpart. Following the revised SLM (Flege & Bohn, 2021), we predict L1, L2, and L3 vowels coexist in a shared phonetic space, and structural similarity among these vowels stimulates CLI. When a CLI occurs, a trilingual speaker who produces a larger duration ratio and a longer ED for a vowel pair in language A is likely to also produce a larger duration ratio and a longer ED for a similar vowel pair in language B. For example, if a CLI occurs between L2 and L3, then a CEG trilingual who produces greater contrasts for L2 English /iː/-/ɪ/ may also produce larger contrasts for L3 German /iː/-/ɪ/. In the current analysis, each trilingual participant’s mean duration ratio and ED of the cross-linguistically similar vowel pairs were calculated, and the resultant mean duration ratio and ED were correlated between trilinguals’ L2 and L3, L1 and L3, and L1 and L2 by Pearson correlation tests.

Table 2.

Vowel Pairs for Correlation Analyses.

		Pair 1	Pair 2	Pair 3
L2 and L3	German	/iː/-/ɪ/	/uː/-/ʊ/	/aː/-/a/
L2 and L3	English	/iː/-/ɪ/	/uː/-/ʊ/	/ɑː/-/ʌ/
L1 and L3	German	/eː/-/ɛ/	/oː/-/ɔ/	/aː/-/a/
L1 and L3	Cantonese	/ɛː/-/e/	/ɔː/-/o/	/aː/-/ɐ/
L1 and L2	English	/ɑː/-/ʌ/
L1 and L2	Cantonese	/aː/-/ɐ/

3 Results

3.1 An Overview of the Long- and Short-Vowel Qualities

Figure 2 provides an overview of the vowel qualities produced by each group. The normalized F1 and F2 of tense and lax vowels were plotted to visualize the height and frontness of the vowels.

Figure 2.

Vowel plots of German, English and Cantonese in the F1–F2 space.

Based on a visual inspection of the German vowel plots, we could observe that the NG long-short vowel minimal pairs appeared clearly separated in the vowel space. In contrast, the CEG and MEG trilinguals’ productions appeared less differentiated than the native productions. Furthermore, the MEG trilinguals’ long–short pairs appeared even less contrastive than those of the CEG trilinguals.

Visual inspection of the English vowel plots showed that NE long–short pairs were dispersed in the vowel space. A broadly similar pattern was observed for the CE bilinguals and the CEG trilinguals, although the trilinguals’ /uː/ productions appeared slightly more back than those of the bilinguals. Only the MEG trilinguals’ /uː/ and /ʊ/ appeared to be relatively close to each other.

Visual inspection of the Cantonese vowel qualities suggested that CE bilinguals and CEG trilinguals exhibited largely comparable L1 production patterns.

These patterns suggest that participants with different language backgrounds may vary in their vowel length production. To examine whether these apparent differences were statistically reliable, we conducted further quantitative analyses comparing vowel productions across the participant groups.

3.2 Group Comparison in Duration and Quality Contrasts of Vowel Length Pairs

To test whether the above visually observed patterns were statistically significant, mixed models were built on the duration ratios and the EDs of vowel minimal pairs in each language, with Group and Vowel Type as fixed factors. For the model on German, the Group factor had three levels: CEG trilinguals, MEG trilingual, and NG speakers. The Vowel Type factor had seven levels: /iː/-/ɪ/, /yː/-/ʏ/, /uː/-/ʊ/, /eː/-/ɛ/, /øː/-/œ/, /oː/-/ɔ/, and /aː/-/a/. For the model on English, the Group factor contained four levels: CEG trilinguals, MEG trilinguals, NE speakers, and CE bilinguals. The Vowel Type factor had three levels: /iː/-/ɪ/, /ɑː/-/ʌ/, and /uː/-/ʊ/. For the model on Cantonese, the Group factor had two levels: CEG trilinguals and CE bilinguals. The Vowel Type factor had three levels: /aː/-/ɐ/, /ɛː/-/e/, and /ɔː/-/o/. Following recommendations by Barr et al. (2013), we initially specified a maximal random-effects structure, including random intercepts and slopes for both subjects and items. However, the model failed to converge, which is likely due to the relatively small number of observations and the complexity of the random-effects structure. We therefore simplified the model stepwise, first removing correlation parameters and then random slopes. This approach is supported by Matuschek et al. (2017), as fitting an overly complex random-effects structure can lead to a substantial and unnecessary decrease in statistical power. Each iteration of the model was compared with the previous best-fit model using the Akaike information criterion (AIC). The final models were the simplest models that provided the best fit. Here, only the final model results are reported. The duration ratio and ED data are plotted in Figure 3.

Figure 3.

Group differences in temporal and spectral vowel production across German, English, and Cantonese. (a) Mean duration ratios for each vowel pair (top). (b) Mean Euclidean distances for each vowel pair (bottom).

The best-fitting model for the German duration ratios was the one with Group, Vowel, and their interactions as fixed effects. The by-item random intercept was initially included but was removed because each item is associated with a vowel category and contributed negligible variance, resulting in singular model fits. Based on the final model (DR~ group * vowel + (1 + vowel|subj)), we conducted further Tukey-adjusted within-vowel pairwise comparisons. As summarized in Table 3, the MEG group consistently showed reduced duration contrasts relative to the NG group across all vowel pairs (all ps < .05). In contrast, the CEG group patterned more closely with the NG group, showing no significant differences from the NG group except for /aː/-/a/ and /øː/-/œ/, while exhibiting significantly larger duration ratios than the MEG for /eː/-/ɛ/, /iː/-/ɪ/, /oː/-/ɔ/, and /yː/-/ʏ/. Therefore, the CEG trilinguals did not differ significantly from NG speakers in producing most of the L3 vowel duration contrasts, whereas the MEG trilinguals consistently produced smaller duration contrasts than NG speakers for the L3 vowel pairs.

Table 3.

Pairwise Comparisons Across Vowels and Groups for German Duration Ratio (DR) and Euclidean Distance (ED).

Vowel	Contrast	DR estimate	DR SE	DR t	DR p	ED estimate	ED SE	ED t	ED p
/aː/-/a/	CEG-MEG	0.00	0.14	0.00	1.000	6.52	37.30	0.18	.983
/aː/-/a/	CEG-NG	−0.50	0.15	−3.34	.003	−31.89	40.30	−0.79	.709
/aː/-/a/	MEG-NG	−0.50	0.15	−3.35	.003	−38.42	40.10	−0.96	.604
/eː/-/ɛ/	CEG-MEG	0.66	0.17	3.96	<.001	67.62	49.80	1.36	.364
/eː/-/ɛ/	CEG-NG	0.14	0.17	0.83	.683	−202.65	52.70	−3.85	<.001
/eː/-/ɛ/	MEG-NG	−0.51	0.15	−3.43	.002	−270.27	41.50	−6.52	<.001
/iː/-/ɪ/	CEG-MEG	0.71	0.14	5.02	<.001	25.65	37.40	0.69	.772
/iː/-/ɪ/	CEG-NG	0.31	0.15	2.12	.091	−278.06	41.20	−6.75	<.001
/iː/-/ɪ/	MEG-NG	−0.40	0.15	−2.70	.021	−303.70	40.90	−7.43	<.001
/oː/-/ɔ/	CEG-MEG	0.48	0.15	3.21	.005	28.92	39.70	0.73	.747
/oː/-/ɔ/	CEG-NG	−0.03	0.16	−0.21	.976	−278.70	43.00	−6.48	<.001
/oː/-/ɔ/	MEG-NG	−0.51	0.15	−3.42	.002	−307.62	40.30	−7.64	<.001
/øː/-/œ/	CEG-MEG	0.16	0.16	1.07	.537	9.11	42.30	0.22	.975
/øː/-/œ/	CEG-NG	−0.55	0.16	−3.38	.003	−39.43	45.50	−0.87	.661
/øː/-/œ/	MEG-NG	−0.71	0.15	−4.73	<.001	−48.54	40.90	−1.19	.462
/uː/-/ʊ/	CEG-MEG	0.31	0.15	2.17	.080	−14.16	38.50	−0.37	.928
/uː/-/ʊ/	CEG-NG	−0.13	0.15	−0.85	.674	−147.30	41.20	−3.57	.001
/uː/-/ʊ/	MEG-NG	−0.44	0.15	−2.97	.010	−133.14	40.20	−3.31	.003
/yː/-/ʏ/	CEG-MEG	0.51	0.16	3.25	.004	121.73	44.80	2.72	.019
/yː/-/ʏ/	CEG-NG	−0.27	0.16	−1.69	.213	−94.97	47.70	−1.99	.115
/yː/-/ʏ/	MEG-NG	−0.78	0.15	−5.21	<.001	−216.70	41.10	−5.27	<.001

Signiﬁcant results (p < .05) are presented in bold text.

The best-fitting model for German EDs (ED ~ group * vowel + (1 + vowel|subj)) also suggested Group, Vowel, and their interaction as significant predictors. The post hoc comparison output in Table 3 revealed that differences between the CEG and MEG groups were generally not significant across most vowel pairs. A significant difference between these two trilingual groups emerged only for /yː/-/ʏ/, for which the CEG group produced larger values than the MEG group. In contrast, both trilingual groups differed reliably from the NG group for several vowels. Specifically, for /eː/-/ɛ/, /iː/-/ɪ/, /oː/-/ɔ/, and /uː/-/ʊ/, both the CEG and MEG groups produced significantly smaller values than the NG group. A similar pattern was observed for /yː/-/ʏ/, where the MEG group also produced significantly smaller values than the NG group, although the difference between the CEG and NG groups did not reach significance. These results indicate that the two trilingual groups exhibited largely comparable vowel quality contrasts, while both groups tended to produce reduced contrasts relative to the native speakers for several L3 vowel length pairs.

The final model for English duration ratios (DR ~ group * vowel + (1 + vowel|subj)) also suggested Group, Vowel, and their interaction to be significant predictors. The post hoc test results are shown in Table 4. For /ɑː/-/ʌ/, significant group separation was observed mainly for the MEG group, who had lower values than the NE, CEG, and MEG groups. In contrast, no significant group differences were found for /iː/-/ɪ/. Group effects emerged for /uː/-/ʊ/, with the CE group exhibiting higher values than the CEG, MEG, and NE groups. Therefore, the MEG trilinguals produced smaller duration contrasts than the native speakers in their L2 English vowel length production, whereas the CE bilinguals produced larger duration contrasts in their L2 English. In contrast, the CEG trilinguals did not differ significantly from the native speakers in producing L2 English vowel duration contrasts.

Table 4.

Pairwise Comparisons Across Vowels and Groups for English Duration Ratio (DR) and Euclidean Distance (ED).

Vowel	Contrast	DR estimate	DR SE	DR t	DR p	ED estimate	ED SE	ED t	ED p
/ɑː/-/ʌ/	CE-CEG	−0.15	0.1	−1.48	.458	−10.50	34.80	−0.30	.990
/ɑː/-/ʌ/	CE-MEG	0.31	0.09	3.32	.009	22.10	32.60	0.68	.905
/ɑː/-/ʌ/	CE-NE	−0.17	0.1	−1.72	.324	36.80	33.80	1.09	.698
/ɑː/-/ʌ/	CEG-MEG	0.46	0.1	4.51	<.0001	32.60	35.30	0.92	.792
/ɑː/-/ʌ/	CEG-NE	−0.02	0.1	−0.17	.998	47.30	36.40	1.30	.567
/ɑː/-/ʌ/	MEG-NE	−0.48	0.1	−4.85	<.0001	14.60	34.30	0.43	.974
/iː/-/ɪ/	CE-CEG	−0.09	0.14	−0.64	.92	88.10	63.60	1.38	.514
/iː/-/ɪ/	CE-MEG	−0.23	0.14	−1.65	.357	46.10	62.90	0.73	.883
/iː/-/ɪ/	CE-NE	0.11	0.14	0.76	.874	−21.40	64.20	−0.33	.987
/iː/-/ɪ/	CEG-MEG	−0.14	0.15	−0.98	.762	−42.00	64.20	−0.65	.914
/iː/-/ɪ/	CEG-NE	0.2	0.15	1.35	.536	−109.50	65.50	−1.67	.348
/iː/-/ɪ/	MEG-NE	0.34	0.15	2.34	.101	−67.50	64.80	−1.04	.725
/uː/-/ʊ/	CE-CEG	0.49	0.13	3.86	.002	66.20	63.60	1.04	.727
/uː/-/ʊ/	CE-MEG	0.46	0.13	3.53	.005	141.60	64.50	2.19	.137
/uː/-/ʊ/	CE-NE	0.74	0.13	5.75	<.0001	−168.80	64.50	−2.62	.053
/uː/-/ʊ/	CEG-MEG	−0.03	0.13	−0.26	.994	75.50	63.40	1.19	.636
/uː/-/ʊ/	CEG-NE	0.25	0.13	1.98	.208	−235.00	63.40	−3.71	.003
/uː/-/ʊ/	MEG-NE	0.29	0.13	2.19	.138	−310.50	64.30	−4.83	<.001

Signiﬁcant results (p < .05) are presented in bold text.

In terms of the English vowel ED model, the selected best model was the one including only Vowel, Group, and their interaction as the fixed effects (ED ~ vowel *group + (1 + vowel|subj)). The pairwise comparison results in Table 4 showed that no significant group differences were observed for /iː/-/ɪ/ or /ɑː/-/ʌ/ contrasts. For /uː/-/ʊ/, however, significant differences emerged: both the CEG and MEG groups produced significantly smaller values than the NE group. The results indicate that group differences were largely absent for /iː/-/ɪ/ or /ɑː/-/ʌ/, whereas for /uː/-/ʊ/, the trilingual groups exhibited smaller contrasts than the native group.

As for the Cantonese vowel duration ratio model, the final selected model included Vowel as the sole fixed effect (DR ~ vowel + (1 + vowel| subj) + (1| item)). Model comparisons indicated that adding Group as a main effect, χ²(1) = 0.04, p = .846, or a Group × Vowel interaction, χ²(2) = 4.90, p = .086, did not yield a significant improvement. Thus, vowel duration ratio differences differed across vowels but not groups.

In terms of the Cantonese vowel ED, the null model including only random effects (ED ~ 1 + (1 + vowel| subj) + (1| item)) was the chosen model. Adding Group, χ²(1) = 3.63, p = .06; Vowel, χ²(1) = 1.51, p = .47; or their interaction, χ²(2) = 8.90, p = .11, did not result in a significant improvement in model fit, while increasing model complexity. These results indicate that vowel EDs did not differ significantly across vowels or between groups. Any observed variability in EDs was primarily attributable to individual differences between speakers and items.

In summary, the CEG trilinguals were more nativelike than the MEG trilinguals in their German vowel length production. The CEG trilinguals produced similar duration differences for most of the German vowel length pairs as native speakers, whereas the MEG trilinguals consistently produced smaller duration differences than native speakers. Both groups of trilinguals exhibited smaller quality differences than native speakers for /eː/-/ɛ/, /iː/-/ɪ/, /oː/-/ɔ/, and /uː/-/ʊ/, but the MEG trilinguals additionally produced smaller quality contrasts for /yː/-/ʏ/. Regarding English production, the CEG trilinguals distinguished themselves as the most proficient non-native cohort in terms of temporal contrast. These trilinguals were near-native in realizing English vowel durations, while both MEG trilinguals and CE bilinguals deviated from natives in several instances of duration differentiation. Notwithstanding their temporal accuracy, the CEG trilinguals were less precise in their spectral execution, as evidenced by their compressed quality contrasts for /uː/-/ʊ/ relative to native speakers. Finally, in Cantonese production, all speakers performed similarly regardless of whether they were bilingual or trilingual.

3.3 CLIs and Individual Variations

The duration ratios and EDs of the cross-linguistic similar vowel pairs (Table 2) were subjected to Pearson correlation analyses to examine individual variations and potential CLIs. In terms of L2 English and L3 German, there was no significant correlation in duration ratios, r(38) = .00, p = .99. In contrast, as shown in Figure 4, a significant positive correlation was observed for EDs, r(38) = .47, p = .002. Therefore, aggregated EDs of the tested vowels reveal that trilinguals who maximized the quality differentiation for their L2 vowel contrasts, /iː/–/ɪ/, /uː/–/ʊ/, and /ɑː/–/ʌ/, manifested more robust quality distinctions for similar L3 contrasts, /iː/–/ɪ/, /uː/–/ʊ/, and /aː/–/a/.

Figure 4.

The mean values of vowel length contrasts produced by each Cantonese–English–German trilingual speaker. (a) Euclidean distances of L2 English and L3 German vowel pairs (left). (b) duration ratios of L1 Cantonese and L3 German vowel pairs (right).

As for L1 Cantonese and L3 German, there was a significant negative correlation in duration ratios between L1 Cantonese and L3 German vowels, r(19) = −.62, p = .002. Based on the aggregate distances of the tested vowel pairs, we could infer that the trilinguals who produced more contrastive L1 Cantonese vowel pairs, /ɛː/-/e/, /ɔː/ - /o/, and /aː/-/ɐ / also produced smaller duration contrasts in L3 German vowel pairs, /eː/-/ɛ/, /oː/-/ɔ/, and /aː/-/a/. On the other hand, there was no significant correlation in ED between the tested L1 Cantonese and L3 German vowel length pairs, r(19) = .36, p = .10. The mean vowel ED values of each participant were plotted in Figure 4.

Regarding L1 Cantonese and L2 English, the correlation between L1 and L2 did not reach significance for either duration ratio, r(9) = −.50, p = .12, or ED, r(9) = .27, p = .41.

4 Discussion

4.1 The Realization of L1, L2, and L3 Vowel Length Pairs in Duration and Quality

In support of H1 and P1, the study shows that the CEG trilinguals are more nativelike in their production of L3 German vowel length contrasts than the MEG trilinguals. The duration ratio analysis indicates that the CEG trilinguals were similar to natives in most cases, while the MEG trilinguals always produced smaller vowel duration contrasts than native speakers. This finding is compatible with the Feature Hypothesis (McAllister et al., 2002), which proposes that the prominence of a feature in L1 phonology plays an important role in the acquisition of an L2. In McAllister et al. (2002), L2 Swedish vowel duration was better acquired by L1 Estonian speakers than by L1 Spanish speakers because duration played a central role in Estonian phonology but only a minor role in Spanish phonology. Similarly, Nimz (2016) found that L2 German vowel length pairs were produced with smaller duration contrasts by Polish L1 speakers than by NG speakers, as duration was not important for distinguishing Polish vowels. In this study, L3 German vowel durations were produced by L1 Cantonese speakers with nativelike accuracy, as duration cues are used to distinguish Cantonese vowels. In contrast, L3 German vowel durations were produced by L1 Mandarin speakers with smaller contrasts than those of native speakers, because duration cues are not phonologically meaningful in distinguishing Mandarin vowels. Taken together, these results extend the Feature Hypothesis to L3 acquisition, suggesting that L3 features not used to signal phonological contrasts in the L1 are also difficult to acquire and produce.

Concerning the relative weightings of quality and duration, H2 predicted that both trilingual groups would rely more heavily on duration than quality in L3 vowel length production. Specifically, P2 posited that they would produce duration distinctions comparable to those of native speakers, but less spectral differentiations. These predictions are partially supported by the production patterns of the CEG trilinguals; compared with native speakers, these trilinguals produced lower duration ratios in two of seven vowel pairs (/aː/-/a/ and /øː/-/œ/) and compressed EDs in four of seven (/eː/-/ɛ/, /iː/-/ɪ/, /oː/-/ɔ/, and /uː/-/ʊ/), indicating a better performance in duration than quality production. However, the MEG trilinguals did not exhibit the expected predominance of duration over quality. Instead, they produced lower duration ratios than native speakers across all seven vowel pairs, as well as reduced EDs for six pairs: /eː/-/ɛ/, /iː/-/ɪ/, /oː/-/ɔ/, /uː/-/ʊ/, and /yː/-/ʏ/, thereby failing to fully exploit either temporal or spectral cues. Interestingly, the MEG trilinguals appeared to weigh quality more heavily than duration; for the /øː/-/œ/ pair, they achieved ED values commensurate with those of NG speakers, yet produced lower duration ratios. Consequently, contrary to our predictions, the two trilingual groups did not uniformly prioritize temporal cues over spectral cues in L3 vowel length production.

The discrepancy between our predictions and the present results may stem from fundamental differences between speech production and perception. The predictions were formulated on the basis of the Desensitization Hypothesis (Bohn, 1995), a perceptual model which proposes that when L2 learners are unfamiliar with the spectral cues of L2 vowels, they tend to rely more heavily on duration cues in perception. However, in speech production, no comparable bias toward duration cues was observed. Further evidence for the divergence between production and perception emerges from a comparison between this study and prior research on L3 vowel perception. In the perception study conducted by Luo et al. (2020), trilingual participants achieved high accuracy in discriminating L3 Cantonese vowel length contrasts even at the initial stage of acquisition. By contrast, participants in the present production study were unable to produce L3 vowel length contrasts with comparable accuracy, despite having studied the L3 for 2 years. Moreover, Luo et al. (2020) argued that their findings challenged the Feature Hypothesis while lending support to the Desensitization Hypothesis. In contrast, the present L3 production study provides evidence in favour of the Feature Hypothesis and runs counter to the Desensitization Hypothesis. These findings suggest that the Feature Hypothesis may be more applicable to L3 speech production, whereas the Desensitization Hypothesis appears to better account for L3 speech perception.

Several factors contribute to the discrepancy between L3 production and perception. First, perception requires the listeners to attend to the relevant acoustic cues, whereas production necessitates the mastery of precise motor configurations. Consequently, even if a speaker can perceptually distinguish a vowel length pair, they may still struggle to accurately coordinate the articulatory gestures to realize that difference in speech. In addition, perception tasks, such as identification or discrimination, are often highly controlled and allow the participant to focus solely on a specific acoustic cue. Production tasks are more demanding as the speaker must manage lexical retrieval, syntactic encoding, and the physical articulation to complete the task. These factors all contribute to the asymmetry between L3 vowel length production and perception.

4.2 CLIs Among the L1, L2, and L3 Vowel Systems

Regarding RQ2 concerning the CLI effects, H3 assumed that CLIs would occur dynamically among L1 Cantonese, L2 English, and L3 German vowel systems. Such multidirectional interactions are shown in both group comparison results and individual variation patterns.

First, comparison between two trilingual groups reveals CLIs from L1 to L2 and L3, consistent with P3. Specifically, MEG trilinguals in general produced smaller duration contrasts than the CEG trilinguals in both their L2 (e.g., English /ɑː/-/ʌ/) and L3 (e.g., German /eː/-/ɛ/). This result aligns with previous research on L3 vowels (e.g., Kamiyama, 2007; Lipińska, 2015; Sypianska, 2016), which suggests that L3 vowel quality constitutes a fully interconnected system. These studies primarily focused on the collective influence of both the L1 and L2 on the L3. However, existing literature on L3 vowel production frequently relies on comparing trilingual vowels with baseline L1 and L2 values reported in previous studies rather than direct control groups. By incorporating multiple control groups, this study underscores the possibility that L2 vowels are already predisposed to L1 interference. Consequently, L3 production may be filtered through an L2 system that has already been modified by the native language, i.e., accented L2. This perspective, which identifies an L1-shaped articulatory routine, aligns with a growing body of research in L3 phonology. For instance, recent studies indicate that dominant motor patterns from the L1 play a crucial role in the production of L3 rhotics (Kopečková et al., 2023; Patience & Qian, 2022). Furthermore, a primary L1 influence on L3 phonological acquisition has been observed in production tasks involving vowels, consonants and stress contrasts (Norman et al., 2025). Therefore, the findings suggest a fundamental L1 influence that permeates the acquisition of all subsequent languages.

Second, an analysis of individual variation indicated both L1–L3 and L2–L3 interactions. Regarding L2–L3 interaction, the ED analyses revealed a positive correlation: trilinguals who produced robust quality contrasts in their L2 were more predisposed to manifest greater spectral contrasts within the related L3 vowel length pairs. This suggests that speakers may have transferred L2 phonetic rules to L3 vowel length differentiation, likely driven by an L2–L3 category assimilation process. According to SLM-r (Flege & Bohn, 2021), when an L2 sound is perceived as similar to an L1 sound, speakers tend to merge them into a single category. This study suggests that L2 and L3 vowels are likewise stored within a common phonetic space and undergo analogous processes. This finding resonates with previous L3 research (e.g., Lipińska, 2015), which provides evidence for category assimilation in trilingual vowel production. Conversely, regarding L1–L3 interaction, the duration ratio analyses indicated a negative correlation; specifically, speakers who produced narrower duration contrasts in their L1 tended to realize more pronounced contrasts in the corresponding L3 pairs. In this instance, the similarity between vowel pairs appeared to trigger category dissimilation rather than assimilation, whereby overlapping categories are actively separated to mitigate cross-linguistic interference.

As posited by SLM-r, category formation hinges on whether learners perceive distinctions between cross-linguistically similar sounds: failure to detect such differences results in the assimilation of L2 sounds into existing L1 categories, whereas successful differentiation prompts the establishment of new phonemic representations dissimilar from the L1 categories. It is possible that the trilinguals discerned the difference between similar L1 and L3 vowel pairs, which led to a conscious effort to maximize contrast, yet they failed to perceive the difference between similar L2 and L3 pairs, resulting in category merger. Further evidence of SLM-r’s role in L3 phonology is observed in the production of the /yː/-/ʏ/ contrast, the only pair consistently produced with accuracy in both duration and quality by the CEG trilinguals. Among the seven L3 contrasts examined, the /yː/-/ʏ/ pair stands alone in having no correspondence in either Cantonese or English. Due to this phonetic divergence, L3 learners can readily distinguish /yː/-/ʏ/ from L1 and L2 vowel length contrasts within the trilingual phonetic space. Since discerning cross-linguistic differences is a prerequisite for new category formation, the trilinguals possess a superior baseline for acquiring /yː/-/ʏ/, ultimately leading to higher production accuracy compared with other vowel pairs. To fully elucidate L3 speech development within the SLM-r framework, future research should incorporate perceptual tasks to determine the correlation between perceived similarity and the production accuracy of vowel length contrasts across three languages.

Furthermore, the results demonstrate that CLI can manifest as reverse transfers, though this occurs less frequently than forward transfers. Partially supporting P4, the data provided evidence of L3-to-L2 influence, whereas L3-to-L1 influence was absent. Specifically, the English duration ratio analysis reveals that the CE bilingual group produced higher values than the CEG trilingual group for the /uː/-/ʊ/ contrast, suggesting a backward influence from L3 German. Reverse transfer has been widely documented in previous L2 studies (Baker & Trofimovich, 2005; Seo & Dmitrieva, 2024). However, this study did not observe reverse transfer from L3 to L1, as no significant differences were found in either vowel quality or duration between the Cantonese productions of CEG trilinguals and those of CE bilinguals. The absence of L3-to-L1 influence may be attributable to limited immersive exposure to the L3. Except for one or two participants who had attended a 1-month summer programme in Germany, most participants learnt German through classroom instruction conducted twice a week in their home country. Evidence from a longitudinal study on trilingual vowel production (Kartushina & Martin, 2019), indicates that L3-to-L1 reverse transfer emerges only after 4 months of overseas L3 language programme. Moreover, this effect dissipated 4 months after learners returned to their native country. These findings suggest that L1 modification following L3 acquisition is possible, but only under conditions of intensive L3 input and use. Participants in this study may not have reached a sufficient level of L3 exposure or usage frequency to allow the L3 to exert a measurable reverse influence on the L1. Finally, the current finding that L3 exerts a stronger influence on L2 than on L1 is consistent with previous research on L3 phonology and morphosyntax (Amaro & Rothman, 2010; Stoehr et al., 2024), which demonstrates that L2 systems remain more permeable than L1 systems to influence from subsequently acquired languages.

5 Conclusion

This study investigates CLIs on the production of L3 German vowel length contrasts by CEG trilinguals. Results demonstrate a pervasive L1 role in third language acquisition. The L1 exerts a significant impact on both the L2 and L3; furthermore, the L1 proves more resilient than the L2 against backward CLI from the L3. CLIs between the L1/L3 and L2/L3 dyads can manifest as both category assimilation and dissimilation, ultimately reshaping the trilingual vowel space. These findings align with Dynamic Systems Theory, which emphasizes the fluidity and interdependence of linguistic systems. This research contributes to a deeper understanding of L3 phonological acquisition and calls for further studies to explore the interplay of linguistic features across diverse typological contexts.

Footnotes

Appendix

Table A1.

The Reading Materials in German, English, and Cantonese.

German	Ich habe “bieten”/ biːtn/ gesagt; jetzt sage ich “bitten” /bɪtn/.
	Ich habe “Beet” /beːt/ gesagt; jetzt sage ich “Bett” /bɛt/.
	Ich habe “Polen” /poːlən/ gesagt; jetzt sage ich “Pollen” /pɔlən/.
	Ich habe “Staat” /ʃta: t/ gesagt; jetzt sage ich “Stadt” /ʃtat/.
	Ich habe “Höhle” /høːlə/ gesagt; jetzt sage ich “Hölle” /hœlə/.
	Ich habe “Mus” /muːs/ gesagt; jetzt sage ich “muss” /mʊs/.
	Ich habe “Mühle” /myːlə/ gesagt; jetzt sage ich “Müller” /mʏlɐ/.
	Ich habe “Mieter” /miːtɐ/ gesagt; jetzt sage ich “Mitte” /mɪtə/.
	Ich habe “stehlen” /ʃteːlən/ gesagt; jetzt sage ich “Stelle” /ʃtɛlə/.
	Ich habe “Ofen” /oːfən/ gesagt; jetzt sage ich “offen” /ɔfən/.
	Ich habe “kam” /kaːm/ gesagt; jetzt sage ich “Kamm” /kam/.
	Ich habe “Goethe” /ɡøːtə/ gesagt; jetzt sage ich “Götter” /ɡœtɐ/.
	Ich habe “Mut” /muːt/ gesagt; jetzt sage ich “Mutter” /mʊtɐ/.
	Ich habe “Hüte” /hyːtə/ gesagt; jetzt sage ich “Hütte” /hʏtə/.
English	I said “bark” /bɑːk/ first, and now I say “buck” /bʌk/.
	I said “tooth” /tuːθ/ first, and now I say “took” /tʊk/.
	I said “peak” /piːk/ first, and now I say “pick” /pɪk/.
	I said “dark” /dɑːk/ first, and now I say “duck” /dʌk/.
	I said “food” /fuːd/ first, and now I say “foot” /fʊt/.
	I said “beat” /biːt/ first, and now I say “bit” /bɪt/.
Cantonese	我讀”卡” /kʰaːt˥/字, 我讀”咳”/ kʰɐt˥/字。
	我讀”間”/ kan˥/字, 我讀”巾”/ kɐn˥/字。
	我讀”石” /sɛːk˨/字, 我讀”食” /sek˨/字。
	我讀”驚”/kɛːŋ˥/字, 我讀”經”/keŋ˥/字。
	我讀”湯”/t^hɔːŋ˥/字, 我讀”通” /t^hoŋ˥/字。
	我讀”度”/tɔːk˨/字, 我讀”讀”/tok˨/字。

ORCID iDs

Yanjiao Zhu

Peggy Pik Ki Mok

Ethics considerations

This study was approved by the Ethical Assessment Committee Linguistics (ETCL) of the Utrecht Institute of Linguistics OTS (Ethics Code: zhu00021-01-2018) on March 20, 2018. This research was conducted ethically in accordance with the World Medical Association Declaration of Helsinki.

Consent to participate

All participants provided written informed consent prior to participating.

Author contributions

Yanjiao Zhu conceptualized the research, conducted the investigation and analysis, and was responsible for drafting and revising the manuscript. Peggy Mok conceptualized the research, provided supervisory mentorship, and reviewed and edited the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The first author received the PhD studentship and the PhD Student Exchange Programme Award from the Chinese University of Hong Kong for the research and authorship of this article.

References

Adank

Smits

van Hout

(2004). A comparison of vowel normalization procedures for language variation research. Journal of the Acoustical Society of America, 116, 3099–3107.

Amaro

J. C.

Rothman

(2010). On L3 acquisition and phonological permeability: A new test case for debates on the mental representation of non-native phonological systems. IRAL - International Review of Applied Linguistics in Language Teaching. https://doi.org/10.1515/iral.2010.012

Baker

Trofimovich

(2005). Interaction of native- and second-language vowel system(s) in early and late bilinguals. Language and Speech, 48(Pt 1), 1–27. https://doi.org/10.1177/00238309050480010101

Bardel

Falk

(2007). The role of the second language in third language acquisition: The case of Germanic syntax. Second Language Research, 23(4), 459–484. https://doi.org/10.1177/0267658307080557

Barr

D. J.

Levy

Scheepers

Tily

H. J.

(2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68(3), 255–278. https://doi.org/10.1016/j.jml.2012.11.001

Bates

Maechler

Bolker

Walker

(2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1–48. https://doi.org/10.18637/jss.v067.i01

Bauer

Benedict

(1997). Modern Cantonese phonology. Mouton de Gruyter.

Boersma

Weenink

(2015). Praat: Doing phonetics by computer (version 6.3.08). [Computer software]. http://www.praat.org/

Bohn

O.-S.

(1995). Cross-language speech perception in adults: First language transfer doesn’t tell it all. In Strange

(Ed.), Speech perception and linguistic experience: Theoretical and methodological issues in cross-language speech research (pp. 275-300). York Press.

10.

Chao

(1968). A grammar of spoken Chinese. University of California Press.

11.

Chen

Zhang

Mccollum

A. G.

Wayland

(2017). Statistical modelling of phonetic and phonologised perturbation effects in tonal and non-tonal languages. Speech Communication, 88, 17–38. https://doi.org/10.1016/j.specom.2017.01.006

12.

Chen

(2006). Production of tense-lax contrast by Mandarin speakers of English. Folia Phoniatrica et Logopaedica, 58(4), 240–249. https://doi.org/10.1159/000093181

13.

Cheung

H. N. S.

(1972). Xianggang Yueyu yufa de yanjiu (Cantonese as spoken in Hong Kong). The Chinese University Press.

14.

Cheung

(1986). The phonology of present day Cantonese. University College London.

15.

De Bot

Lowie

Verspoor

. (2007). A dynamic systems theory approach to second language acquisition. Bilingualism, 10(1), 7–21. https://doi.org/10.1017/S1366728906002732

16.

Flege

J. E.

(1995). Second language speech learning: Theory, findings, and problems. In Strange

(Ed.), Speech perception and linguistic experience: Issue in cross-language research (pp. 233-277). York Press.

17.

Flege

J. E.

Bohn

O.-S.

(2021). The Revised Speech Learning Model (SLM-r). In Wayland

(Ed.), Second language speech learning: Theoretical and empirical progress (pp. 3-83). Cambridge University Press.

18.

Flynn

Foley

Vinnitskaya

(2004). The Cumulative-Enhancement Model for language acquisition: Comparing adults’ and children’s patterns of development in first, second and third language acquisition of relative clauses. International Journal of Multilingualism, 1(1), 3–16. https://doi.org/10.1080/14790710408668175

19.

Grosjean

(2001). The bilingual’s language modes. In Nicol

J. L.

(Ed.), One mind, two languages: Bilingual language processing (pp. 1-22). Wiley-Blackwell.

20.

Harrington

Hoole

Kleber

Reubold

(2011). The physiological, acoustic, and perceptual basis of high back vowel fronting: Evidence from German tense and lax vowels. Journal of Phonetics, 39(2), 121–131. https://doi.org/10.1016/j.wocn.2010.12.006

21.

Hillenbrand

Getty

L. A.

Clark

M. J.

Wheeler

(1995). Acoustic characteristics of American English vowels. Journal of the Acoustical Society of America, 97(5), 3099–3111. https://doi.org/10.1121/1.411872

22.

Kamiyama

(2007). Acquisition of French vowels by Japanese-speaking learners: Close and close-mid rounded vowels. Paper presented at the L3 Phonology Satellite Workshop of ICPhS XVI.

23.

Kartushina

Martin

C. D.

(2019). Third-language learning affects bilinguals’ production in both their native languages: A longitudinal study of dynamic changes in L1, L2 and L3 vowel production. Journal of Phonetics, 77, 100920. https://doi.org/10.1016/j.wocn.2019.100920

24.

Kohler

(1999). German. In Handbook of the International Phonetic Association (pp. 86–89). Cambridge University Press.

25.

Kopečková

Marecka

Wrembel

Gut

(2016). Interactions between three phonological subsystems of young multilinguals: The influence of language status. International Journal of Multilingualism, 13(4), 426–443.

26.

Kopečková

Gut

Wrembel

Balas

(2023). Phonological cross-linguistic influence at the initial stages of L3 acquisition. Second Language Research, 39(4), 1107–1131. https://doi.org/10.1177/02676583221123994

27.

Kuznetsova

Brockhoff

P. B.

Christensen

R. H. B.

(2017). LmerTest package: Tests in linear mixed effects models. Journal of Statistical Software, 82(13), 1–26.

28.

Lee

T. H.

(1983). The vowel system in two varieties of Cantonese. UCLA Working Papers in Phonetics, 57, 97–114.

29.

Lee

W. S.

Zee

(2003). Standard Chinese (Beijing). Journal of the International Phonetic Association, 33(1), 109–112. https://doi.org/10.1017/S0025100303001208

30.

Lipińska

(2015). Production of L3 Vowels: Is it possible to separate them from L1 and L2 sounds? Research in Language, 13(1), 77–92. https://doi.org/10.1515/rela-2015-0011

31.

Luo

V. G.

Mok

P. P. K.

(2020). The perception of Cantonese vowel length contrast by Mandarin speakers. Language and Speech, 63(3), 635–659. https://doi.org/10.1177/0023830919879471

32.

Matuschek

Kliegl

Vasishth

Baayen

Bates

(2017). Balancing Type I Error and Power in Linear Mixed Models. Journal of Memory and Language, 94, 305–315. https://doi.org/10.1016/j.jml.2017.01.001

33.

McAllister

Flege

J. E.

Piske

(2002). The influence of L1 on the acquisition of Swedish quantity by native speakers of Spanish, English and Estonian. Journal of Phonetics, 30(2), 229–258. https://doi.org/10.1006/jpho.2002.0174

34.

Norman

Prior

Degani

(2025). Contribution of prior linguistic knowledge to L3 phonological perception and production. Journal of Memory and Language, 141(March 2024). https://doi.org/10.1016/j.jml.2024.104600

35.

Nimz

(2016). Sound perception and production in a foreign language. Universitätsverlag Potsdam.

36.

Guion-Anderson

Aoyama

Flege

J. E.

Akahane-Yamada

Yamada

(2011). A one-year longitudinal study of English and Japanese vowel production by Japanese adults and children in an English-speaking setting. Journal of Phonetics, 39(2), 156–167. https://doi.org/10.1016/j.wocn.2011.01.002

37.

Patience

Qian

(2022). The role of task complexity and dominant articulatory routines in the acquisition of L3 Spanish. Languages, 7(2). https://doi.org/10.3390/languages7020090

38.

Predeck

(2022). Vowel length in German: Use of quality and quantity in the perception of long and short vowels in German [UC Davis]. https://doi.org/ProQuestID:Predeck_ucdavis_0029D_21215

39.

R Core Team. (2024). R: A language and environment for statistical computing (version 4.4.1). R Foundation for Statistical Computing.

40.

Roach

(2004). British English: Received pronunciation. Journal of the International Phonetic Association, 34(2), 239–245. https://doi.org/10.1017/S0025100304001768

41.

Rothman

(2011). L3 syntactic transfer selectivity and typological determinacy: The typological primacy model. Second Language Research, 27, 107–127. https://doi.org/10.1177/0267658310386439

42.

Rothman

(2015). Linguistic and cognitive motivations for the Typological Primacy Model (TPM) of third language (L3) transfer: Timing of acquisition and proficiency considered. Bilingualism: Language and Cognition, 18(2), 179–190. https://doi.org/10.1017/S136672891300059X

43.

Seo

Dmitrieva

(2024). L2 cross-linguistic influence on L1 perception: Evidence from heritage speakers and long-term immigrants. Journal of Phonetics, 104, 101314. https://doi.org/10.1016/j.wocn.2024.101314

44.

Slabakova

(2017). The scalpel model of third language acquisition. International Journal of Bilingualism, 21(6), 651–665. https://doi.org/10.1177/1367006916655413

45.

Sypianska

(2016). L1 vowels of multilinguals: The applicability of SLM in multilingualism. Research in Language, 14(1), 79–94. https://doi.org/10.1515/rela-2016-0003

46.

Tomaschek

Truckenbrodt

Hertrich

(2015). Discrimination sensitivities and identification patterns of vowel quality and duration in German / u / and / o / instances (January).

47.

Wassink

A. B.

Wright

R. A.

Franklin

A. D.

(2007). Intraspeaker variability in vowel production: An investigation of motherese, hyperspeech, and Lombard speech in Jamaican speakers. Journal of Phonetics, 35(3), 363–379. https://doi.org/10.1016/j.wocn.2006.07.002

48.

Westergaard

Rodina

(2017). Crosslinguistic influence in the acquisition of a third language: The Linguistic Proximity Model. International Journal of Bilingualism, 2(6), 666–682. https://doi.org/10.1177/1367006916648859

49.

Wiener

Goss

(2019). Second and third language learners’ sensitivity to Japanese pitch accent is additive. Studies in Second Language Acquisition, 41(4), 897–910. https://doi.org/10.1017/S0272263119000068

50.

Yue-Hashimoto

A. O.

(1972). Cantonese (the Yue dialects I). Cambridge University Press.

51.

Zee

(1999). Chinese (Hong Kong Cantonese). In Handbook of the International Phonetic Association (pp. 58-65). Cambridge University Press.

Cross-Linguistic Influences on the Production of Third Language German Vowel Length Contrasts by Cantonese–English Bilingual Learners

Abstract

Keywords

1 Introduction

1.1 Vowel Length Features in German, English and Cantonese

1.2 The Acquisition of L3 Vowel Length Contrasts

1.3 The Acquisition of L2 Vowel Length Contrasts

1.4 The Present Study

2 Method

2.1 Participants

2.2 Materials

2.3 Procedures

2.4 Analysis

3 Results

3.1 An Overview of the Long- and Short-Vowel Qualities

3.2 Group Comparison in Duration and Quality Contrasts of Vowel Length Pairs

3.3 CLIs and Individual Variations

4 Discussion

4.1 The Realization of L1, L2, and L3 Vowel Length Pairs in Duration and Quality

4.2 CLIs Among the L1, L2, and L3 Vowel Systems

5 Conclusion

Footnotes

Appendix

ORCID iDs

Ethics considerations

Consent to participate

Author contributions

Funding

References