Tonal language experience facilitates pitch perception in L2/3 Japanese

2.1 Tonal advantage in Lx perception

Tonal L1 speakers often display an advantage in perceiving pitch cues over speakers of languages where pitch is primarily used for intonation. While the origin of this advantage is debated, it has been observed in crosslinguistic studies on various target languages (e.g., Kaan et al. 2008; Wu et al. 2012). Two prominent models of speech learning posit that L2 sounds are filtered through L1 phonological knowledge. The Speech Learning Model (SLM) focuses on how listeners categorize L2 speech segments on the basis of L1 categories (Flege 1995). Its revised form, the SLM-r, maintains the emphasis on sound categorization but shifts its focus to how cue weighting patterns develop between L1 and L2 (Flege et al. 2021).

The Perceptual Assimilation Model (PAM) also treats the L1 as a filter for L2 sounds but considers the goodness-of-fit of phonological categories between L1 and Lx as the primary driver of perceptual behavior (PAM: Best 1995; Best and Tyler 2007). That is, listeners are better at discriminating or categorizing Lx tones when they assimilate to existing tonal or intonational categories in the L1 in a one-to-one manner, and worse when the mapping is many-to-one. Of the two models, only PAM has been applied to suprasegmentals (PAM for Suprasegmentals (PAM-S); So and Best 2014). Prior studies broadly framed in terms of category assimilation of suprasegmentals have found cross-linguistic assimilation in the perception of Mandarin and Cantonese tones (So and Best 2010; Wu et al. 2014), as well as English stress (Chan 2008). Categorical assimilation has also been reported in the perception and production of Japanese pitch accent (e.g., Ayusawa 2003; Hirano-Cook 2011; Laméris and Graham 2020; Shport 2016a).

In contrast, other studies have approached the question of tonal advantage through the lens of functional load of tone in the L1 (e.g., Surendran and Niyogi 2006). In other words, the greater the role that lexical tones play in word selection in the L1, the more weight that listeners tend to give these prosodic cues when perceiving tonal contrasts in a new language. While never formalized as a single model, functional load-driven advantage has been shown when the target language is tonal (Mandarin, Cantonese (Chang et al. 2017; Choi and Chiu 2023); Thai (Chan and Leung 2020; Wayland and Guion 2004)), pitch accent (Wiener and Goss 2019), or nontonal (i.e., stress or intonation perception; Choi et al. 2019). To be certain, the acoustic properties of tones themselves in the target language play a role in perceptual accuracy (Tong et al. 2015; Zhu et al. 2023), and the tonal advantage is not universal (Cooper and Wang 2012; Francis et al. 2008), but a substantial body of evidence points toward the facilitative role of tonal L1.

Behavioral studies have operationalized the functional load of F0 by grouping listeners (e.g., Laméris et al. 2023; Schaefer and Darcy 2014) by tonal status in their L1 (e.g., high, medium, and low), or by calculating informativeness of F0 based on corpus estimates of pitch, stress, and tone minimal pairs (Shibata and Shibata 1990; Wiener and Goss 2019). Here, we also assume the functional view that experience with L1/Lx cues is a primary driver of facilitative transfer of pitch in a new language.

In the current study, we address our first research question by comparing listeners from two tonal languages – Mandarin and Cantonese – on their perception of Japanese pitch accent. To our knowledge, no prior studies have compared two groups of tonal (Sinitic) language speakers on their perception of Japanese pitch accent. To refine the notion of facilitative transfer – that tonal language speakers can perceive unfamiliar pitch cues as well as or better than target language L1 speakers – we compare L1 speakers of these two languages, where pitch is highly informative, but nonetheless differs in tone inventory (i.e., six Cantonese tones vs. four Mandarin tones) and functional load. Specifically, Cantonese listeners may be better at Japanese pitch discrimination because of the slightly higher functional load of tone than in Mandarin (Oh et al. 2013) or its larger tone inventory (for evidence on tone inventory, see Hu et al. 2020). Here, we treat functional load and tone inventory as indexing the same construct. That is, we assume that the larger the lexical tone inventory a language possesses, the greater the functional load that tone carries in the language. This aligns with Wiener and Goss’ (2019) minimal pair-based index of functional load, which they termed informativeness of F0, as well as other information-theory approaches to phonological contrasts (e.g., Oh et al. 2013; Surendran and Niyogi 2006).

Conversely, from the vantage of feature assimilation (i.e., PAM-S), Mandarin tone 2 (rising F0) would assimilate to the Japanese LHH (unaccented) pitch pattern over a three-mora unit, while Mandarin tone 4 (falling F0) would correspond to the H*LL (initial accent) pattern. Cantonese features two rising tones (tones 2 and 5) and a falling tone (tone 4), in addition to three level tones (Yip 2002). However, neither language contains a tone category that resembles the Japanese LH*L (medial accent) pattern. Thus, given the presence of rising and falling tone categories in both languages, the two groups would perform similarly under an assimilation account.

2.2 Lx learning-derived tonal advantage

Tonal advantage in F0 perception can also be an outcome of Lx learning. Few studies have attempted to depict the complex, yet very real, picture of how learning additional languages influences L3 perception (for cumulative enhancement in morphosyntax, see Flynn et al. 2004). For nontonal L1 speakers, learning a tonal Lx can enhance perceptual accuracy for F0 cues in an unfamiliar language. For instance, Qin and Jongman (2016) found that L1 English learners of L2 Mandarin were more sensitive to certain Cantonese tone contrasts than L1 English monolinguals. Likewise, a recent replication of this study with L1 Korean speakers found that those who had learned L2 Mandarin could discriminate Cantonese tones on par with L1 Mandarin speakers (Qin et al. 2024).

For Japanese pitch perception, evidence exists for both category-specific assimilation, which would align with PAM’s predictions, as well as general enhancement resulting from Lx classroom learning of a tonal language. For example, L1 English-L2 Mandarin learners may assimilate Japanese pitch patterns (particularly the falling, H*-L pattern) to a Mandarin tonal category (tone 4, falling F0) after only one semester of Mandarin study, leading to higher accuracy on the discrimination of patterns that resemble Mandarin tones (Wiener and Goss 2018). On the other hand, a follow-up study also involving classroom learners and using the same Japanese stimuli found that L1 English listeners’ sensitivity to Japanese pitch accent was best predicted by informativeness of F0 in their L2 (Mandarin), rather than Japanese-specific experience (Wiener and Goss 2019). That is, these L1 English-L2 Mandarin learners discriminated all Japanese pitch patterns in the stimuli (HLL, LHL, LHH) equally well, while outperforming L2 Japanese learners who had classroom exposure to these patterns. This result suggests that category assimilation may not be the primary driver of perceptual behavior, mirroring recent findings that show extralinguistic factors such as non-linguistic tone processing ability (Laméris et al. 2023), or acoustic properties of tone contours themselves (Shport 2016b), contributing more to cross-linguistic tone perception than category assimilation.

However, the question remains as to whether the differences exhibited by tonal L2 learners in these studies were the result of facilitative tonal L2 to Lx transfer, or whether classroom input in any L2 might have triggered the cue reweighting (Wang et al. 1999), and thus greater sensitivity to Japanese pitch patterns. Two pieces of evidence motivate this question. First, L1 speakers of languages with rich intonational systems, but limited lexical prosody, such as German, show greater sensitivity to Mandarin tone mismatches than both L1 Japanese and French speakers, languages with fewer intonational shapes than German (Braun et al. 2014). This suggests that L2 German classroom learners may be similarly advantaged in their perception of Japanese pitch. German prosodically resembles English in that both are stress-accented languages with a tendency toward initial syllable stress. Both languages also display a similar range of pitch shapes at the utterance level. For example, both languages use falling pitch in statements and rising pitch in yes/no questions. Statement intonation may approximate the Japanese HLL accent pattern, and question intonation the rising LHH pattern. Yet in laboratory tasks German speakers are more sensitive to rising intonation than English speakers, perhaps because of their greater attention to F0 movement before a tonal accent (Kember et al. 2017). This heightened attention to rising intonation could potentially transfer to language-like tasks as well, giving German learners an advantage on the LHH (unaccented) pitch pattern in Japanese.

Second, classroom input does not necessarily reflect normal distributional patterns of language (Ellis 2002), including the “frequency dimension” of intonation (Mennen 2015). This allows for the possibility that classroom input on intonational features in a language like German could enhance learners’ perception of pitch or tone in a new language (but see Thomson and Derwing (2015) for discussion of L2 pronunciation instruction).

Because the previous design (in Wiener and Goss 2019) did not account for the possibility that classroom practice with the intonational features of any Lx may increase sensitivity to F0, we include a group of nontonal L1 English-L2 German learners, who were undergoing classroom German training at the time of the study. Thus, our second research question asks whether learning any additional language results in enhanced pitch discrimination ability or if the Lx learning must be tonal. To test this, we compare L1 English speakers currently studying three L2s – Mandarin, Japanese and German.

2.3 Proficiency and pitch perception

The final research question pertains to the role of Japanese-specific experience with pitch cues. We ask – do learners increase their sensitivity to pitch cues as their Japanese ability increases, and does this experience with Japanese pitch cues bring them into parity with L1 tonal speakers? The question of target-language proficiency is relevant to the current study because the notion of facilitative transfer predicts that tonal L1 status would continue to outweigh Japanese-specific experience with pitch, at least in pre-lexical pitch (i.e., nonword) discrimination. If cue-specific experience enables Japanese learners to close the gap with L1 tonal speakers, then L1 tonal status may not have such a privileged position in pitch perception.

Findings are mixed as to whether Japanese learners from nontonal backgrounds improve their pitch perception ability in step with overall proficiency gains. Some studies report gains in accuracy (Wu et al. 2017), yet others do not, even among learners above the beginner level (Shport 2008). Wiener and Goss (2019) found tonal language experience to be the best predictor of sensitivity to Japanese pitch patterns, but did not include L1 English speakers who had attained advanced L2 Japanese proficiency. Here, we compare L2/3 Japanese learners from both tonal and non-tonal backgrounds at two proficiency levels (beginner and advanced) on their pitch discrimination ability. This comparison enables us to see whether Japanese learner proficiency modulates pitch discrimination ability irrespective of the L1.

3.1 Participants

Nine participant groups (15 participants per group; n = 135) took part in the experiment (Table 1). Of the nine groups, four (n = 60) were newly recruited for this study. Data from the remaining five groups (n = 75) were incorporated from Wiener and Goss (2019), which followed identical procedures to the current study. All stimuli, data, and R code are available on the Open Science Framework (https://osf.io/9328s/).

In this study, we define L1 as the first language a speaker acquired as an infant. L2 and L3 are defined as languages learned after the L1 was already in place, with the L3 being learned after L2 was acquired (De Angelis 2007). “Native speaker” status was not asked as recent studies have shown it can be ambiguous and mean different things for different populations of speakers (Brown et al. 2023). Non-L1-English participants were matched on their L2 English proficiency by self-reported abilities and standardized English proficiency test scores (Test of English as a Foreign Language (TOEFL) or International English Language Testing System (IELTS). L1-Mandarin participants reported only speaking standard Mandarin and no other tonal dialect. L1-Cantonese speakers were from Hong Kong, were Cantonese-dominant, and reported basic Mandarin ability. The L1-English group not learning or exposed to an L2 were functionally monolingual and had 2 years or less of secondary school instruction in Spanish, French, Latin, or German.

For Lx learners, L1-English groups engaged in L2-Mandarin, L2-Japanese, or L2-German classroom learning were controlled for their respective level of L2 proficiency based on self-reported abilities, length of L2 study, and placement in comparable university courses. No participant had studied abroad in the target country. L1-Mandarin + L3-Japanese learners were recruited from an advanced Japanese class in the US and screened by placement test scores, which mainly measured grammatical and lexical knowledge, but also included a short aural comprehension task. L1-Cantonese + L3 Japanese learners were enrolled in a 1st-year Japanese class at a university in Hong Kong.

Participants were screened for their nonlinguistic pitch perception using an adaptive pitch test (Mandell 2018) or frequency discrimination test (Lutman 2004). In both tasks, listeners heard a 500 Hz reference tone followed by a tone that differed by hertz intervals. Listeners judged by button press whether the second tone was of higher or lower pitch than the first tone. The between-stimulus frequency difference either increased or decreased depending on the accuracy of participants’ responses. All 135 participants were able to reliably differentiate two tones at a 20 Hz interval or lower (range: 1.4–20 Hz). Whereas any cutoff is arbitrary, 20 Hz provided us with sufficient power and screened for potential congenital amusia (e.g., Zhu et al. 2023).

3.2 Materials

We employed an ABX discrimination task with Japanese-like nonwords. Ten nonwords with a segmental structure of CV.CV.CV were used as stimuli (see our Open Science Framework repository for all items and acoustic measurements). Each trimoraic stimulus was recorded in a carrier sentence by a female native speaker of Tokyo Japanese with three existent pitch accent patterns: HLL, LHL, and LHH. Nonwords were then extracted in Praat (Boersma and Weenink 2020). The nonword tazana, for example, was produced as TAzana (HLL), taZAna (LHL), and taZANA (LHH). Acoustic parameters for each category aligned with accent categories used in previous studies (e.g., Cutler and Otake 1999; Shport 2015, 2016a) and, besides volume normalization, were presented to listeners without phonetic manipulation. The ABX stimuli consisted of three consecutive trimoraic nonwords with a 250 ms interstimulus interval. The first two sounds, A and B, only differed in pitch accent pattern (i.e., HLL, LHL, or LHH). Participants were asked to decide whether the third sound, X, matched the pitch pattern of the first (A) or second (B) sound. Participants completed a total of 120 trials (12 ABX patterns per target × 10 targets). See Appendix for a list of the nonword stimuli.

3.3 Procedures

Participants were instructed to indicate by button press whether the final sound (X) was similar to the first (A) or second (B) sound as quickly and accurately as possible. If participants did not respond within 2.5 s from the offset of X, the next trial would proceed automatically. AB ordering was counterbalanced across all trials with a 1 s intertrial interval. Two practice trials were presented using a practice stimulus of similar moraic structure and accent pattern as the stimuli. Stimuli were presented to listeners via headphones using either Superlab 5 or Eprime 2.0 software.

3.4 Statistical analysis

All analyses and visualizations were carried out in R (version 4.4.0; R Core Team 2024). Trials in which participants timed out (n = 263; <2 %) were first removed. Accuracy of the ABX trials (coded as 1 correct; 0 incorrect) was modeled using the lme4 package (Bates et al. 2015). A generalized linear mixed effects model (with a logit link function and the “bobyqa” optimizer) was built with items and participants as random intercepts. The independent variables included pitch accent pattern type as a three-level categorical variable (with LHH as the reference) and listener group, which changed given the research question with pairwise comparisons and adjusted p-values obtained using estimated marginal means (emmeans package; Lenth 2021). For RQ1, the L1 Cantonese and L1 Mandarin groups were compared to one another with the L1 Cantonese group as the reference level. For RQ2, the monolingual L1 English group was coded as the reference level and the three beginner L2 groups (L1 English-L2 Mandarin (Beg), L1 English-L2 German (Beg), L1 English-L2 Japanese (Beg)) were compared. For RQ3, four groups studying Japanese as an L2/L3 were examined. These groups were Beginner/Advanced pairs. The L1 English-L2 Japanese (Beg) was coded as the reference level and the L1 English-L2 Japanese (Adv), L1 Mandarin-L3 Japanese (Beg), and L1 Mandarin-L3 Japanese (Adv) were compared.

4.1 Research question 1: listeners with different tonal L1s do not show differences in Japanese pitch perception

Figure 1 shows the results from the two tonal L1 groups. Both groups performed similarly as indicated by their respective 95 % confidence intervals: L1 Cantonese [0.90, 0.95], L1 Mandarin [0.90, 0.94]. The logistic regression model, which had moderate explanatory power (conditional R ² = 0.22), revealed no difference between the two groups (β = −0.01, Z = −0.01, SD = 0.29, p = 0.99). Neither a main effect of pitch accent type nor a two-way interaction with accent type and listener group was found with an alpha-level of 0.05 indicating the two groups performed similarly across the three pitch accent types.

Figure 1:

ABX discrimination accuracy for L1 Cantonese and L1 Mandarin participants. The figure shows each participant’s mean (points), and group box plots and density plots.

4.2 Research question 2: only learning an additional tonal language improves Japanese pitch perception relative to the non-learning control group

Figure 2 shows the results from the four L1 English groups. Groups performed differently as indicated by their respective 95 % confidence intervals: L1 English [0.76, 0.79], L1 English-L2 German (Beg) [0.77, 0.81], L1 English-L2 Japanese (Beg) [0.83, 0.86], L1 English-L2 Mandarin (Beg) [0.85, 0.88]. The logistic regression model, which had moderate explanatory power (conditional R ² = 0.22), revealed that relative to the L1 English group as the reference level, the L1 English-L2 German (Beg) performed similarly (β = 0.08, Z = 0.32, SD = 0.25, p = 0.98), the L1 English-L2 Japanese (Beg) performed similarly (β = 0.49, Z = 1.92, SD = 0.25, p = 0.22) and the L1 English-L2 Mandarin (Beg) performed significantly better (β = 0.72, Z = 2.81, SD = 0.25, p = 0.03). Estimated marginal means indicated that all three groups learning an L2 performed similarly (ps > 0.05). Overall, no difference in pitch accent types was found (ps > 0.05), however, a two-way interaction was found between pitch accent types and listener group, which indicated that for the HLL stimuli the L1 English group and L1 English-L2 Mandarin (Beg) group differed (β = 1.14, Z = 3.85, SD = 0.30, p = 0.007). In other words, the difference observed between these two groups was driven by performance on the HLL words only; LHH and LHL behavior did not differ (ps > 0.05).

Figure 2:

ABX discrimination accuracy for four different L1 English groups. The figure shows each participant’s mean (points), and group box plots and density plots.

4.3 Research question 3: improved learner proficiency only helps tonal L1 listeners’ Japanese pitch perception

Figure 3 shows the results from the four L2/L3 Japanese learner groups. Groups performed differently as indicated by their respective 95 % confidence intervals: L1 English-L2 Japanese (Beg) [0.83, 0.86], L1 Mandarin-L3 Japanese (Beg) [0.88, 0.91], L1 English-L2 Japanese (Adv) [0.84, 0.87], L1 Mandarin-L3 Japanese (Adv) [0.95, 0.96]. The logistic regression model, which had substantial explanatory power (conditional R ² = 0.28), revealed that relative to the L1 English-L2 Japanese (Beg) group as the reference level, the L1 English-L2 Japanese (Adv) performed similarly (β = 0.15, Z = 0.57, SD = 0.26, p = 0.94). The L1 Mandarin-L3 Japanese (Beg) performed similarly to the L1 English-L2 Japanese (Beg) group (β = 0.52, Z = 1.98, SD = 0.26, p = 0.19). The L1 Mandarin-L3 Japanese (Adv) outperformed the L1 English-L2 Japanese (Beg) group (β = 1.46, Z = 5.31, SD = 0.27, p < 0.001). Estimated marginal means indicated that the L1 Mandarin-L3 Japanese (Adv) outperformed the L1 Mandarin-L3 Japanese (Beg) (β = 0.94, Z = 3.38, SD = 0.28, p = 0.004). The L1 Mandarin-L3 Japanese (Adv) outperformed the L1 English-L2 Japanese (Adv) group (β = 1.31, Z = 4.77, SD = 0.28, p < 0.001). The L1 Mandarin-L3 Japanese (Beg) performed similarly to the L1 English-L2 Japanese (Adv) group (β = 0.37, Z = 1.42, SD = 0.26, p = 0.49). Overall, no difference in pitch accent types was found (ps > 0.05). A two-way interaction was found between pitch accent types and listener group for all three pitch accent types for differences found between L1 English-L2 Japanese (Beg) and L1 Mandarin-L3 Japanese (Adv) and for L1 English-L2 Japanese (Adv) and L1 Mandarin-L3 Japanese (Adv) (ps < 0.05). In other words, the differences were consistent across all three pitch accent types. Estimated marginal means of the interactions revealed that the L1 Mandarin-L3 Japanese (Adv) and L1 Mandarin-L3 Japanese (Beg) differences were null (ps > 0.05) when examined individually by pitch accent type. That is, the difference between the two L1 Mandarin groups with different L3 Japanese proficiencies was the result of an aggregate effect rather than a specific pitch accent type. In contrast, the differences between the L1 Mandarin-L3 Japanese (Adv) group and the two L1 English groups with different L2 Japanese proficiencies were consistent differences across all three pitch accent types.

Figure 3:

ABX discrimination accuracy for four different Lx Japanese learner groups. The figure shows each participant’s mean (points), and group box plots and density plots.