Investigating the role of musical experience in lexical tone perception: non-musicians and amateur musicians’ perception of Mandarin tones

1 Introduction

Different languages use pitch differently. All languages use pitch variation to convey different kinds of information, including linguistic information (Ladefoged and Johnson 2014: 264–266). In tonal languages like Mandarin, pitch variation (F0) is used to distinguish lexical meanings through the use of contrastive tones (Yip 2002: 229). For example, Mandarin’s four tones – high-level (Tone1), rising (Tone2), low-dipping (Tone3), and falling (Tone4) – can be used to assign distinct meanings to otherwise identical syllables (Ladefoged and Johnson 2014: 264–266; Pfordresher and Brown 2009). In contrast, non-tonal languages like English utilize pitch prosodically (i.e., at the utterance-level) to signal pragmatic rather than lexical meaning (Cruttenden 1997; Wennerstrom 2001). Consequently, speakers of non-tonal languages often struggle to perceive lexical tones (Krishnan et al. 2005).

Pitch also plays an important role in music (e.g., McDermott and Oxenham 2008), and extensive evidence from both behavioural and neurophysiological research suggests that non-tone language native speakers with music training may be able to transfer their ability with pitch in music to perceive lexical tone in a second language (L2; Marie et al. 2011). Neuroimaging studies have shown that there are both structural (Hyde et al. 2009) and functional (Lappe et al. 2008) brain differences between musicians and non-musicians. For example, musicians are better at tracking linguistic pitch changes than non-musicians; musicians’ brainstem responses show more faithful representation of the F0 contours and more robust neural phase-locking for Tone2 and Tone3, the most acoustically complex tone (Wong et al. 2007). Both professional and amateur musicians consistently outperform non-musicians in lexical tone discrimination, even without prior tone language experience (Alexander et al. 2005; Marie et al. 2011; Wong and Perrachione 2007). Likewise, musicians who have been playing for longer have been shown to be better at identifying lexical tone than those with fewer years’ training (Chua and Brunt 2014).

One explanation for why musicians are better than non-musicians at perceiving tone languages is that through musical practice, they have developed enhanced pitch sensitivity vis-à-vis non-musicians. Evidence suggests that musical pitch perception shares neural and cognitive mechanisms with linguistic pitch processing (Bidelman et al. 2011; Krishnan et al. 2005), and that musical training thus enhances pitch discrimination across domains (Moreno et al. 2009; Schön et al. 2004). This implies that domain-general pitch acuity is enhanced by formal music training, and that their increased ability to track pitch direction (i.e., onset-to-offset contour comparisons) facilitates tone perception.

Another potential explanation is that in addition to pitch processing, music training enhances general linguistic and cognitive abilities and that these in turn, support L2 lexical tone perception. For example, musicians have been shown to outperform non-musicians in phonological categorisation (Chobert and Besson 2013; Kraus and Chandrasekaran 2010) and music training has been shown to enhance segmental (Degé and Schwarzer 2011) and suprasegmental (Moreno et al. 2009) phonological processing. However, as far as we know, no study has investigated whether the influence of music training on L2 lexical tone perception is mediated by general segmental speech processing ability. Previous research has found that English listeners treat tones and consonants as perceptually separable (Lin and Francis 2014), and that a listener’s first language (L1) can play a vital role in attention distribution when processing a tonal second language (Liu and Ning 2021). As segmental and tonal information are presented simultaneously in a Mandarin word, enhanced segmental processing ability may allow a listener to attribute more attention to processing tonal information, in turn leading to better tone identification.

Music training may also lead to better cognitive processing as measured in terms of verbal working memory (e.g., Talamini et al. 2017). Both music performance (requiring memorisation of phrases during execution) and L2 tone perception (requiring retention of unfamiliar segmental and suprasegmental features) rely on working memory (e.g., Bidelman et al. 2013), and so music training may support L2 tone perception through enhancing verbal or melodic working memory. For example, Jaschke et al. (2018) found that children who were musically trained showed significant gains in verbal working memory. Additionally, Perrachione et al. (2011) tested the working memory capacity and Mandarin-like tone perception ability of native English speakers with no prior exposure to tone languages. Both studies found that learners with better working memory were faster and more accurate in tone categorisation, likely because they can better maintain and manipulate tonal information in memory.

However, whilst an explanation based on musical training might be appealing, individuals without formal music training can also exhibit highly developed pitch sensitivity (e.g., Mankel and Bidelman 2018). This suggests that advanced pitch discrimination may not require musical expertise or extensive music training. Individuals with no musical training but high baseline pitch aptitude (tested via the Montreal Battery of Evaluation of Amusia, MBEA) matched musicians in identifying L2 lexical tones, indicating that innate pitch sensitivity can offset lack of formal training (Delogu et al. 2010). This indicates that music training is not necessary to enhance perception of pitch in speech (Bidelman and Alain 2015).

Similarly, while music training enhances rhythm perception accuracy and melodic memory for unfamiliar songs (Chen et al. 2020; Matthews et al. 2016), untrained individuals can also exhibit exceptional musical perceptual abilities (Correia et al. 2022). Non-musicians with strong melodic contour identification skills (e.g., detecting rising/falling pitch patterns) performed comparably to musicians in discriminating Mandarin tones (Tone2 vs. Tone3), and neural evidence showed that these individuals had enhanced brainstem encoding of pitch contours (FFR) (Giuliano et al. 2011). Such evidence indicates that innate musical ability, rather than music training itself, may facilitate non-native lexical tone perception.

Notably, the direct or indirect influence of music training on Mandarin tone perception appears tone-specific (e.g., Delogu et al. 2010). For instance, Wong et al. (2007) measured frequency-following responses (FFRs) in non-tone-language-speaking musicians and non-musicians as they passively listened to Mandarin /mi/ synthesized with Tones 1, 2, and 3. While musicians exhibited more faithful neural tracking of F0 contours overall, this advantage was significant for Tone3 and Tone2. Moreover, enhanced neural phase-locking in musicians was observed solely for Tone3, with no group differences in phase-locking across tones. Although the exclusion of Tone4 arguably precludes conclusions about the effect of music training on the full Mandarin tone inventory, these findings suggest that musicians’ superior tone perception may not generalise uniformly across all Mandarin tones.

To sum up, music training enhances a broad range of auditory and cognitive abilities, including pitch discrimination, musical ability, phonological processing and working memory, all of which could collectively support L2 tone perception. However, existing research seldom disentangles the contributions of music training, musical ability, and general pitch processing in Mandarin tone perception. Given that individuals without formal music training can exhibit strong musical ability and perform well in tone perception tasks, it remains unclear whether observed performance advantages in musicians reflect training-induced plasticity or pre-existing auditory processing abilities that may exist independently of training (Mankel and Bidelman 2018; Wong and Perrachione 2007). To fill this gap, this study investigated how non-professional music training and related cognitive abilities influence Mandarin tone identification in native, monolingual English speakers with no prior experience of Mandarin or any tone language. Participants completed a battery of tasks assessing pitch discrimination, verbal/melodic working memory, and L1/L2 phonological processing. Musical experience was measured using the Gold-MSI questionnaire (Müllensiefen et al. 2014). We hypothesised that music training would improve tone identification, mediated by the abilities that it can potentially enhance, that Tone3 (dipping contour) would be the most challenging to identify and that this tone would show the strongest dependence on musical ability, consistent with its acoustic complexity.

2 Methods

3 Results

Participants’ performance in all 6 tasks and the Gold-MSI Questionnaire are plotted in Figure 1.

Figure 1:

Boxplots of all observed variables: Tone Identification (ToneID), reading span (Reading span), melodic memory (Melodic memory), L1 consonant categorisation (L1), L2 consonant discrimination (L2), Pitch Discrimination (PD), Musical Ability (Musical ability), and Music Training (Music training). Melodic memory and L1 scores were Min–Max normalised to match the scales of the other measures for visual comparison. The scores of the other tasks are percentage correct.

To investigate if performance on the different tasks predicted tone identification performance, path analysis was used. To determine the factors for the path analysis, a correlation analysis was carried out first. Given that the scores for reading span, Pitch Discrimination, Musical Ability, and Music Training were not normally distributed, a Spearman correlation was used to investigate potential correlations between ToneID and all other tasks, i.e., predicting variables. The Benjamini-Hochberg procedure was applied to adjust for multiple comparisons. Results showed that ToneID was significantly correlated with Pitch Discrimination (mean = 83.44, SD = 15.21; ρ = 0.56, p < 0.05), Musical Ability (mean = 46.85, SD = 8.06; ρ = 0.41, p < 0.05), and Music Training (mean = 21.56, SD = 10.73; ρ = 0.46, p < 0.05), but not with the other measures (ρ = 0.04 to 0.36, p > 0.05; see Figure 2 for a summary of results). Therefore, these three variables were used in the path analysis to further investigate their relationship with tone identification. Significant correlations are illustrated in the scatterplots shown in Figure 3.

Figure 2:

Results of the Spearman correlations between all 8 variables – Tone Identification (ToneID), Reading span, Melodic memory, L1 consonant categorisation (L1), L2 consonant discrimination (L2), Pitch Discrimination (PD), Gold-MSI music training score (Music training) and Gold-MSI musical ability score (Musical ability). Correlation coefficients are given in each box and non-significant correlations marked with a cross.

Figure 3:

Scatterplots for all significant correlations. Each datapoint (i.e., a participant) is represented by a black dot and the linear regression line is shown in red.

A path model (Figure 4) was fitted with Pitch Discrimination, Musical Ability, and Music Training as continuous variables to predict performance on the ToneID task, i.e., tone identification ability. The multivariate normality assumption was assessed using the Henze-Zirkler test given the moderately small but sufficient sample size. Due to the violation of the multivariate normality assumption, maximum likelihood estimation with robust standard errors and the Satorra-Bentler corrected test statistic were used (Barbeau et al. 2019; Savalei 2019). The model exhibited a good fit (χ²(1) = 0.01, p = 0.92; Standardized Root Mean Square Residual [SRMR] = 0.004; Robust Root Mean Square Error of Approximation [RMSEA] = 0.00; Robust Comparative Fit Index [CFI] = 1.00). Although the RMSEA and CFI suggested potential of the model overfitting the data due to the model’s minimal complexity (df = 1), the non-significant chi-square value, lower than 0.08 SRMR indicated a good fit of the mode, and the strong theoretical justification supported the model’s reliability and generalizability.

Figure 4:

Path model using scores from the ToneID task (TID), Pitch Discrimination (PD) and Musical Ability (MsA) as measured using the Gold-MSI as endogenous variables (whose variance is dependent on at least one of the other variables in the model). Gold-MSI Music Training score (MsT) is the exogenous variable, meaning that its variance is not dependent on any other variable in the model. The paths in the model are directional. The significance-level of the path coefficients is shown as follows; p < 0.0001 ‘***’, p < 0.001 ‘**’, p < 0.01 ‘*’, p < 0.05.

The path coefficients showed that Pitch Discrimination (β = 0.37, p < 0.05) and Musical Ability (β = 0.27, p < 0.05), but not Music Training (β = 0.17, p = 0.28), had a significant direct effect on Tone Identification. However, Music Training appeared to significantly influence Musical Ability (β = 0.46, p < 0.05) and Pitch Discrimination (β = 0.46, p < 0.05), and had a significant indirect influence on Tone Identification, with Pitch Discrimination (β = 0.17, p < 0.05) and Musical Ability (β = 0.12, p < 0.05) both acting as mediating variables.

As displayed in Figure 5, participants were able to identify the tones at better than chance level. Although there were some differences in how they performed with individual tones, these differences were relatively small. Tone1 was the best identified (mean = 53.33, SD = 28.50) and was identified slightly more accurately than Tones 2 and 3, both of which were identified with a similar level of accuracy (Tone2, mean = 48.72, SD = 24.62; Tone3, mean = 50.00, SD = 25.24). Tone1 was most commonly confused with Tone2 (mean = 33.33, SD = 24.85) though Tone2 itself was rarely confused with Tone1 (mean = 12.05, SD = 14.54) and instead, was most commonly confused with Tone3 (mean = 25.38, SD = 17.30). Tone3 was most commonly confused with Tone4 (mean = 18.97, SD = 14.65). The least well-identified tone was Tone4 (mean = 42.82, SD = 29.29) which was misidentified most frequently as Tone1 (mean = 33.08, SD = 22.38) or 2 (mean = 16.41, SD = 17.24). A one-way repeated measures ANOVA with the tone type as the independent variable and the percentage of correct identification of each tone as the dependent variable was used to investigate if there were any significant differences in identification of individual tones. Results showed no significant differences in identification of the four tones [F(3,152) = 1.03, p > 0.05].

Figure 5:

Confusion matrix of the stimuli tone (i.e., Stimulus) and participants’ responses (i.e., Response) in tone identification task for each tone (percentage correct).

To further examine the effects of Pitch Discrimination and Musical Ability on the accuracy of identification of each tone, a multivariate multiple linear regression was conducted. Results showed that Pitch Discrimination had a significant positive overall effect on the combined outcome [F(4, 33) = 3.31, p < 0.05, Pillai’s trace = 0.29], but that Musical Ability did not [F(4, 33) = 1.48, p > 0.05, Pillai’s trace = 0.15].

Separate linear regression models for each tone showed that Pitch Discrimination significantly predicted accuracy with Tone3 (β = 0.61, t(36) = 2.32, p < 0.05) and Tone4 (β = 0.84, t = 3.12, p < 0.05), while Musical Ability predicted performance for Tone1 (β = 0.66, t = 2.15, p < 0.05) and Tone4 (β = 0.58, t = 2.04, p < 0.05). The models for Tone4 (R² = 0.29, F(2, 36) = 8.80, p < 0.05) and Tone1 (R² = 0.13, F(2, 36) = 3.95, p < 0.05) were significant, indicating a good fit. The model for Tone3 was marginally significant (R² = 0.10, F(2, 36) = 3.08, p = 0.06), while the model for Tone2 was not significant (R² = 0.04, F(2, 36) = 1.87, p > 0.05).

4 Discussion

This study examined how musical expertise and associated cognitive skills (pitch discrimination, verbal and melodic memory, phonological processing) affect the ability of English speakers’ with no previous experience of tone languages to identify Mandarin tones. Individuals can develop musical expertise without formal training (Zendel and Alexander 2020): we were interested in whether music training per se (i.e., formal or self-guided training in playing an instrument or singing) or whether the cognitive abilities that are potentially been enhanced by music training confers an advantage in perceiving and potentially learning lexical tone. We recruited participants with a range of music experience; amateur musicians with formal training, those with informal musical engagement (e.g., instrument playing or choir singing) and those with no musical training at all. In contrast to previous studies (e.g., Alexander et al. 2005; Bidelman et al. 2013; Musacchia et al. 2007), we assessed music training holistically using the Gold-MSI questionnaire and treated music training as a continuous variable to avoid grouping participants based on duration of music training.

Results showed that lexical tone identification was significantly correlated with Music Training, Pitch Discrimination, and Musical Ability, but not with working memory or general segmental speech perception abilities. Path analysis revealed that Music Training indirectly influenced Tone Identification with Pitch Discrimination and Musical Ability as mediators. Although there were no significant differences in identification of individual tones, there was an asymmetrical confusion pattern; Tone1 was most commonly misidentified as Tone2, Tone2 as Tone3, Tone4 as Tone1, and Tone3 was confused equally with Tone2 and Tone4. Pitch Discrimination and Musical Ability predicted the accuracy of identification of different tones, with Pitch Discrimination a better predictor than Musical Ability.

It is generally well-accepted that musical training facilitates identification and discrimination of lexical tone for non-tone language speakers (e.g., Chua and Brunt 2014; Delogu et al. 2010). Our results present a more nuanced picture. Only Pitch Discrimination and Musical Ability (listening skills not limited to formal training) had a significant direct effect on Tone Identification. In contrast, Music Training did not have a direct effect on Tone Identification. Instead, Music Training significantly influenced Musical Ability and Pitch Discrimination. Music Training therefore had an indirect influence on Tone Identification via Pitch Discrimination and Musical Ability. This supports the idea that perception of pitch in music and in L2 tone are transferable (e.g., Marie et al. 2011), and further indicates that the transfer of pitch ability between music and language does not require music training, in particular the large amounts of music training typical of professional musicians.

As illustrated in the path analysis, both Pitch Discrimination and Musical Ability are important mediators between Music Training and Tone Identification, but they appear to have differing effects on tone identification. Multivariate multiple regression results demonstrated that performance on our Pitch Discrimination task explained more variance in the identification of each tone than did Musical Ability. While Musical Ability was not statistically significant as a predictor in the overall regression model, it had a significant influence on the mean score of the identification of all four tones. Results from individual linear regression models for each tone indicated that music training modulates tone processing in a selective manner, with stronger facilitatory effects on some specific tones compared to others. Specifically, Pitch Discrimination performance primarily predicted the identification of Tone3 and Tone4, whereas Musical Ability predicted the identification of Tone1 and Tone4. Neither Pitch Discrimination nor Musical Ability was a good predictor of Tone2 identification. This suggests that individuals with strong pitch discrimination and musical ability, regardless of the amount of music training they have had, may perform better in Mandarin tone identification, primarily due to enhanced accuracy in perceiving Tones 1, 3, and 4.

It is widely accepted that some tones are inherently more difficult to identify than others and so it was surprising that we found no significant differences in identification of the different tones. Among the 4 Mandarin tones, Tone1 and Tone4 are commonly considered to be easier to identify for listeners with no prior experience with a tonal language while Tone3 is thought to be the hardest to identify (e.g., Kirkham et al. 2011). There was a trend in this direction: all participants performed best with Tone1 but they performed worst with Tone4 not Tone3. In contrast to previous studies (e.g., So and Best 2010), we used natural stimuli. One possibility is that variation in the natural speech recordings masked difficulty with individual tones (cf. Wong et al. 2007), especially as our study used fewer trials per tone than previous studies that have shown this hierarchy (cf. So and Best 2010; Wang et al. 1999). A smaller number of trials may also have reduced our ability to detect subtle differences in performance. Still another explanation is that unlike our study which used identification, much of the research showing differences in performance between Mandarin tones has relied on discrimination tasks. For example, discrimination of Tone2-Tone3 has been shown to be harder than Tone1-Tone4 or other possible tone pairs (e.g., Hao 2012, 2018; Tsukada and Kondo 2019). Discrimination assesses perceptual sensitivity to differences between tones, while identification requires categorisation. Specifically, in our study, it means mapping individual tones to their pitch contours as represented by the tone marks used for the response options. This task requires listeners to track pitch changes and to match them to linguistic pitch categories, which may equalise difficulty across tones due to greater cognitive engagement. Thus, while Tone2-Tone3 may be perceptually confusable in a discrimination task, identification tasks may mitigate for these differences through categorical knowledge.

Although there was no significant difference in identification of individual tones, the confusion matrix shows that tone confusions were asymmetrical with non-reciprocal confusion patterns among the four tones. While Tone1 was mostly confused with Tone2, Tone2 was mostly confused with Tone3. Interestingly, Tone4 was mostly confused with Tone1, even though it is typically thought to be easy to distinguish from Tone1 (e.g., Hao 2018). This could be related to the fact that Tone4 resembles the intonation change at the end-of-statement sentences in non tonal language, such as English (So and Best 2008, 2014) and English speakers may thus consider Tone4 as neutral and with no special pitch changes, and such match it to Tone1.

Contrary to our expectations, we did not observe a relationship between verbal or melodic working memory on tone identification performance. This was surprising given that our participants had no previous experience with Mandarin and were only familiarised with each tone at the beginning of the experiment. In our task, participants heard a single word and identified which of the 4 tones they had heard. To do our task successfully, they would have needed to access stored representations held in memory from earlier in the experiment. On the other hand, they only heard a single word on each trial and responses were given by clicking on tone marks which depict tone direction. This may have acted as a prompt and thus reduced working memory load, leading to the absence of correlations with verbal or melodic working memory.

It was also surprising that while Music Training correlated with L1 phonological processing ability, it showed no such association with L2 phonological processing. This dissociation may stem from differences in the linguistic features used in each task. The L1 task assessed sensitivity to voice onset time (VOT), whereas the L2 task focused on place of articulation. Given that VOT and musical rhythm both involve temporal processing, it is possible that music training facilitated performance in our L1 perception task. In contrast, place of articulation is specific to the speech domain (Sadakata et al. 2010), which may explain the absence of a correlation with L2 performance. Critically, neither VOT nor place of articulation sensitivity directly relates to pitch perception, which is essential to Mandarin tone identification. Thus, despite the fact that we initially hypothesised that better segmental processing (in both L1 and L2) might free attentional resources for tone identification (Liu and Ning 2021), our results did not support this. One possible explanation is that our tone identification task used Mandarin syllables with segments shared between English and Mandarin, minimising segmental difficulty for naïve English listeners. Consequently, individual differences in segmental processing may not have significantly influenced tone identification performance. Future studies could test this by comparing tone identification performance using stimuli with shared versus Mandarin-unique segments.

In sum, the present study extends prior work that has demonstrated systematic differences in Mandarin tone identification between amateur musicians and non-musicians. It investigated the relationship between music training, L2 Mandarin tone identification and the cognitive abilities that music training can enhance. Our results highlight that musical training influences Mandarin tone perception specifically through enhanced pitch discrimination and musical ability, rather than through general linguistic-phonological processing or working memory capacity. Together the results imply that music and speech share perceptual resources but that the direct influence of pitch discrimination and musical ability on tone identification is tone-specific rather than uniform across all lexical tones. These findings demonstrate that music training can support, but is not necessary, for successful L2 lexical tone perception, and as such, may confer an initial advantage for a non-tone language speaking learner of Mandarin. Future studies should consider carefully the design and type of the experimental tasks used, for example, including both identification and discrimination tasks, employing more sensitive measures of pitch discrimination and using longer speech stimuli, investigating domain-transfer effects between music and individual tones, as well as extending these findings to natural sentence contexts.