Neural correlates of foreign speech imitation: The effects of age and music

© 2025 The Authors. Published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. Imaging Neuroscience, Volume 3, 2025 https://doi.org/10.1162/IMAG.a.75 Research Article 1. INTRODUCTION Speech production requires complex motor control, involving more than 100 muscles to work in concert. Speech motor control develops early in life ( Perrin & Venance, 2019; Wächter etal., 2009), making the acquisition of a second language speech easier and more native for children than adults, where foreign accents are common ( Flynn & Manuel, 1991; Geschwind & Carterette, 1966; Hack et al., 2012; Long, 1990; Wolfe, 1967). According to Kuhl et al. (2005), the end of the critical period around 7years old is characterized by a reduced cortical plasticity in the motor and auditory circuits, along with lower proficiencies in foreign speech phonetic discrimination and speech production. Although reduced Neural correlates of foreign speech imitation: The effects of age and music Xiaohui Yana, Jiaqi Maob, Zixin Maa, Kyle Perkinsc, Weizheng Lid, Yang Wangd, Fan Caoa aDepartment of Psychology, the University of Hong Kong, Hong Kong, China bBCBL Basque Center on Cognition, Brain and Language, Donostia, Gipuzkoa, Spain cretired professor, Florida International University, Miami, FL, United States dDepartment of Psychology, Sun YatSen University, Guangzhou, China Corresponding Author: Fan Cao ([email protected]) ABSTRACT Adult learners of a foreign speech are often marked by having a foreign accent; however, children and adults with singing training tend to have better pronunciations than adults without music training. The assimilation hypothesis proposes that people tend to assimilate foreign speech to native speech during perception and production, which may explain foreign accent. Unfortunately, the neural mechanisms underlying the age and music effects are still unclear. In this study, we compared brain activation patterns in three groups of participants, namely, children, adults with singing training, and adults without music training (control adults) during native (Chinese) and foreign speech (Spanish) imitation with each word repeated three times. We found greater representational similarity between Chinese and Spanish in both groups of adults than in children during both speech perception and production, supporting the assimilation hypothesis. Furthermore, we found groupspecific effects for the similarity between different times of imitation, suggesting different mechanisms. Specifically, control adults showed greater similarity between different times of Spanish word imitation than the other two groups in the medial orbital frontal cortex involved in adaptive learning/memory; children showed greater similarity than the other two groups in the bilateral inferior premotor/postcentral gyri involved in sensorimotor learning; adults with singing training showed greater similarity than the other two groups in the left superior temporal gyrus involved in auditory feedback. It suggests that singing training facilitates reliance on auditory discrimination, while children rely on somatosensory and speech motor control to learn foreign speech sounds, implicating different mechanisms of age and singing training effects. Our results provide insights in understanding the neural mechanisms of age and music effects in foreign speech learning. Keywords: representational similarity, foreign speech imitation, musicians, speech production Received: 24 January 2025 Revision: 1 May 2025 Accepted: 10 June 2025 Available Online: 25 June 2025 2 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 neural plasticity marks the end of the critical period, the specific underlying neural mechanisms of how learning foreign speech differs between children and adults remain unclear. The Directions into Velocities of Articulators (DIVA) model provides a framework for understanding the neural mechanisms underlying foreign speech imitation ( Tourville & Guenther, 2011), which is essentially a process of adjusting motor control according to somatosensory and auditory feedback. According to the DIVA model ( Tourville & Guenther, 2011), a feedforward control system is responsible for projecting motor commands for executing articulatory gestures. At the same time, a feedback control system generates somatosensory expectations of the articulatory gestures, and compares incoming somatosensory and auditory feedbacks with sensory expectations. If error signals are detected, corrective motor commands are sent to the motor cortex ( Guenther & Vladusich, 2012; Tourville & Guenther, 2011). Specifically, in the motor cortex, the middle precentral gyrus is a larynx area which directly controls muscles in the vocalfold, operation of which determines the opening and closing of the glottal space, as well as tensing and relaxing of vocal folds, modulating the pitch. The inferior precentral gyrus is involved in lip and tongue movement control. Lesion studies found that the middle and inferior precentral gyrus are the most consistent regions associated with foreign accent syndrome in stroke patients ( Higashiyama etal., 2021). While the middle and inferior precentral gyrus are involved in producing acquired speech sounds, the striatum, thalamus, and premotor cortex are more involved in vocal learning of novel speech sounds, according to Jarvis (2004, 2006) based on research of songbirds and humans. Simmonds (2015) further suggested that the vocal learning pathway (e.g., the striatum) becomes inactive too early during vocal learning, and the motor cortex for producing acquired speech sounds is involved instead, which may be why there is foreign accent in late second language learners. Unfortunately, no studies have compared the realtime brain mechanisms underlying foreign speech imitation in children and adults. Previous studies have compared native and nonnative vowels imitation in adults ( Carey, Miquel etal., 2017; Klein etal., 2006), as well as adults with higher L2 aptitude and those with lower L2 aptitude during L1 and L2 production ( Hu, Ackermann etal., 2013). A few studies also concerned how age of acquisition affects speech production in second language ( Berken etal., 2015; Berken, Gracco, etal., 2016; FrenckMestre etal., 2005). Two of them found that early bilinguals tend to show greater activation and greater grey matter volume in the putamen than late bilinguals ( Berken, Gracco, etal., 2016; FrenckMestre etal., 2005), while another study found greater activation in the inferior frontal gyrus in early bilinguals than late bilinguals during speech production of L2 ( Berken, Chai, etal., 2016). However, these studies exclusively scanned adults, with half being early bilinguals and half being late bilinguals, examining L2 as a longterm learning effect, rather than the realtime neural underpinnings of foreign speech acquisition. A few studies have examined the realtime brain activations of adults learning new languages. One study found greater activation in the anterior insula and inferior frontal gyrus when learning to speak a new language compared to native speech production, especially in the first 10min of speaking the new language ( Moser etal., 2009). Training studies examined brain activation during nonnative speech sound perception ( Golestani & Zatorre, 2004) and production ( Simmonds etal., 2014) before and after training. It was found that more efficient processing in the frontal speech areas was correlated with greater success in foreign phonetic identification ( Golestani & Zatorre, 2004). Simmonds etal. (2014) found reduced activation in the anterior striatum over time both within and between scanning sessions, suggesting that the striatum becomes inactive too early during vocal learning. However, no studies have directly compared the learning process of foreign speech imitation between children and adults to understand why children have a reduced foreign accent. One hypothesis for the existence of foreign accents in adults when learning a second language is that they use their first language as reference. Some previous studies have shown assimilation of foreign speech sounds to native sounds during perception in adult learners ( Best & Tyler, 2007; Best etal., 2001).This inaccurate speech representation further causes failure in speech motor control during production ( Ingram & Park, 1997). On the other hand, children confront less interference from native language. For example, Baker etal. (2008) found that children aged 7 to 14 years old performed better at distinguishing similar native and foreign speech sounds than adults. One possible reason for the less assimilation to the native speech in children than in adults is that the native speech representation system is still under development in children ( Zevin, 2012). In fact, the representation of native speech sound categories shows a significant development for the age range of 6– 12years and 12– 18years ( McMurray etal., 2018). However, no neuroimaging evidence is available yet to support this assimilation hypothesis. Moreover, it is unknown whether the greater similarity between foreign speech sounds and native speech sounds in adults than in children exists only in perception or extends to production as well. In the current study, we compared the brain 3 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 activation patterns of native Chinesespeaking children (aged 9– 10) and adults in Spanish speech imitation. The task mimicked a naturalistic speech learning situation in which each Spanish word was repeated three times, with the participant repeating it after each presentation. Under this paradigm, we aimed at comparing how children and adults are different during this foreign speech learning process. We expect greater similarity between Chinese and Spanish representation in adults than in children and faster decline of the striatum in adults than in children during Spanish speech learning. Furthermore, about 5– 15% of the population have the ability to achieve nativelike speech, even when the age of acquisition is late ( Abrahamsson & Hyltenstam, 2008; Birdsong, 2005; Flynn & Manuel, 1991; Wells, 1985). It has been redundantly documented that there is crossdomain transfer from musical expertise to speech perception and production ( Jekiel & Malarski, 2021; Milovanov & Tervaniemi, 2011; Weiss etal., 2015; Wong etal., 2007; Zuk etal., 2013), which could be explained by the common neural correlates involved in musical and speech processing in the auditory pathway and motor cortex ( Ozdemir etal., 2006). Moreover, the difference in musical nodes is more trivial than that in speech phonemes, leading to the fact that musical abilities can be transferred to speech abilities. One study showed that musical ability predicted L2 phonological ability (both receptive and productive) even after controlling for other factors, but did not account for the unique variance in L2 syntax or lexical knowledge ( Slevc & Miyake, 2006). Another study found that musical experience is correlated with greater pronunciation accuracy of English vowels after speech therapy for accent reduction ( Jekiel & Malarski, 2021). Unlike musical instrument training, which mainly fosters speech perception, vocal training is beneficial for both speech perception and speech production, as singing not only enhances individuals’ pitch discrimination, but also trains the vocal motor apparatus necessary for pitch production. As illustrated in a study by Christiner and Reiterer (2015), in which three groups of German adults (i.e., singers, instrumentalists, nonmusicians) were asked to imitate sentences of a foreign language (i.e., Hindi), both singers and instrumentalists had higher performances than nonmusicians in mimicking the Hindi utterances. In addition, as vocalists had more precise vocal control than instrumentalists, singers outperformed instrumentalists in the Hindi sentence imitation task. Therefore, singing may be a strong predictor of good pronunciation skills in foreign speech imitation. Neurologically, researchers have found that professional singers tend to have a greater volume in the ventral primary somatosensory cortex, rostral supramarginal gyrus, and auditory cortex than nonmusicians ( Kleber etal., 2017). In addition, larger volumes of arcuate fasciculus, which is a fiber tract connecting temporal areas and prefrontal regions, were found in musicians than nonmusicians ( Halwani etal., 2011). These brain structural changes in musicians were argued to play a significant role in speech production ( Halwani etal., 2011). However, no published studies have compared brain activation patterns during actual foreign speech learning in singers and nonmusicians to understand the brain differences for singers to outperform nonmusicians. In the current study, we recruited adults with professional vocal music training for more than 2years and we planned to compare the foreign speech learning mechanisms in the brain in singer adults, children and adults without music training (control adults) to understand why adults with vocal training and children have advantages than control adults in foreign speech learning. We expect singers to have more accurate representations than control adults in the feedforward motor control areas, including the key vocal learning regions in the striatum, or somatosensory and auditory feedback areas. 2. METHOD 2.1. Participants We recruited 32 adults without music background (i.e., control adults) (mean age 22.6, range 19– 31), 20 adults with vocal singing training for at least 2years (mean age 19, range 18– 24), and 20 children without music training (mean age 10.3, range 10– 11) in the local city. The adults with singing training were recruited from a local music college. The demographic information of the participants is presented in Table1. The adults with singing training had 4.08years of professional vocal training on average (range: 2– 9 years, std: 2 years), with age of vocal training onset being 7.95 years old on average (range: 6– 17 years old, std: 3.75 years old). The two adult groups were matched on education and English pronunciation according to a native English speaker’s rating on each participant’s language sample (t(50)=0.446, p=.657). All participants also met the following criteria: (a) native Chinese speakers with English as their L2; (b) never learned Spanish, French, Portuguese, or Italian for more than 3months; (c) free of medical implants and other metal accessories; (d) free of claustrophobia, hearing disorders, attention deficit hyperactivity disorder (ADHD), other developmental disabilities, neurological disease, and psychiatric disorders; and (e) righthanded. The present study was approved by the ethics committee at the local univer- 4 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 sity. Informed consent was obtained from all participants/parents of participants before data collection. 2.2. Sensitivity analysis Sensitivity analysis was conducted using Gpower 3.19. In order to achieve α=0.001 and a statistical power of 95%, our current sample size would need an effect size of 1.51 for the comparison between control adults and children, 1.71 for the comparison between children and adults with singing training, and 1.49 for the comparison between control adults and adults with singing training. In the wholebrain analysis, our actual effect size is 3.09 (https://www . sdmproject . com / utilities / ? show = Statistics) for all group comparisons when the voxellevel threshold was set at p<.001, which is much larger than needed. 2.3. Procedures 2.3.1. Behavioral tasks Several behavioral tests were administered before the fMRI scanning. A pseudoword rhyming judgment test and an initial sound deletion test were included to test phonological awareness. The pseudoword rhyming judgment test consists of 40 pairs of singlesyllable English pseudowords, and participants were asked to determine if the two pseudowords in a pair rhymed or not. In the initial sound deletion test, participants were asked to listen to real English words and repeat it out loud without the initial sound. There were 30 words, including 10 singlesyllable words, 10 twosyllable words, and 10 threesyllable words. Furthermore, we measured working memory using a digit span test in forward and reversed order. The digit span test was in Chinese, which is the first language of participants. In this test, experimenters explicitly read random digit strings with an increasing span. All participants were given the same tests. In addition, we measured music aptitude in adult participants using the Advanced Measures of Music Audition (AMMA; Gordon, 1989). In this test, participants listened to 30 pairs of music audios, and after each pair of audio, they were asked to choose one from the following options: the two audios differ in tones, differ in rhythms, same, or not sure. There is a tonal score, a rhythm score, and a composite score in AMMA. 2.3.2. fMRI task In the speech imitation task, there was a Spanish run and a Chinese run that were counterbalanced across participants. For the Spanish run, there were 28 Spanish real words (15 twosyllable words, 10 threesyllable words, and 3 foursyllable words), and for the Chinese run, there were 28 Chinese pseudowords with syllable numbers matched with the Spanish words. Chinese pseudowords were used in order to avoid semantic activation in the native language but not in the foreign language. In both runs, participants were asked to listen to and repeat each word/pseudoword three times consecutively. Audio stimuli played in the scanner were recorded by a native female speaker in Spanish and Chinese respectively. A 5min practice session was conducted prior to the scanning. As illustrated in FigureS1, a red cross and an audio word/pseudoword were presented for 1500ms. Then, the participant was asked to imitate the word/ pseudoword they just heard in the next 1500ms (t1). After a “jitter” screen which was presented for 1000– 5000 ms (3000 ms at average), the same word was played again for 1500ms and the participant was asked to imitate the word for the second time (t2). The same procedure was repeated again for the third time (t3). After the third “jitter” screen, a red cross was displayed for 3000ms without auditory stimulus, which served as the baseline, followed by the fourth “jitter”. Then, the next trial started. After the fMRI scanning, participants were asked to imitate the Spanish speech again outside Table1. Demographic information and behavioral tests results in each group of participants. Control adults Adults with singing training Children N32 (10M) 20 (3M) 20 (6M) Age (years) 22.6 (3.3) [1931] 19.9 (1.5) [1824] 10.3 (0.6) [1011] Rhyming judgment 36.8 (2.5) 35.3 (3.8) 31.0 (4.9) Initial sound Deletion 26.5 (4.3) 23.9 (7.3) 15.5 (8.5) Digit span (forward) 9.0 (1.4) 9.6 (1.3) 8.1 (1.3) Digit span (backward) 6.4 (2.2) 6.2 (2.4) 4.7 (1.3) AMMA (percentile) 61.4 (27.4) 80.5 (15.8) - Numbers in the parenthesis are standard deviations, and in the brackets are ranges. M: males; AMMA: Advanced Measures of Music Audition. 5 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 the scanner in order to have a highquality recording of their pronunciation. 2.4. fMRI data acquisition The fMRI images were acquired using a 3T Siemens Prisma MRI scanner. Participants lay down in the scanner with a standard 20channel head coil, and two foam pads were used to help reduce head movement. Before they entered the scanner, a mock scanner was used for practicing speaking with limited head movement. During scanning, a realtime monitoring of head movement was conducted, and participants were reminded to keep their head still while talking during the break between runs if the head movement was large. A singleshot echo planar imaging (EPI) sequence was adopted to collect functional BOLD signals, with an interleaved acquisition from bottom to top for each volume (repetition time (TR)=2000ms, echo time (TE) = 20.0 ms, flip angle = 80°, matrix size= 128 ×128, field of view (FOV)=220mm, slice thickness=3.0mm, number of slices=34, voxel size =1.7×1.7×3.0mm3). There were 348 volumes collected for each run. Highresolution structural T1weighted 3D images (MPRAGE) were also acquired (TR=2300ms, TE = 3.24 ms, TI = 900 ms, flip angle = 9°, matrix size=256×256, FOV=260mm, slice thickness=1.0mm, number of slices=160). 2.5. Data analysis 2.5.1. Acoustic analysis 2.5.1.1. Voice onset time (VOT) analysis. As a behavioral indicator of the inscanner task performance, we measured the voice onset time (VOT) of /b/ and /d/ in seven Spanish words using Praat ( Boersma, 2001). The VOT is the time interval between a plosive consonant release and voicing onset. The seven Spanish words (“bebé”, “bueno”, “brazo”, “dado”, and “difícil”) we chose contained voiced stops (/b/ and /d/ in Spanish) in which their onsets of phonation occur before the consonant release, resulting in a negative VOT. FigureS2 illustrates the measurement of VOT in Praat. 2.5.1.2. Formant frequency analysis. The formant frequency analysis included the same seven Spanish words as the VOT analysis, covering all vowels in Spanish (/a/, /e/, /i/, /o/, and /u/). In order to quantify the performance on Spanish vowels’ imitation, the first formant (F1) and second formant (F2) of each vowel were calculated in Praat. The F1 and F2 are the first two resonating frequencies in a vowel’s pronunciation, with F1 indicating the opening of lips and F2 indicating the tongue’s position (Wood, 1982). The individual formant values were normalized using the R package NORM (Thomas & Kendall, 2007), in order to eliminate the influence of gender and age (Fabricius etal., 2009). Two independent researchers who were blinded about the participants’ information calculated the VOT and frequency formants, and their evaluation results were correlated (r=.856, p<.001 for the VOT; r=.982, p<.001 for the frequency formant). We averaged their results to serve as the final score. The VOT and formant values of each consonant and vowel were averaged across all words and entered into a repeatedmeasure ANOVA of group (control adults, adults with singing training, children) by time (1st, 2nd, 3rd) for further analysis. 2.5.2. fMRI images preprocessing fMRI data preprocessing was conducted using DPARSF 4.3 (Yan etal., 2016; http://rfmri . org / DPARSF). First of all, slice timing was performed to correct timing difference of the interleaved slices with the middle slice as the reference. Next, functional images were aligned to the first volume to correct head movement. We used ART (Artifact Detection Tools, https://www . nitrc . org / projects / artifact _ detect) to detect head movements that exceeded 3mm for translations or 3° for rotations for each participant. We found that 4 participants (2 from the adults without music training group and 2 from the children group) had excessive head movements for less than 8 volumes across all sessions. Considering that the affected data points were less than 2% of the total data in each participant, we kept these participants and repaired the affected volumes in ArtRepair (https://www.nitrc.org/projects/art _repair/), using the interpolated values from neighboring time points to replace the affected time points. We also deleted two participants due to extensive head movement (1 from the adults without music training group and 1 from the adults with singing training group). The current sample size was after eliminating these two participants. Then, T1weighted structural images were coregistered to the realigned functional images for each individual. A filter was applied so that only signals above 0.01Hz were kept. The images were segmented into gray matter, white matter, and cerebrospinal fluid, before being normalized to the Montreal Neurological Institute (MNI) space. Finally, the normalized images were detrended by regressing out the nuisances using a Friston 24parameter model ( Friston etal., 1996). 2.5.3. Representational similarity analysis (RSA) RSA was conducted to calculate the similarity between Chinese and Spanish, as well as between different times of imitation within a language (i.e., t1 and t2, t2 and t3, t1 6 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 and t3) using the CoSMoMVPA toolbox (http://www . cosmomvpa . org/). Specifically, brain responses for each trial were estimated using unsmoothed preprocessed data. The LeastSquares Separate (LSS) method ( Mumford etal., 2012) was used, with six head movement parameters and all other trials as covariates. A searchlight approach was adopted, where a sphere containing 125 voxels was centered at each voxel and moved across the entire brain. For the crosslanguage similarity analysis, pattern similarity between Chinese and Spanish was calculated using splithalf correlation. Specifically, within each searchlight, a Pearson’s correlation coefficient was calculated between each Chinese trial and each Spanish trial on the beta values of the 125 voxels, and in total there were 84×84=7056 such correlation coefficients, because we had 84 trials (28×3=84) in each language. Then, we averaged these 7056 correlation coefficients to represent similarity between Chinese and Spanish at this voxel. For the similarity between different times of imitation within a language, at each searchlight, a Pearson’s correlation coefficient was calculated between each trial and each other trial in the same imitation order within a language (28 trials in total) on the beta values of the 125 voxels; therefore, we had a 28×28 DSM. Spearman’s correlations were calculated between the DSMs of the first imitation and the second imitation, between the second and the third imitation, and between the first and the third imitation. Following the analysis, the results were ztransformed and subjected to further group analysis. The threshold was set at uncorrected p< .001 at the voxel level and FWEcorrected p<.05 at the cluster level when reporting group analysis. 2.5.4. Machine learning In order to confirm that the three groups have different representation patterns of foreign speech, we adopted a machinelearning approach. After calculating the representational similarity between different times of imitation, three machinelearning models were set up to classify children versus control adults, children versus adults with singing training, and control adults versus adults with singing training separately using similarity between t1 and t2, t2 and t3, t1 and t3 in each language for both perception and production. In total, there were 12 similarity parameters (i.e., three similarities in two languages for both perception and production). The analysis was performed using PRoNTo v2.1 (Pattern recognition for neuroimaging toolbox) (http://www . mlnl . cs . ucl . ac . uk / pronto / prtsoftware . html). Specifically, using training data, the 12 similarity parameters were first averaged and then meancentered. Subsequently, a binary linear Support Vector Machine (SVM) was employed to train the model in a wholebrain gray matter mask, with crossvalidation performed using the leaveonesubjectout approach. To measure whether the classification is successful, 1000 permutations were conducted. For a successful classification, the contributing weight map was computed. The weight map was then thresholded at 30% of the maximum weight value with 100 extended voxels. 2.5.5. Statistical analysis on brain activation A general linear model (GLM) was constructed in SPM12 (Statistical Parametric Mapping, http://www . fil . ion . ucl . ac . uk / spm) after the data were smoothed with an isotropic Gaussian kernel of 4 mm full width half maximum (FWHM). For each participant, the preprocessed functional images from all sessions were entered into a GLM to estimate the wholebrain neural activities for the perception and production stage. In the grouplevel statistical analysis, we conducted flexible factorial ANOVAs of group (control adults, adults with singing training, children) by imitation orders (t1, t2, t3) separately for perception and production for each language in SPM ( Gläscher & Gitelman, 2008). The main effects of order and group as well as the interaction between order and group were calculated. The threshold was set at uncorrected p<.001 at the voxel level and FWEcorrected p<.05 at the cluster level. 3. RESULTS 3.1. Behavioral tests There was a significant main effect of group for both phonological awareness tests (for the rhyming judgment test: F(2, 69)=15.86, p<.001, partial η ²=.261; for the initial sound deletion test: F(2, 69)=18.239, p<.001, partial η ²=.261). The Bonferronicorrected posthoc analysis showed that children had significantly lower scores on the pseudoword rhyming judgment test than control adults (t(69)=5.603, p<.001, partial η ²=.313) and adults with singing training (t(69) = 3.575, p = .017, partial η ²=.156). Children also had lower scores on the initial sound deletion test than control adults (t(69) = 5.97, p<.001, partial η ²=.341) and adults with singing training (t(69)=4.14, p<.001, partial η ²=.199). No differences were found between the two adult groups on the pseudoword rhyming judgment test (t(69) = 1.637, p=.238, partial η ²=.037), or the initial sound deletion test (t(69)=1.371, p=.389, partial η ²=.026). For the working memory test of digit span, a significant group effect was found in the test of forward order (F(2, 69)=5.908, p=.004, partial η ²=.105), but not in the test 7 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 of reversed order (F(2, 69) = 1.761, p = .179, partial η ²=.078). A Bonferronicorrected posthoc test revealed greater digit span in adults with singing training than in children (t(69)=3.345, p=.025, partial η ²=.14) and in control adults than in children (t(69)=2.514, p=.043, partial η ²=.084). No significant difference between the two adult groups (t(69)=1.761, p=.179, partial η ²=.043) was found. For the AMMA test, the adults with singing training significantly outperformed the control adults (t(50)=8.568, p=.005, partial η ²=.595). 3.2. VOTs We ran an ANOVA of group by order for each consonant separately (i.e., /d/ and /b/). We found a significant main effect of group for /d/ (F(2, 65)=5.299, p=.007, partial η ²= .14) and /b/ (F(2, 65)=3.445, p=.038, partial η ²= .096). Bonferronicorrected posthoc analysis revealed that children’s VOT was more negative than control adults (t(45)=3.934, p=.006, partial η ²=.256) for /d/, and marginally more negative than control adults (t(45)=2.847, p=.06, partial η ²=.153) for /b/ (Fig.1). Adults with singing training did not differ significantly from the other two groups in the VOTs of either /d/ (for comparison with children: t(33)=- 1.857, p=.574, partial η ²=.095; for comparison with control adults: t(48)=1.724, p=.248, partial η ²= .058) or /b/ (for comparison with children: t(33)= - .907, p= .645, partial η ²= .024; for comparison with control adults: t(48)=1.955, p=.184, partial η ²=.074). The other main effects or interaction effects were not significant. 3.3. Formant of the vowels To evaluate the imitation performance, we calculated the distance between the participant’s vowel in the frequency space (F1, F2) and the native speaker’s vowel (F1stimulus, F2stimulus). Then, a repeatedmeasure ANOVA of group by order was conducted for each vowel on the distance. For the vowel /o/ and /i/, the ANOVA revealed a significant main effect of group (F(2, 63)=11.559, p<.001, partial η ²=.187 for /o/, and F(2, 63)=5.965, p=.004, partial η ²=.101 for /i/). Bonferronicorrected posthoc analysis revealed that for /o/, control adults had a larger distance than adults with singing training (t(48)=4.679, p<.001, partial η ²=.314) and children (t(43)=2.793, p=.013, partial η ²=.153). For /i/, children and control adults had Fig.1. Control adults showed a poorer performance on VOT and vowel’s frequency formant than singingtrained adults and children. A and B are the VOT results for /d/ and /b/ in each group; C and D are the distance to the native speaker in the frequency formant space for /o/ and /i/. The dotted line in A and B is the VOT for /d/ and /b/ in a native Spanish speaker. 8 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 Table2. Results of the crosslinguistic similarity. Anatomical label H Cluster size (Voxels) MNI coordinate Zx y z (Control adults>children) & (adults with singing training>children) (Perception) Superior occipital gyrus/cuneus, BA 7/17/18/19 B 703 18 - 84 18 3.71 Caudate/ Thalamus B 145 4 14 0 3.48 (Control adults>children) & (adults with singing training>children) (Production) Precuneus/ posterior cingulate gyrus, BA 18/30 L 125 - 20 - 52 2 3.42 Calcarine sulcus/cuneus/posterior cingulate gyrus, BA7/18/19/23 B 1265 20 - 50 6 4.07 Rolandic/ superior temporal gyrus/middle temporal gyrus, BA 13/22/40/41 R 299 54 - 28 20 3.44 Perception>production (main effect) Thalamus, caudate L 131 - 14 - 18 14 4.47 Brain stem R 138 0 - 22 - 8 4.42 Thalamus, caudate head R 57 0 6 2 4.19 Caudate body L 59 - 8 14 10 3.99 Interaction of group by process (perception, production) Cingulate gyrus/supplementary motor area, BA 31 B 138 0 - 28 44 4.92 a larger distance than adults with singing training (for children: t(31)=1.619, p=.007, partial η ²=.078; for control adults: t(48) = 1.694, p = .019, partial η ² = .056) (Fig. 1). However, no significant group difference was detected for the other vowels (/a/: F(2, 63) = 1.419, p=.250, partial η ²=.064; /e/: F(2, 63)=.527, p=.593, partial η ²=.007; /u/: F(2, 63)=.902, p=.411, partial η ²=.09). No main effects of order or interactions between group and order were significant for any vowel. 3.4. Representational similarity analysis 3.4.1. Similarity between Chinese and Spanish We conducted an ANOVA of group by process (perception, production) to examine group differences in the similarity between the two languages during perception and production. We found main effects of group and processes, as well as interactions between group and process. For the group differences, we found that both adult groups showed greater similarity than children, but children did not show greater similarity than adults; therefore, we conducted a conjunction analysis between control adults>children and adults with singing training>children. The conjunction analysis showed that both control adults and adults with singing training showed greater representational similarity between Chinese and Spanish than children in the bilateral caudate/thalamus and superior occipital gyrus, cuneus during perception, and in the bilateral precuneus, cuneus, posterior cingulate and right STG during production (Table2, Fig.2A, B). For the main effect of process, we found greater similarity in the bilateral thalamus and caudate in perception than in production (Table2). There was an interaction between group and process in the cingulate gyrus/SMA, driven by greater crosslinguistic similarity in control adults than in children in perception, and greater similarity in adults with singing training than in children in production (Fig.2C). 3.4.2. Similarity between the 1st, 2nd, and 3rd imitation We also calculated representational similarity between different times of imitation within a language separately for perception and production. Then, we conducted a group (3) by language (2) ANOVA separately for perception and production. We found main effects of group. Then, we conducted a conjunction analysis to identify groupspecific similarity patterns. Specifically, a conjunction between control adults>children and control adults>adults with singing training would reveal regions specific for control adults. A conjunction between children>control adults and children>adults with singing training would reveal regions specific for children. A conjunction between adults with singing training>children and adults with singing training>control adults would reveal regions specific for adults with singing training. We found greater similarity in control adults than in children and adults with singing training in the bilateral medial orbital frontal cortex across both perception and production; and greater similarity in children than in control adults and adults with singing training in the bilateral inferior premotor/postcentral gyrus in both perception and production (Table3, Fig.3). In the conjunction analysis, we did not find greater similarity in adults with singing training than the other two groups. 9 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 For the main effect of language, we found greater similarity in Spanish than in Chinese in bilateral thalamus during perception (Table3). We found interaction effects between group and language for both perception and production (Fig.4). For perception, we found interactions at the left STG and supramarginal gyrus. At the left STG, adults with singing training had greater similarity than the other two groups in Spanish, while control adults had greater similarity than the other two groups in Chinese. At the left supramarginal gyrus, adults with singing training had greater similarity than control adults in Spanish while control adults had greater similarity than adults with singing training in Chinese. Children did not show difference from the other two groups in either language at the left supramarginal gyrus. For production, control adults showed greater similarity than the other two groups in the left insula in Chinese, but no group differences were found in Spanish. In the left STG, control adults showed greater similarity than adults with singing training in Chinese, but no group differences were found in Spanish. 3.4.3. Machine learning results Machine learning yielded significant results when classifying control adults and children (ACC = 74.38%, p=.001) (Fig.5). The bilateral medial frontal gyrus, bilateral superior temporal gyrus, bilateral caudate/thalamus, and bilateral postcentral gyrus/precentral gyrus contributed to the contrast of control adults minus children; in contrast, the bilateral superior frontal gyrus, bilateral Fig.2. Results of the similarity between Chinese and Spanish. Adults showed greater similarity between the two languages than children in both the perception and production processes, while children did not show greater similarity than adults (A, B). (C) is the interaction between group and process for the similarity between Chinese and Spanish. Adults with singing training showed greater similarity in the cingulate gyrus in production than perception, while control adults showed greater similarity in perception than production. 16 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 adaptive learning; children showed greater similarity in bilateral inferior premotor/postcentral gyrus, suggesting sensorimotor learning; and adults with singing training showed greater similarity in the left STG, suggesting reliance on auditory feedback. Taken together, these findings pave the way for understanding why adults confront greater challenge in foreign speech learning than children, and how singing training may help. DATA AND CODE AVAILABILITY All data and code will be available after acceptance of the paper based on request sent to the correspondence author with a possibility of needs for a formal datasharing agreement, and approval from the requesting researcher’s local ethics committee. AUTHOR CONTRIBUTIONS X.Y.: investigation, methodology, data curation, data validation, visualization, and formal data analysis. J.M.: investigation, methodology, data curation, data validation, formal data analysis, and draft writing. Z.M.: writing, editing, and reviewing. K.P.: data validation, writing, and editing and reviewing. W.L.: investigation, methodology, and data curation. Y.W.: investigation, methodology, and data curation. F.C.: formation of ideas, supervision of investigation, methodology, formal data analysis, draft writing, editing, and reviewing. FUNDING This work was funded by “General Research Fund (17605925), the Research Grants Council, Hong Kong” awarded to F.C. DECLARATION OF COMPETING INTEREST The authors claim no competing interests. SUPPLEMENTARY MATERIALS Supplementary material for this article is available with the online version here: https://doi . org / 10 . 1162 / IMAG . a . 75 REFERENCES Abrahamsson, N., & Hyltenstam, K. (2008). The robustness of aptitude effects in nearnative second language acquisition. Stud Second Lang Acquis, 30(4), 481–509. https://doi . org / 10 . 1017 / s0272263108080339 Baker, W., Trofimovich, P., Flege, J.E., Mack, M., & Halter, R. (2008). Child— adult differences in secondlanguage phonological learning: The role of crosslanguage similarity. Lang Speech, 51(4), 317–342. https://doi . org / 10 . 1177 / 0023830908099068 Bamiou, D.E., Musiek, F.E., & Luxon, L.M. (2003). The insula (Island of Reil) and its role in auditory processing. Literature review. Brain Res Brain Res Rev, 42(2), 143– 154. https://doi . org / 10 . 1016 / s0165 - 0173(03)00172 - 3 Berken, J.A., Chai, X., Chen, J.K., Gracco, V.L., & Klein, D. (2016). Effects of early and late bilingualism on restingstate functional connectivity. J Neurosci, 36(4), 1165–1172. https://doi . org / 10 . 1016 / j . neuropsychologia . 2016 . 08 . 031 Berken, J.A., Gracco, V.L., Chen, J.K., & Klein, D. (2016). The timing of language learning shapes brain structure associated with articulation. Brain Struct Funct, 221(7), 3591–3600. https://doi . org / 10 . 1007 / s00429 - 015 - 1121 - 9 Berken, J.A., Gracco, V.L., Chen, J.K., Watkins, K.E., Baum, S., Callahan, M., & Klein, D. (2015). Neural activation in speech production and reading aloud in native and nonnative languages. Neuroimage, 112, 208–217. https://doi . org / 10 . 1016 / j . neuroimage . 2015 . 03 . 016 Best, C.T., McRoberts, G.W., & Goodell, E. (2001). Discrimination of nonnative consonant contrasts varying in perceptual assimilation to the listener’s native phonological system. J Acoust Soc Am, 109(2), 775–794. https://doi . org / 10 . 1121 / 1 . 1332378 Best, C.T., & Tyler, M.D. (2007). Nonnative and secondlanguage speech perception: Commonalities and complementarities. In O.- S. Bohn & M.J. Munro (Eds.), Language learning & language teaching (Vol. 17, pp. 13–34). John Benjamins Publishing Company. https://doi . org / 10 . 1075 / lllt . 17 . 07bes Birdsong, D. (2005). Interpreting age effects in second language acquisition. In J.F. Kroll & A.M.B. de Groot (Eds.), Handbook of bilingualism: Psycholinguistic approaches, 109, 127. Oxford University Press. https:// doi . org / 10 . 1093 / oso / 9780195151770 . 003 . 0007 Boersma, P. (2001). Praat, a system for doing phonetics by computer. Glot Int, 5(9), 341–345. https://doi . org / 10 . 1097 / aud . 0b013e31821473f7 Brown, S., Ngan, E., & Liotti, M. (2008). A larynx area in the human motor cortex. Cereb Cortex, 18(4), 837–845. https://doi . org / 10 . 3410 / f . 1104671 . 560738 Carey, D., Miquel, M.E., Evans, B.G., Adank, P., & McGettigan, C. (2017). Functional brain outcomes of L2 speech learning emerge during sensorimotor transformation. Neuroimage, 159, 18–31. https://doi . org / 10 . 1016 / j . neuroimage . 2017 . 06 . 053 Christiner, M., & Reiterer, S.M. (2015). A Mozart is not a Pavarotti: Singers outperform instrumentalists on foreign accent imitation. Front Hum Neurosci, 9, 482. https://doi . org / 10 . 3389 / fnhum . 2015 . 00482 Du, Y., & Zatorre, R.J. (2017). Musical training sharpens and bonds ears and tongue to hear speech better. Proc Natl Acad Sci, 114(51), 13579–13584. https://doi . org / 10 . 1073 / pnas . 1712223114 Fabricius, A., Watt, D., & Johnson, D. E. (2009). A comparison of three speaker-intrinsic vowel formant frequency normalization algorithms for sociophonetics. Language Variation and Change, 21(3), 413–435. https:// doi.org/10.1017/S0954394509990160 FrenckMestre, C., Anton, J.L., Roth, M., Vaid, J., & Viallet, F. (2005). Articulation in early and late bilinguals’ two languages: Evidence from functional magnetic resonance imaging. Neuroreport, 16(7), 761–765. https://doi . org / 10 . 1097 / 00001756 - 200505120 - 00021 Flinker, A., Korzeniewska, A., Shestyuk, A. Y., Franaszczuk, P. J., Dronkers, N. F., Knight, R. T., & Crone, N. E. (2015). 17 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 Redefining the role of Broca’s area in speech. Proc Natl Acad Sci U S A, 112(9), 2871–2875. https://doi .org/10.1073/pnas.1414491112 Flynn, S., & Manuel, S. (1991). Agedependent effects in language acquisition: An evaluation of critical period hypotheses. In L. Eubank (Ed.), Point counterpoint: Universal grammar in the second language, pp. 117–145. John Benjamins Publishing Company. https://doi . org / 10 . 1075 / lald . 3 . 06fly Friston, K.J., Williams, S., Howard, R., Frackowiak, R.S., & Turner, R. (1996). Movementrelated effects in fMRI timeseries. Magn Reson Med, 35(3), 346–355. https://doi . org / 10 . 1002 / mrm . 1910350312 Garrett, D.D., Kovacevic, N., McIntosh, A.R., & Grady, C.L. (2011). The importance of being variable. J Neurosci, 31, 4496–4503. https://doi . org / 10 . 1523 / jneurosci . 5641 - 10 . 2011 Garrett, D.D., SamanezLarkin, G.R., MacDonald, S.W.S., Lindenberger, U., McIntosh, A.R., & Grady, C.L. (2013). Momenttomoment brain signal variability: A next frontier in human brain mapping? Neurosci Biobehav Rev, 37, 610– 624. https://doi . org / 10 . 1016 / j . neubiorev . 2013 . 02 . 015 Gaser, C., & Schlaug, G. (2003). Brain structures differ between musicians and nonmusicians. J Neurosci, 23(27), 9240–9245. https://doi . org / 10 . 1523 / jneurosci . 23 - 27 - 09240 . 2003 Geschwind, N., & Carterette. (1966). A discussion of “Speech development: Its anatomical and physiological concomitants.” In E.H. Lenneberg & Carterette (Eds.), Brain function, 3. https://doi . org / 10 . 1525 / 9780520333819 - 006 Gläscher, J., & Gitelman, D. (2008). Contrast weights in flexible factorial design with multiple groups of subjects. Sml Editor, 1, 12. https://doi . org / 10 . 1097 / 01434893 - 200808000 - 00010 Golestani, N., & Zatorre, R.J. (2004). Learning new sounds of speech: Reallocation of neural substrates. Neuroimage, 21(2), 494–506. https://doi . org / 10 . 1016 / j . neuroimage . 2003 . 09 . 071 Gordon, E.E. (1989). Manual for the advanced measures of music audiation. G. I. A. Publications, Inc. https://doi . org / 10 . 2307 / 3399589 Gourley, S.L., Zimmermann, K.S., Allen, A.G., & Taylor, J.R. (2016). The medial orbitofrontal cortex regulates sensitivity to outcome value. J Neurosci, 36(16), 4600– 4613. https://doi . org / 10 . 1523 / jneurosci . 4253 - 15 . 2016 Guenther, F.H., & Vladusich, T. (2012). A neural theory of speech acquisition and production. J Neurolinguistics, 25(5), 408–422. https://doi . org / 10 . 1016 / j . jneuroling . 2009 . 08 . 006 Hack, J., MarinovaTodd, S.H., & May Bernhardt, B. (2012). Speech assessment of Chinese– English bilingual children: Accent versus developmental level. Int J Speech Lang Pathol, 14(6), 509–519. https://doi . org / 10 . 3109 / 17549507 . 2012 . 718361 Haefner, R.M., Berkes, P., & Fiser, J. (2016). Perceptual decisionmaking as probabilistic inference by neural sampling. Neuron, 90, 649–660. https://doi . org / 10 . 1016 / j . neuron . 2016 . 03 . 020 Halwani, G.F., Loui, P., Ruber, T., & Schlaug, G. (2011). Effects of practice and experience on the arcuate fasciculus: Comparing singers, instrumentalists, and nonmusicians. Front Psychol, 2, 156. https://doi . org / 10 . 3389 / fpsyg . 2011 . 00156 Higashiyama, Y., Hamada, T., Saito, A., Morihara, K., Okamoto, M., Kimura, K., Joki, H., Kishida, H., Doi, H., Ueda, N., Takeuchi, H., & Tanaka, F. (2021). Neural mechanisms of foreign accent syndrome: Lesion and network analysis. Neuroimage Clin, 31, 102760. https:// doi . org / 10 . 1016 / j . nicl . 2021 . 102760 Hu, X., Ackermann, H., Martin, J.A., Erb, M., Winkler, S., & Reiterer, S.M. (2013). Language aptitude for pronunciation in advanced second language (L2) learners: Behavioural predictors and neural substrates. Brain Lang, 127(3), 366–376. https://doi . org / 10 . 1016 / j . bandl . 2012 . 11 . 006 Ingram, J.C., & Park, S.- G. (1997). Crosslanguage vowel perception and production by Japanese and Korean learners of English. J Phonetics, 25(3), 343–370. https:// doi . org / 10 . 1006 / jpho . 1997 . 0048 Jarvis, E.D. (2004). Learned birdsong and the neurobiology of human language. Ann N Y Acad Sci, 1016, 749–777. https://doi . org / 10 . 1196 / annals . 1298 . 026 Jarvis, E.D. (2006). Selection for and against vocal learning in birds and mammals. Ornithol Sci, 5, 5–14. https://doi . org / 10 . 2326 / osj . 5 . 5 Jekiel, M., & Malarski, K. (2021). Musical hearing and musical experience in second language English vowel acquisition. J Speech Lang Hear Res, 64(5), 1666–1682. https://doi . org / 10 . 1044 / 2021 _ jslhr - 19 - 00253 Kleber, B., Friberg, A., Zeitouni, A., & Zatorre, R. (2017). Experiencedependent modulation of right anterior insula and sensorimotor regions as a function of noisemasked auditory feedback in singers and nonsingers. Neuroimage, 147, 97–110. https://doi . org / 10 . 1016 / j . neuroimage . 2016 . 11 . 059 Klein, D., Watkins, K.E., Zatorre, R.J., & Milner, B. (2006). Word and nonword repetition in bilingual subjects: A PET study. Hum Brain Mapp, 27(2), 153–161. https://doi . org / 10 . 1002 / hbm . 20174 Kuhl, P.K., Conboy, B.T., Padden, D., Nelson, T., & Pruitt, J. (2005). Early speech perception and later language development: Implications for the “critical period.” Lang Learn Dev, 1(3– 4), 237–264. https://doi . org / 10 . 1207 / s15473341lld0103 & 4 _ 2 Long, M.H. (1990). Maturational constraints on language development. Stud Second Lang Acquis, 12(3), 251–285. https://doi . org / 10 . 1017 / s0272263100009165 McMurray, B., Danelz, A., Rigler, H., & Seedorff, M. (2018). Speech categorization develops slowly through adolescence. Dev Psychol, 5(8), 1472–1491. https://doi . org / 10 . 1037 / dev0000542 Milovanov, R., & Tervaniemi, M. (2011). The interplay between musical and linguistic aptitudes: A review. Front Psychol, 2, 321. https://doi . org / 10 . 3389 / fpsyg . 2011 . 00321 Moser, D., Fridriksson, J., Bonilha, L., Healy, E.W., Baylis, G., Baker, J.M., & Rorden, C. (2009). Neural recruitment for the production of native and novel speech sounds. Neuroimage, 46(2), 549–557. https://doi . org / 10 . 1016 / j . neuroimage . 2009 . 01 . 015 Mumford, J.A., Turner, B.O., Ashby, F.G., & Poldrack, R.A. (2012). Deconvolving BOLD activation in eventrelated designs for multivoxel pattern classification analyses. NeuroImage, 59(3), 2636–2643. https://doi . org / 10 . 1016 / j . neuroimage . 2011 . 08 . 076 Nomi, J.S., Bolt, T.S., Ezie, C.E.C., Uddin, L.Q., & Heller, A.S. (2017). Momenttomoment BOLD signal variability reflects regional changes in neural flexibility across the lifespan. J Neurosci, 37, 5539–5548. https://doi . org / 10 . 1523 / jneurosci . 3408 - 16 . 2017 Ozdemir, E., Norton, A., & Schlaug, G. (2006). Shared and distinct neural correlates of singing and speaking. NeuroImage, 33(2), 628–635. https://doi . org / 10 . 1016 / j . neuroimage . 2006 . 07 . 013 18 X. Yan, J. Mao, Z. Ma etal. Imaging Neuroscience, Volume 3, 2025 Perrin, E., & Venance, L. (2019). Bridging the gap between striatal plasticity and learning. Curr Opin Neurobiol, 54, 104–112. https://doi . org / 10 . 1016 / j . conb . 2018 . 09 . 007 Raichle, M.E. (2015). The brain’s default mode network. Ann Rev Neurosci, 38, 433–447. https://doi . org / 10 . 1146 / annurev - neuro - 071013 - 014030 Raja Beharelle, A., Kovačević, N., McIntosh, A.R., & Levine, B. (2012). Brain signal variability relates to stability of behavior after recovery from diffuse brain injury. Neuroimage, 60, 1528–1537. https://doi . org / 10 . 1016 / j . neuroimage . 2012 . 01 . 037 Schoenbaum, G., Roesch, M., Stalnaker, T., & Takahashi, Y.K. (2009). A new perspective on the role of the orbitofrontal cortex in adaptive behaviour. Nat Rev Neurosci, 10, 885–892. https://doi . org / 10 . 1038 / nrn2753 Schwartz, B.D., & Sprouse, R.A. (1996). L2 cognitive states and the full transfer/full access model. Second Lang Res, 12(1), 40–72. https://doi . org / 10 . 1177 / 026765839601200103 Shuster, L. I., & Lemieux, S. K. (2005). An fMRI investigation of covertly and overtly produced monoand multisyllabic words. Brain Lang, 93(1), 20–31. https://doi.org/10.1016 /j.bandl.2004.07.007 Simmonds, A.J. (2015). A hypothesis on improving foreign accents by optimizing variability in vocal learning brain circuits. Front Hum Neurosci, 9, 606. https://doi . org / 10 . 3389 / fnhum . 2015 . 00606 Simmonds, A.J., Leech, R., Iverson, P., & Wise, R.J. (2014). The response of the anterior striatum during adult human vocal learning. J Neurophysiol, 112(4), 792–801. https://doi . org / 10 . 1152 / jn . 00901 . 2013 Slevc, L.R., & Miyake, A. (2006). Individual differences in secondlanguage proficiency: Does musical ability matter? Psychol Sci, 17(8), 675–681. https://doi . org / 10 . 1111 / j . 1467 - 9280 . 2006 . 01765 . x Thomas, E. R., & Kendall, T. (2007). NORM: The vowel normalization and plotting suite. http://lingtools.uoregon .edu/norm/ Torrico, T.J., Munakomi, S., & Neuroanatomy, T. (2023). In: StatPearls [Internet]. Treasure Island, FL: StatPearls Publishing. https://doi . org / 10 . 1080 / 15424065 . 2024 . 2389325 Tourville, J.A., & Guenther, F.H. (2011). The DIVA model: A neural theory of speech acquisition and production. Lang Cogn Process, 26(7), 952–981. https://doi . org / 10 . 1080 / 01690960903498424 Wächter, T., Lungu, O.V., Liu, T., Willingham, D.T., & Ashe, J. (2009). Differential effect of reward and punishment on procedural learning. J Neurosci, 29(2), 436–443. https:// doi . org / 10 . 1523 / jneurosci . 4132 - 08 . 2009 Waschke, L., Kloosterman, N.A., Obleser, J., & Garrett, D.D. (2021). Behavior needs neural variability. Neuron, 109(5), 751–766. https://doi . org / 10 . 1016 / j . neuron . 2021 . 01 . 023 Weiss, Y., Katzir, T., & Bitan, T. (2015). Many ways to read your vowels— Neural processing of diacritics and vowel letters in Hebrew. Neuroimage, 121, 10–19. https://doi . org / 10 . 1016 / j . neuroscience . 2021 . 12 . 025 Wells, G. (1985). Language development in the preschool years (Vol. 2). CUP Archive. https://doi . org / 10 . 1177 / 014272378700702008 Wolfe, D.L. (1967). Some theoretical aspects of language learning and language teaching. Lang Learn, 17(3‐4), 173–188. https://doi . org / 10 . 1111 / j . 1467 - 1770 . 1967 . tb00924 . x Wong, P.C., Perrachione, T.K., & Parrish, T.B. (2007). Neural characteristics of successful and less successful speech and word learning in adults. Hum Brain Mapp, 28(10), 995–1006. https://doi . org / 10 . 1002 / hbm . 20330 Wood, S. (1982). X-ray and model studies of vowel articulation, Working Papers in Linguistics, 23, Department of Linguistics, University of Lund. https:// journals.lub.lu.se/LWPL/article/view/16897/15276 Yan, C. G., Wang, X. D., Zuo, X. N., & Zang, Y. F. (2016). DPABI: Data processing & analysis for (resting-state) brain imaging. Neuroinform, 14(3), 339–351. https://doi .org/10.1007/s12021-016-9299-4 Zevin, J.D. (2012). A sensitive period for shibboleths: The long tail and changing goals of speech perception over the course of development. Dev Psychobiol, 54(6), 632–642. https://doi . org / 10 . 1002 / dev . 20611 Zuk, J., Andrade, P.E., Andrade, O.V., Gardiner, M., & Gaab, N. (2013). Musical, language, and reading abilities in early Portuguese readers. Front Psychol, 4, 288. https://doi . org / 10 . 3389 / fpsyg . 2013 . 00288