Preschoolers’ production of L2 vowels is affected by input quality: A longitudinal study

M. Kučerová, Š. Šimáčková

L2 input

Learners initially use L1 representations to process L2 speech, representations of L2 words affected
Learning: updating and forming L2 representations
Quality of L2 input impacts creation of new representations and their updating [1]
FLA: less L2 input, few L2 speakers, variable accents(?)
Variable native input affects vowel production[2], [3]

[1] Llompart, M. (2021). Phonetic categorization ability and vocabulary size contribute to the encoding of difficult second-language phonological contrasts into the lexicon.

[2] Bohn, O.-S., & Bundgaard-Nielsen, R. L. (2009). Second language speech learning with diverse inputs.

[3] Bosch, L., & Ramon-Casas, M. (2011). Variability in vowel production by bilingual speakers: Can input properties hinder the early stabilization of contrastive categories?

Hi, my name is Monika Kucerova and together with Sarka Simackova, we looked at L2 vowel acquisition by preschoolers, so that is 3-5 year olds. And they were learning L2 English in a classroom setting.

We were interested in how language experience shapes L2 sound learning. So, early stages of L2 learning, learners process L2 words using their existing L1 phonology, they have no alternative. So, their memory representations of the sounds of L2 words will differ from a more experienced speaker’s representations.

Updating the representations is what we call learning, and it also involves forming new representations, which depends on the ability of learners to notice and store the phonetic detail in the input, and it also depends on how much input the learner gets, and what that input is like.

In the situation that a child L2 learner is exposed to the L2 in an L1 dominant environment, some specific constraints on the input may affect the phonological updating in L2 learning. Learners in this setting generally get less L2 input, often from few speakers, who may have very variable accents, which was the case for this study.

Assumptions

Speakers store words as strings of phonemes, and as phonetically-detailed tokens in exemplar clouds. [4] This is also true for L2 learners [6]
Phonological categories emerge from sound exemplar clouds [5]

[4] Nijveld, A., Ten Bosch, L., & Ernestus, M. (2022). The use of exemplars differs between native and non-native listening.

[5] Wilder, R. J. (2018). Investigating hybrid models of speech perception.

Assumptions

Schweitzer, A. (2019). Exemplar-theoretic integration of phonetics and phonology

Exemplar-based Phonetic Space

Schweitzer, A. (2019). Exemplar-theoretic integration of phonetics and phonology: Detecting prominence categories in phonetic space. Journal of Phonetics, 77, 100915.

Assumptions: exemplars and categories

Perceived L2 sounds can be integrated into L1 clouds

compatibility with equivalence classification in SLM-r [7]

Language experience modulates categories:

shift of cloud centre: phonetic drift
creation of new cloud: category formation

[6] Goldinger, S. D. (2007). A complementary-systems approach to abstract and episodic speech perception.

[7] Flege, J. E., & Bohn, O.-S. (2021). The revised speech learning model (SLM-r).

Children’s non-native language phonology

Shared L1~L2 sound system, bidirectional influences (interference [8], drift [9])
One hour of articulatory feedback training [10]
Weaker compactness of L1 categories may lead to stronger drift [10]
This study: children exposed to L2 around age 3 \(\rightarrow\) less developed L1 phonology, weaker crosslinguistic influences, increased likelihood of new category formation.
- BUT no immersion, creating L2 categories may be delayed or prevented

[8] Lee, S. A. S., & Iverson, G. K. (2012). Stop consonant productions of Korean–English bilingual children.

[9] Yang, J., & Fox, R. A. (2017). L1–L2 interactions of vowel systems in young bilingual Mandarin-English children.

[10] Kartushina, N., Hervais-Adelman, A., Frauenfelder, U. H., & Golestani, N. (2016). Mutual influences between native and non-native vowels in production: Evidence from short-term visual articulatory feedback training.

Assuming that learners’ L1 and L2 phonology coexist in a single shared system, we expect to see bidirectional influences. That is, along with interference in the L2, we expect to see phonetic drift in the L1. Our participants received a very limited amount of L2 input (45 minutes weekly for a couple mounts), but there have been laboratory studies showing drift in adults with as little as one hour of exposure, in this case articulatory feedback training. Further, children have been reported to have more variable categories, and more variable categories have been shown to be more susceptible to phonetic drift in adults. These were the main points that made us hypothesise that we could observe drift in these young children, even though their input was limited. When it comes to new category formation, we are looking at children who are 3 to 5 years old, so their L1 phonology is less developed, hence we may see weaker crosslinguistic influences, that would increase the likelihood of new category formation. But these learners have never been immersed, and their limited input may delay or prevent category formation.

This study: input during experiment

In-class input from a teacher with a SSBE-like accent, separated target vowels in quality.
L2 input from parents: reported their English to be Czech-accented
We assumed that input from parents included vowel mergers typical of Moravian Czech: /ɛ, æ/ as /e/, and possibly /i,ɪ/ as /i/

Questions

How does in-class input influence pre-literate FL learners’ production of L2 and L1 vowels in terms of quality?

Are /i,ɪ,ɛ,æ,ʌ/ better separated in the learners’ pronunciation of words heard in class compared to words heard mainly from parents?
Does pronunciation of /i,ɪ,ɛ,æ,ʌ/ in words heard mainly from parents change in time, do the vowels become separated better?
Does production of L2 vowels change over time?

Participants and L2 exposure

7 girls, 3;9 to 5;9
L1 Moravian Czech monolinguals (+ 1 Czech-Slovak bilingual)
AOL between 1;9 and 3;4 yrs; low proficiency in English
Weekly 45 min EFL classes (role-playing, daily activities, dancing, drawing, problem solving); teacher: L1 Czech, L2 English with SSBE-like accent

School-provided audio-visual materials featuring SSBE speakers, media exposure

Data collection

8 recording sessions: 2 in Czech (10 weeks apart), 6 in English
Picture naming task, repetition after the teacher when word not recalled
Mono-/disyllabic words, initial stress
- 39 English words: SSBE /i,ɪ,ɛ,æ,ʌ/ in stressed syllable
- 16 Czech words: /iː,i,e,a/
- CVC context
- Fillers with back vowels (14 English, 8 Czech)
New words from teacher (n=20), Old words from parents (n=19)

We conducted eight recording sessions. We recorded Czech words in two recording sessions, which were ten weeks apart. During fourteen weeks, we had six sessions in which we recorded English words. With each participant, the words were recorded immediately after the English lesson. For both Czech and English words, we used a picture naming task with mono- and disyllabic words that were initially stressed. When the child could not recall a word, the teacher produced it and the child repeated it immediately. The depicted objects corresponded to either words which we labelled „old“ or „new“. New words were those that the children were taught during class throughout the experiment. Old words they learned in class but at least a month before the experiment began, and so had the opportunity to practice with parents. The target vowels in English words were the FLEECE, KIT, DRESS, TRAP and STRUT. In Moravian Czech words, they were the long high front /i:/, short high front /i/, short mid front /e/, short low central /a/.

Measurement and analysis

Manual segmentation in Praat [11]; F0, F1, and F2 measurements from segments of 25–50 ms inside vowel intervals
F1, F2 extracted using Praat built-in LPC with the Burg algorithm
Normalization: ERB in phonR [13]
Linear mixed effects models in R [14] using lme4 [15], fitted by REML using BOBYQA optimization [16]
Vowel height as F1e-F0e, vowel retraction as F2e-F0e

[11] Boersma, P., & Weenink, D. (2022). Praat: doing phonetics by computer.

[13] McCloy, D. R. (2016). phonR: Tools for phoneticians and phonologists.

[14] R Core Team (2023). R: A language and environment for statistical computing.

[15] Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4.

[16] Powell, M. J. (2009). The BOBYQA algorithm for bound constrained optimization without derivatives.

English vowels

F1e-F0e~Vowel*Time*Input+(1|Participant)+(1|Word)
F2e-F0e~Vowel*Time*Input+(1|Participant)+(1|Word)
Time (levels: T1, T2), Input (levels: New, Old), Vowel: factor predictor with levels /i,ɪ,ɛ,æ,ʌ/
Intercept: /æ/ in New words at Time 2
771 vowel tokens from 39 words spoken by 7 children were provided to each of two models

Are the learners’ L2 vowels better separated acoustically in newly learned words?
/æ/ and /ʌ/, /i/ and /ɪ/ separated to a similar degree in all words (words form class and from home were pronounced similarly)
Better separation of /æ/ and /ɛ/ in height in New words (/ɛ/ was raised in New words)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: F1e - F0e ~ Vowel * Time * Input + (1 | Speaker) + (1 | Word)
   Data: efl_En
Control: lmerControl(optimizer = "bobyqa")

REML criterion at convergence: 2910.7

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-4.2344 -0.5884  0.0313  0.6212  4.6127 

Random effects:
 Groups   Name        Variance Std.Dev.
 Word     (Intercept) 0.007136 0.08447 
 Speaker  (Intercept) 0.547916 0.74021 
 Residual             2.506805 1.58329 
Number of obs: 771, groups:  Word, 39; Speaker, 7

Fixed effects:
                      Estimate Std. Error        df t value Pr(>|t|)    
(Intercept)            8.17908    0.42215  25.34662  19.375  < 2e-16 ***
Vowelɛ                -1.65913    0.42750  65.25649  -3.881 0.000245 ***
Voweli                -6.81817    0.42429  70.57411 -16.070  < 2e-16 ***
Vowelɪ                -5.83383    0.41642  63.21942 -14.010  < 2e-16 ***
Vowelʌ                -0.60507    0.44187  77.53509  -1.369 0.174854    
TimeT1                 0.18895    0.44501  97.81900   0.425 0.672060    
Input2                -0.67679    0.36795  56.71397  -1.839 0.071101 .  
Vowelɛ:TimeT1          0.07196    0.63277  64.55410   0.114 0.909812    
Voweli:TimeT1          1.25329    0.62458  93.20023   2.007 0.047687 *  
Vowelɪ:TimeT1          0.39512    0.62386  82.57810   0.633 0.528257    
Vowelʌ:TimeT1          0.65275    0.63346  87.78026   1.030 0.305625    
Vowelɛ:Input2          1.33467    0.50248  50.62762   2.656 0.010543 *  
Voweli:Input2          0.54214    0.60875  45.82272   0.891 0.377804    
Vowelɪ:Input2          0.68234    0.59754  37.39647   1.142 0.260749    
Vowelʌ:Input2          0.45141    0.59827  41.91904   0.755 0.454745    
TimeT1:Input2         -0.01154    0.50845 140.28707  -0.023 0.981925    
Vowelɛ:TimeT1:Input2  -0.09302    0.71815  94.29086  -0.130 0.897221    
Voweli:TimeT1:Input2  -0.70617    0.83835 240.29708  -0.842 0.400441    
Vowelɪ:TimeT1:Input2  -0.43778    0.81398 201.45291  -0.538 0.591288    
Vowelʌ:TimeT1:Input2   0.35259    0.82006 202.79438   0.430 0.667686    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: F2e - F0e ~ Vowel * Time * Input + (1 | Speaker) + (1 | Word)
   Data: efl_En
Control: lmerControl(optimizer = "bobyqa")

REML criterion at convergence: 2605.1

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-4.7869 -0.5532  0.0743  0.6249  2.7322 

Random effects:
 Groups   Name        Variance Std.Dev.
 Word     (Intercept) 0.05589  0.2364  
 Speaker  (Intercept) 0.12434  0.3526  
 Residual             1.66113  1.2888  
Number of obs: 771, groups:  Word, 39; Speaker, 7

Fixed effects:
                      Estimate Std. Error        df t value Pr(>|t|)    
(Intercept)           14.58216    0.31940  38.38354  45.655  < 2e-16 ***
Vowelɛ                 0.46909    0.39701  30.08963   1.182  0.24664    
Voweli                 3.00112    0.39195  32.52439   7.657 8.98e-09 ***
Vowelɪ                 1.80708    0.38756  29.14001   4.663 6.41e-05 ***
Vowelʌ                -1.27638    0.40621  34.54353  -3.142  0.00343 ** 
TimeT1                 0.27377    0.40103  55.62807   0.683  0.49764    
Input2                 0.09848    0.34311  30.00992   0.287  0.77607    
Vowelɛ:TimeT1          0.38320    0.58404  37.11399   0.656  0.51579    
Voweli:TimeT1         -0.23492    0.56383  54.85623  -0.417  0.67856    
Vowelɪ:TimeT1          0.18168    0.56722  47.83518   0.320  0.75014    
Vowelʌ:TimeT1         -0.25691    0.57428  49.75271  -0.447  0.65656    
Vowelɛ:Input2          0.35508    0.47224  26.98158   0.752  0.45862    
Voweli:Input2         -0.64233    0.57795  23.02039  -1.111  0.27787    
Vowelɪ:Input2         -0.49674    0.57554  19.71051  -0.863  0.39848    
Vowelʌ:Input2         -0.79904    0.57098  21.97747  -1.399  0.17565    
TimeT1:Input2         -0.01022    0.45007  79.40295  -0.023  0.98194    
Vowelɛ:TimeT1:Input2  -0.63259    0.64889  52.12673  -0.975  0.33413    
Voweli:TimeT1:Input2   0.21649    0.72600 135.83916   0.298  0.76601    
Vowelɪ:TimeT1:Input2  -0.58373    0.71009 109.80786  -0.822  0.41283    
Vowelʌ:TimeT1:Input2   0.76005    0.71520 110.14986   1.063  0.29024    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Our first question was whether the learners’ vowels were better separated acoustically in newly learned words.

The green TRAP and red STRUT vowels were differentiated to a similar degree in both new and old words, we did not see a significant difference between STRUT in old and new words. There was also no sufficient evidence of differences between the FLEECE and KIT vowels in new compared to old words.

However, the TRAP and DRESS vowels were better separated in newly learned words. You can see this in the plot: this are vowels in Old words, and these are vowels in New words. In New words, /ef/ is higher, which is how that larger separation of TRAP and DRESS was achieved. This suggests that class input contributed to a better separation of the two vowels.

The e-raising that we see here reflects the teacher’s pronunciation (which we also recorded). The raising suggests that the children were able to attend to the phonetic detail in the input and also refer to it in their own production. However, even in the New words, we see indicators of lexical misrepresentation: some /ae/ tokens we [ef] like. This has been reported in Czech learners. The fact that we also see this in New words, which we assume the children have heard from the teacher primarily, suggests that the children’s production may not be entirely exemplar based, or that if it is, that the exemplars contribute to a larger less specified category possibly consisting of DRESS and TRAP tokens, which the children may have processed as free variants rather than belonging to distinct categories.

L2 /ɛ, æ, ʌ/ produced by the learners in Old words. Ellipses include 68% of tokens.

There was no sufficient evidence of differences between the FLEECE and KIT vowels in new compared to old words.

Here we have the low vowels. The green TRAP and red STRUT vowels were differentiated to a similar degree in both new and old words, we did not see a significant difference between STRUT in old and new words.

However, the TRAP and DRESS vowels were better separated in newly learned words. The DRESS V raised. This suggests that class input contributed to a better separation of the two vowels.

The e-raising that we see here reflects the teacher’s pronunciation (which we also recorded). The raising suggests that the children were able to attend to the phonetic detail in the input and also refer to it in their own production. However, even in the New words, we see indicators of lexical misrepresentation: some /ae/ tokens we [ef] like. This has been reported in Czech learners. The fact that we also see this in New words, which we assume the children have heard from the teacher primarily, suggests that the children’s production may not be entirely exemplar based, or that if it is, that the exemplars of both DRESS and TRAP contribute to a larger less specified category possibly consisting of DRESS and TRAP tokens, which the children may have processed as free variants rather than belonging to distinct categories.

L2 /ɛ, æ, ʌ/ produced by the learners in New words. Ellipses include 68% of tokens.

Input effects: low vowels

/ɛ/-raising reflects the teacher’s pronunciation, children able to attend to the phonetic detail in the input as well as to refer to the recently perceived input in their own production
Lexical misrepresentation (some /æ/ tokens [ɛ]-like, also in [17]
Even in New words (exemplar vs. abstract representations)

[17] Šimáčková, Š., & Podlipský, V. J. (2018). Production accuracy of L2 vowels: Phonological parsimony and phonetic flexibility.

The e-raising that we see here reflects the teacher’s pronunciation (which we also recorded). The raising suggests that the children were able to attend to the phonetic detail in the input and also refer to it in their own production. However, even in the New words, we see indicators of lexical misrepresentation: some /ae/ tokens we [ef] like. This has been reported in Czech learners. The fact that we also see this in New words, which we assume the children have heard from the teacher primarily, suggests that the children’s production may not be entirely exemplar based, or that if it is, that the exemplars of both DRESS and TRAP contribute to a larger less specified category possibly consisting of DRESS and TRAP tokens, which the children may have processed as free variants rather than belonging to distinct categories.

Time effects: low vowels

Does production of L2 vowels change over time?

No sufficient evidence for change in /ɛ, æ, ʌ/

Our second question was whether production of L2 vowels changes over time. We didn’t find evidence of DRESS and TRAP changing in time. This indicates that the more separated production of DRESS and TRAP in new words did not contribute to a better separation of the two vowels in old words. For the old words, children heard them produced mainly from the parents, but learned them from the teacher. They may have considered TRAP and DRESS tokens to be free variants, because they heard them in a the same word spoken by both input sourcea, so e.g. “bat” would be produced as “bet” by parents and “bat” by the teacher. For TRAP and STRUT, it’s also possible that the teacher’s production didn’t provide tokens that were different enough for the learners to learn from that input.

Low English vowels at T1 (blue) and T2 (red)

Time effects: low vowels

No evidence of /ɛ/ raising: no generalisation of new experience to old representations?
Spectral distinction in teacher’s input not substantial enough to change the learners’ production of /æ, ʌ/? (consistent separation in all words throughout the experiment)

Time effects: high vowels

/i/ raised at Time 2
/ɪ/ became more variable at Time 2

L2 /i, ɪ/ produced by the learners at Time 1 (blue) and Time 2 (pink) across all words. Ellipses include 68% of tokens.

Czech Vowels: phonetic drift

Results
summary(): V height
summary(): V retraction

Does production of L1 vowels change over time?

Vowel height: F1e-F0e~Vowel*Time+(1|Participant)+(1|Word)
Vowel retraction: F2e-F0e~Vowel*Time+(1|Participant)+(1|Word)
4 vowels, 19 word types, 201 tokens total
/i/ raised, /iː/ raised
/e/ retracted, /a/ retracted
the relative distance between /e/ and /a/ remained similar

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: F1e - F0e ~ Vowel * Time + (1 | Speaker) + (1 | Word)
   Data: dataCz
Control: lmerControl(optimizer = "bobyqa")

REML criterion at convergence: 692.7

Scaled residuals: 
     Min       1Q   Median       3Q      Max 
-2.87716 -0.70084  0.05063  0.69690  2.13584 

Random effects:
 Groups   Name        Variance Std.Dev.
 Word     (Intercept) 0.1074   0.3277  
 Speaker  (Intercept) 0.2007   0.4480  
 Residual             1.7279   1.3145  
Number of obs: 201, groups:  Word, 19; Speaker, 7

Fixed effects:
               Estimate Std. Error      df t value Pr(>|t|)    
(Intercept)      6.9630     0.3548 16.7963  19.625 5.14e-13 ***
Vowela           1.7117     0.4621  9.9595   3.704  0.00411 ** 
Voweli          -3.9927     0.4565 10.9725  -8.746 2.82e-06 ***
Voweliː         -4.8652     0.4699 11.4315 -10.353 3.75e-07 ***
TimeT2           1.0453     0.4878 12.4126   2.143  0.05257 .  
Vowela:TimeT2   -1.1694     0.6879  9.8533  -1.700  0.12044    
Voweli:TimeT2   -2.6379     0.6753 11.1618  -3.907  0.00238 ** 
Voweliː:TimeT2  -1.8968     0.7096 11.4791  -2.673  0.02097 *  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1


Correlation of Fixed Effects:

            (Intr) Vowela Voweli Vowelː TimeT2 Vowela:TmT2 Voweli:TmT2
Vowela      -0.593                                                    
Voweli      -0.597  0.457                                             
Voweliː     -0.580  0.445  0.466                                      
TimeT2      -0.557  0.425  0.431  0.420                               
Vowela:TmT2  0.396 -0.668 -0.305 -0.298 -0.707                        
Voweli:TmT2  0.403 -0.307 -0.673 -0.315 -0.722  0.510                 
Vowelː:TmT2  0.382 -0.291 -0.306 -0.662 -0.687  0.486       0.504

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: F2e - F0e ~ Vowel * Time + (1 | Speaker) + (1 | Word)
   Data: dataCz
Control: lmerControl(optimizer = "bobyqa")

REML criterion at convergence: 592.5

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-3.6424 -0.6003  0.0871  0.5569  1.9062 

Random effects:
 Groups   Name        Variance Std.Dev.
 Word     (Intercept) 0.02976  0.1725  
 Speaker  (Intercept) 0.24645  0.4964  
 Residual             1.02201  1.0109  
Number of obs: 201, groups:  Word, 19; Speaker, 7

Fixed effects:
               Estimate Std. Error      df t value Pr(>|t|)    
(Intercept)     15.6148     0.2864 16.2353  54.513  < 2e-16 ***
Vowela          -1.6422     0.3143 10.5500  -5.225 0.000325 ***
Voweli           1.0790     0.3122 11.4783   3.457 0.005051 ** 
Voweliː          2.4007     0.3220 12.0443   7.455 7.52e-06 ***
TimeT2          -0.9183     0.3366 13.9409  -2.728 0.016376 *  
Vowela:TimeT2   -0.2385     0.4675 10.5500  -0.510 0.620478    
Voweli:TimeT2    0.4483     0.4624 11.9056   0.969 0.351600    
Voweliː:TimeT2   0.8686     0.4871 12.5815   1.783 0.098677 .  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1


Correlation of Fixed Effects:

            (Intr) Vowela Voweli Vowelː TimeT2 Vowela:TmT2 Voweli:TmT2
Vowela      -0.521                                                    
Voweli      -0.521  0.472                                             
Voweliː     -0.504  0.459  0.470                                      
TimeT2      -0.480  0.434  0.439  0.426                               
Vowela:TmT2  0.346 -0.667 -0.314 -0.307 -0.716                        
Voweli:TmT2  0.350 -0.315 -0.671 -0.316 -0.727  0.521                 
Vowelː:TmT2  0.331 -0.298 -0.307 -0.660 -0.690  0.495       0.506

Vowels produced in Czech words at T1 (green) and T2 (red). Ellipses use 68% CIs.

Phonetic drift: high L1 and L2 vowels

L1 /iː, i/ came closer to L2 /i/ in height (large overlap)
the distance between L1 /iː, i/ and L2 /ɪ/ increased (category centres)

English (blue) and Czech (red) high vowels produced by learners at T1.

English (blue) and Czech (red) high vowels produced by learners at T2.

Phonetic drift: low vowels

L1 /e/ assimilated to L2 /æ/
L1 /e/ dissimilated from L2 /ɛ/ in New words
L1 /a/ dissimilated from L2 /æ/
L1 /a/ assimilated to L2 /ʌ/

For the low vowels, The L1 low central vowel retracted, and so became more dissimilar from L2 TRAP, and more similar to L2 STRUT. Which might be an effect of input. The L1 mid front vowel assimilated to the L2 TRAP by retracting.

With the DRESS vowel, the picture becomes more complicated, because we observed a significant difference between the DRESS in new words, which was higher, and DRESS in old words, which was lower. So, because the L1 mid front vowel lowered, it dissimilated from the L2 DRESS but only in new words.

We can interpret this as Exemplar effects: DRESS and TRAP may have formed one category, and DRESS was being pulled to a lower position by the incoming TRAP exemplars, but input effects caused DRESS to raise, as the chidlren noticed the differences between TRAP and DRESS in the new input (and could not move TRAP any lower)

English /ɛ, æ, ʌ/ in New words at T1 and Czech /e/ at T1

English /ɛ, æ, ʌ/ in New words at T2 and Czech /e/ at T2.

Summary

In-class input influences pre-literate FL learners’ production of L1 and L2 vowels - Recent input affected the learners’ L2 output - New words: /ɛ/ and /æ/ were separated in quality due to the raising of /ɛ/.

No evidence for a cumulative effect of classroom exposure on the production of Old words. - /ɛ/-raising from New words not present in Old words

Production of L1 vowels shifted over time. - learners’ vowels undergo continual changes throughout L2 exposure, even in time constrained FL classroom exposure

Limitations

Parental input not controlled
Amount of exposure necessary for updating representations of Old words?
Children were worse at remembering New words than Old words
Perception test
No tracking of New word development

There are some important limitations to this study, that may help us guide in future research. We did not control the exposure to old words. We dont know much exposure to old words produced by the teacher would sufficie for the hcildren to produce a higher DRESS in old words, so to generalsie the split across their vocabulary.

We did not see generalisation of new experience to vowels in old words. So, does it také many more exemplars of new words to cause generalisation (because of a cloud shift) or does it take more old words heard from the teacher to bias production towards vowels similar those in new words?

We also had less data on New words, because the children failed to remember those much more often than Old words. And we did not track New word development, we only recorded each New word twice in three weeks and then moved onto newer New words.

It would be interesting to see how perceptual categories look and evolve.

References

[1] Llompart, M. (2021). Phonetic categorization ability and vocabulary size contribute to the encoding of difficult second-language phonological contrasts into the lexicon. Bilingualism: Language and cognition, 24(3), 481-496. doi:10.1017/S1366728920000656

[2] Bohn, O.-S., & Bundgaard-Nielsen, R. L. (2009). Second language speech learning with diverse inputs. In T. Piske, & M. Young-Scholten (Eds.), Input matters in SLA (pp. 207-218). Multilingual Matters. https://pure.au.dk/portal/en/publications/second-language-speech-learning-with-diverse-inputs

[3] Bosch, L., & Ramon-Casas, M. (2011). Variability in vowel production by bilingual speakers: Can input properties hinder the early stabilization of contrastive categories? Journal of Phonetics, 39(4), 514–526. doi:10.1016/j.wocn.2011.02.001

[4] Nijveld, A., Ten Bosch, L., & Ernestus, M. (2022). The use of exemplars differs between native and non-native listening. Bilingualism: Language and Cognition, 25(5), 841-855. doi:10.1017/S1366728922000116

References

[5] Wilder, R. J. (2018). Investigating hybrid models of speech perception. Publicly Accessible Penn Dissertations. 3202. https://repository.upenn.edu/edissertations/3202

[6] Goldinger, S. D. (2007). A complementary-systems approach to abstract and episodic speech perception. In Proceedings of the 16th international congress of phonetic sciences (pp. 49-54).

[7] Flege, J. E., & Bohn, O.-S. (2021). The revised speech learning model (SLM-r). In R. Wayland (Ed.), Second language speech learning: Theoretical and empirical progress (pp. 3–83). Cambridge University Press. doi:10.1017/9781108886901

[8] Lee, S., Potamianos, A., & Narayanan, S. (1999). Acoustics of children’s speech: Developmental changes of temporal and spectral parameters. The Journal of the Acoustical Society of America, 105(3), 1455-1468. doi:10.1121/1.426686

References

[9] Yang, J., & Fox, R. A. (2017). L1–L2 interactions of vowel systems in young bilingual Mandarin-English children. Journal of Phonetics, 65, 60-76. doi:10.1515/phon-2021-2006

[11] Boersma, P., & Weenink, D. (2022). Praat: doing phonetics by computer [Computer software]. Version 6.2.23. http://www.praat.org/

[13] McCloy, D. R. (2016). phonR: Tools for phoneticians and phonologists. R package version 1.0-7. Online: https://cran.r-project.org/web/packages/phonR/phonR.pdf

[14] R Core Team (2023). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/

References

[15] Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1-48. doi:10.18637/jss.v067.i01

[16] Powell, M. J. (2009). The BOBYQA algorithm for bound constrained optimization without derivatives. Cambridge NA Report NA2009/06. University of Cambridge, Cambridge, 26. https://www.damtp.cam.ac.uk/user/na/NA_papers/NA2009_06.pdf

[17] Šimáčková, Š., & Podlipský, V. J. (2018). Production accuracy of L2 vowels: Phonological parsimony and phonetic flexibility. Research in Language, 16(2). doi:10.2478/rela-2018-0009

Thank you for your attention

kucerova@psu.cas.cz