WEB-BASED HIGH VARIABILITY PHONETIC TRAINING ON L2 CODA .

WEB-BASED HIGH VARIABILITY PHONETIC TRAININGON L2 CODA IDENTIFICATIONNa-Young Ryu and Yoonjung KangUniversity of Torontonayoung.ryu@mail.utoronto.ca, yoonjung.kang@utoronto.caABSTRACTThis study investigates the effects of web-based highvariability phonetic training (HVPT) on thedevelopment of L2 coda identification.45 Mandarin learners of Korean were randomlyassigned to receive coda-focused or vowel-focusedtraining (both of which contained identical auditoryinput), or no training. Performance was assessedusing an identification task before and after training,and a generalization test after training.The results demonstrate that the training groupsoverall improved their perception of Korean codascompared to the no-training group at post-test,suggesting that L2 learners benefit from online HVPT.Between the two training groups, the coda-focusedgroup performed significantly better than the vowelfocused group. This finding indicates that perceptiontraining for L2 sounds can lead to greaterimprovement when learners are consciously aware oftheir target sounds than by passive exposure to thetarget sounds. It was also found that learning effectsgained from HVPT can be successfully transferred tonew phonetic contexts.Keywords: High variability phonetic training,identification task, trained vs. untrained L2 codas1. INTRODUCTIONA vast literature has reported that adult secondlanguage (L2) learners have difficulty perceiving andproducing non-native sounds. In this regard, there hasbeen a great deal of interest in developing perceptualtraining methods and examining to what extent adultL2 learners can improve their perception andproduction of L2 sounds. Specifically, it has beenfound that L2 learners can benefit from highvariability phonetic training (HVPT) in improvingtheir perception and production of L2 speech sounds[1, 3, 4, 9, 11, 17]. The HVPT method proposed byLogan et al. [11] uses multiple voices to providenatural phonetic variability within a phoneticcategory so that L2 learners are able to detect crucialphonetic cues associated with target sounds as well ascues relevant to their category identification. As aresult of HVPT, it was also found that L2 learnerswho receive training are able to generalize their482learning to untrained speech materials [1, 3, 4, 9, 11,17].Previous HVPT studies have been mostlylaboratory-based, where the learners participate intraining in a controlled condition while beingmonitored by the experimenter. The current studyexamines whether positive effects can also beobserved in web-based training, where the learnerscan complete the training sessions at a time, location,and pace of their own choice. Developing onlineHVPT for this study was motivated by meta-analyticreview of 25 years of perception training research bySakai and Moorman [16], who reported that thelocation of training encourages greater gains whencompleted at home versus in the laboratory.In terms of the effectiveness of L2 instruction,Norris and Ortega [13] demonstrated that explicittypes of instruction are more effective than implicittypes. However, recent studies have found that adultlearners can acquire new category contrasts throughpassive exposure to distributional patterns in thelearning input [6, 7, 8]. The current study comparesthe relative effectiveness of explicit and implicitinstruction in L2 perception.The meta-analysis of L2 pronunciation reportedby Lee et al. [10] showed that 97% of the trainingstudies analyzed English as a first or target language.In order to determine how HVPT works among thediversity in language learning, it is imperative toincrease the range of languages. Extending previousresearch, therefore, the current study examines howMandarin learners of Korean improve theirperception accuracy of Korean codas through twotypes of web-based identification HVPT.This study addresses the following specificquestions:1. To what extent do Mandarin L2 learnersimprove their perceptual accuracy inidentifying Korean codas as a result of onlineHVPT?2. Do different types of HVPT (coda-focused vs.vowel-focused) affect outcomes of L2 codaperception differently, even when the sameL2 input is used?

Table 1: Design of the study.3. Can learning effects gained from HVPT besuccessfully transferred to novel stimuli?Phase2. METHOD2.1. Participants45 native Mandarin speakers (39 females, 6 males,mean age 20.7 years old, SD 2.3 years) who wereenrolled in beginner Korean language coursesparticipated in the experiment. They were randomlyassigned to one of the three groups: coda-focusedtraining, vowel-focused training and no training (n 15 per group). The coda-focused group was trained toidentify Korean codas, while the vowel-focusedgroup was exposed to the same input as the codafocused group, but instead participated in a vowelidentification task, in order to direct their attention tovowels. Crucially, this design allowed us to maintainequivalent training input and time between the twotraining conditions, while differing only the type oftraining. The no-training group acted as a controlgroup. For their participation, each member of thetraining groups was paid 65 (CAD), and eachmember of the no-training group was paid 25 (CAD).None of the participants reported hearingimpairments.2.2. Experimental designThe current study consisted of four phases: (1) apre-test, (2) online training, (3) a post-test and (4) ageneralization test. All tests and training involved thesame identification task utilizing the full sets ofKorean codas and vowels as suggested by Nishi andKewley-Port [12], who found that training larger setsof phonemes demonstrated greater gains than trainingonly the most difficult subset of phonemes.Participants individually participated in pre-, postand generalization tests built in PsychoPy [14].Participants were given no feedback regarding theaccuracy of their responses for the tests. Onlinetraining was programmed with jsPsych, which is aJavaScript library for running experiments in a webbrowser [5].Between the pre- and post- test, the two traininggroups completed eight online training sessions in acomfortable setting over a 2-week period. Eachsession was self-paced and typically took betweenfifteen and twenty minutes to complete. Immediatefeedback was provided after each token indicatingwhether their answer was “correct” or “incorrect”. Ifincorrect, they had to try again until they gave acorrect response. Other than the different responsechoices (coda vs. vowel), all aspects of theexperimental design, including input and feedback,were identical between the two training groups. Table1 summarizes the design of this experiment.GroupFeedbackPlatform& locationPre-testAllNoPsychoPygroupsin the oPsychoPygroupsin the labGeneralizCODA,NoPsychoPyationNONEin the labNote. CODA Coda-focused group, VOWEL vowel-focused group, NONE No-training group.2.3 StimuliPre-test, post-test and training stimuli contained 49monosyllabic CVC words in which the initialconsonant was /h/, the vowel was one of the sevenKorean vowels /a e i o u ɨ ʌ/, and the final consonantwas one of the seven codas /t k p n m ŋ l/. In thegeneralization task, the stimuli consisted of KoreanCVC words in which the initial consonant was /k/.The stimuli for the online training were recordedby four native Korean speakers (2 females, 2 males)for a total of 196 tokens (49 stimuli * 4 speakers) tocreate a high-variability training condition. Thestimuli for the pre-, post- and generalization testswere recorded by two native Korean speakers (1female, 1 male, 98 tokens 49 stimuli * 2 speakers).These speakers were different from the trainingstimuli speakers to ensure that the perceptionimprovement gained in the training would not bespecific to the tokens or speakers that the learnerswere exposed to in the training, but rather theimprovement, if any, would be more general.2.4. ProcedureThe participants were instructed to listen to wordsthrough headphones and were asked to identify thestimulus that matched one of seven visual targetspresented in Hangul, the Korean alphabet, on acomputer screen by pressing the correspondingnumber from 1 to 7 on the keyboard. For example,participants in the coda-focused group heard [han]and were asked to identify what the coda was (targetresponse: n ), while the vowel-focused group wasasked to identify the vowel (target response: a )during online training. Participants had to press oneof the seven options before the next stimulus waspresented.483

Figure 1: Identification task in pre-test, post-test,online training and generalization test.Korean vowelsKorean codas2.4. Statistical analysisThe perception performance of Korean codas wasassessed using a mixed-effects logistic regressionmodel with the lme4 package [2] in R [15]. Thedependent variable was whether the response wascorrect or incorrect. As fixed effects, we includedGROUP (coda-focused, vowel-focused, or no training)and TEST (pre-, post-, or generalization test) and theirinteraction. TEST was dummy-coded and GROUP wasHelmert-coded to (i) compare the explicit and implicittraining groups and (ii) compare training groupsagainst the no-training group. As random effects, wehad intercepts for subjects and items, as well as a bysubject random slope for TEST and by-item randomslopes for TEST and GROUP.3. RESULTS AND DISCUSSIONimproved significantly compared to the no-traininggroup (β 0.368, z 2.410, p 0.015), whichsuggests that training groups yielded significantimprovement as a result of online HVPT. Thispositive effect of web-based HVPT is consistent withprevious studies of laboratory-based HVPT [1, 3, 4, 9,11, 17].In addition, there was a significant difference inidentification improvement between the two traininggroups (β 0.651, z 3.651, p 0.001), showing thatthe coda-focused training was more effective than thevowel-focused training in improving perception ofKorean codas. This finding indicates that differenttypes of training on L2 sounds lead to differentdegrees of improvement. More specifically, the codatrained group significantly improved their Koreancoda perception from the pre-test to the post-test (byapproximately 10%), whereas the vowel-trainedgroup, which was only passively exposed to Koreancodas, had less improvement of coda perception(approximately 3%). These results provide promisingevidence that perception training of L2 sounds ismore effective through deliberate listening withfeedback than through passive exposure.3.2 Development of L2 coda perception during onlineHVPTFigure 2 shows the pre- and post-test results beforeand after the web-based training for each group.Figure 3 presents a visual illustration of the codafocused group’s degree of improvement in Koreancoda perception accuracy from session 1 to session 8of online HVPT.Figure 2: Boxplots of identification accuracy forKorean codas at pre-test (white boxes) and post-test(grey boxes) after eight training sessions by group.Figure 3: Coda-focused group’s development ofKorean coda identification accuracy during eighttraining sessions.As seen in Figure 2, at pre-test, the three groups didnot differ significantly in the identification scores ofKorean codas (two training groups vs. no-traininggroup: β -0.225, z -1.024, p 0.305, coda-focusedgroup vs. vowel- focused group: β 0.003, z 0.014,p 0.989), indicating that participants had similarproficiency in perceiving the target sounds beforethey received training.There were significant interactions between TESTand GROUP. The two training groups as a wholeOverall, there was a significant increase in perceptionaccuracy of Korean codas between session 1 andsession 3, followed by a stable upward trend acrossthe remaining sessions. This might be due to the factthat immediate feedback was helpful for L2 learnersin noticing their errors and shifting attention torelevant acoustic cues to improve the ability ofidentification of their target sounds during HVPT.3.1 Effects of online HVPT on L2 codas484