parling_dida

1
Pronunciation improvement by means of a CALL
system with ASR technology for
young foreign language learners
Ornella Mich, Ambra Neri, Diego
Giuliani Interactive Sensory Systems, ITC-irst,
Trento, Italy Centre for Speech and Language
Technologies, Radboud University Nijmegen, The
Netherlands
2
Outline

• Background
• Problem
• Goal of study
• Experiment
• Discussion

3
Background

• CALL / CAPT for children is increasingly popular
  among children educators
• multimedia engaging, game-like activities
• task-based learning
• self-paced learning
• abundant oral input
• opportunities for practice
• automatic feedback by means of ASR technology
• All the more important for pronunciation training
  and in foreign language learning context

4

• Classroom FL learning has practical constraints
• Most advanced systems offer feedback that is
  comparable to human feedback

5
Problem

• We also know that ASR technology is not 100
  error-free
• We do NOT know whether CALL systems with ASR are
  actually pedagogically effective Lack of studies
  investigating this issue

6
Goal of study

• To establish whether CALL systems with ASR
  technology can offer training on global
  pronunciation quality whose effectiveness is
  comparable to teacher-led training
• Comparison of experimental group using CALL
  system with ASR-based feedback and control group
  following regular teacher-fronted classes

7
Experiment

• CALL system Parling
• Method
• Subjects Training
• Testing procedure 2 types of ratings
• Results
• Reliability of ratings
• Global pronunciation quality
• Pronunciation quality of specific words
• Conclusions

8
CALL system PARLING (1)

• Target users Italian primary school children
  learning English
• Design based on literature study, indications by
  teachers, 2 usability tests with end users (8-12)

9
CALL system PARLING (2)

• Content 1 visual dictionary, 1 tool to create
  personal dictionary with self-recorded words
  pictures. Divided into modules.
  Each module contains 1 story, 1 set of
  active words (retrieve meaning, try
  pronunciation), 1 adaptive word game.
• ASR trained on speech of 8-10 yrs old
• All operations are logged

10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
Method Subjects Training
Italian 11-year-old children with 4 yrs of
British English FL training, attending same
public school, same English teacher

• Control group (n15) followed 4 teacher-led 60
  classes focussing on 1 story, training vocabulary
  pronunciation of words with teacher
• Experimental group (n13) followed 4 individual
  30 training sessions in language lab with
  Parling, studying 1 story playing word game

14
Method Testing - 2 types of ratings

• Design pretest/posttest
• Stimuli 28 isolated words (read for
  comparability) from story, covering most frequent
  British English phonemes. Some words unknown to
  subjects before training.
• Recordings dedicated software. Children could
  re-do until happy with pronunciation
• 3 expert NS scored recordings (N1656)
  independently on 10-point scale.
• 2 scores requested a) single-word score,
• b) speaker score
• 32 duplicates, anchors.

15
Results Reliability of ratings
16
Results global pronunciation quality

• Correlation between mean of single-word scores
  and speaker score r.884, plt.01
• Speaker scores are a good indication of overall
  quality and can be used for further analyses.
• Group mean speaker scores not significantly
  different at pretest (t.321, p.754)
• Improvement in both groups
• ANOVA (RM) main effect for Test time, no
  significant Training group x Test time
  interaction
• Both groups improved, improvements were
  comparable.

17
Results improvement on specific words

• Known/easy words (n19) e.g. birds, cage, fire,
  food, leave etc.
• Unknown/difficult words (n5) breadcrumbs,
  pushes, woodcutter, etc.
• Group mean improvement on 2 types of words
• ANOVA (RM) with Test time and Word type (WS) and
  Training group (BS) significant effect for
  Test, for Word type, significant Test x Word
  interaction, but NO Test x Word x Training group
  interaction
• the two Training groups make significant and
  comparable improvements on unknown/difficult words

18
Conclusions

• CALL systems including ASR technology and
  stimulating learners to actively use the language
  by pronouncing words appear to be effective in
  improving pronunciation quality. They can thus be
  used as a complement to classroom teaching.
• Because of unfavourable teacher/student ratio,
  CALL system can offer more opportunities for
  practice than traditional classroom teaching and
  undivided attention.

19
Limitations - future research

• Small sample size.
• No investigation of long-term retention
• No investigation of specific aspects of
  pronunciation
• Improve the ASR algorithm to provide specific
  feedback on pronunciation quality
• Possible application for children that are
  lagging behind

20
THANK YOU!
Any questions?
http//tev.itc.it/people/mich/Parling/
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)