Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Rapid accent adaptation and constraints on cross-talker generalization!Kodi Weatherholtz, Linda Liu, and T. Florian Jaeger, University of Rochester{kweatherholtz,, lliu, fjaeger} @bcs.rochester.edu!
To achieve robust language understanding, listeners must adjust for variability in speech (e.g., dialect, foreign accent). Listeners can adapt to systematic variability across talkers [1,2]. To the extent that listeners can generalize learning from familiar talkers to new similar talkers, they are able to benefit from prior accent experience when encountering a new talker [3]. Such cross-talker generalization requires that listeners distinguish accent-general patterns from talker-specific idiosyncracies.!
Acknowledgments:!This work was supported by an NIHCD R01 (HD075797) to TFJ. We are grateful to Ann Bradlow for providing access to the ALLSSTAR corpus for stimulus materials.!
References:![1] Bradlow, A., & Bent, T. (2008). Cognition, 106(2).![2] Sidaras, S., Alexander, J., & Nygaard, L. (2009). JASA, 125 (5).[3] Kleinschmidt, D. & Jaeger, T. F. (2015). PsychReview, 122(2.!!
Previous findings:!Exposure to multiple talkers with the same foreign accent facilitates—and is perhaps necessary for—generalization to new talkers with that accent (Bradlow & Bent, 2008). !By contrast, exposure to a single talker appears to result in talker-specific adaptation (i.e., better performance on the trained talker, but no generalization to new talkers). !
Exposure-test paradigm:!Task: Sentence intelligibility in noise (following Bradlow & Bent, 2008), conducted on Mechanical Turk.!
Stimuli: Simple declarative sentences, each containing 2-4 keywords, produced by native English or Mandarin-accented English talkers.! - Somebody stole the money.!
Exp1: Replicating foreign accent adaptation via the web! Exp 2 (N = 156): Is multi-talker training necessary for generalization?!
0.4
0.6
0.8
1.0
0 20 40 60 80Trial
Prop
ortio
n ke
ywor
ds c
orre
ctly
tran
scrib
ed
Condition●
●
●
●
Control (5 AmEng)Single talker trainingMulti−talker trainingTalker−specific training
Training data by trial
0.4
0.6
0.8
1.0
Control(training =5 AmEng)
Singletalker
training
Multi−talker
training
Talker−specifictraining
Accuracy at Test
0.4
0.6
0.8
1.0
0 20 40 60 80Trial
Prop
ortio
n ke
ywor
ds c
orre
ctly
tran
scrib
ed
Condition●
●
Control (5 AmEng)Talker−specific training
Training data by trial
●
●
0.4
0.6
0.8
1.0
Control(training = 5 AmEng)
Talker−specifictraining
Accuracy at Test
Results!- Training: participants who heard foreign-accented speech were
initially less accurate (relative to control) but improved with exposure.!- Test: higher accuracy in the talker-specific condition, relative to
control, indicating accent adaptation (beyond task adaptation)!- Effect size for talker-specific adaptation ~10%; comparable to B&B2008!
Training materials. 16 sentences repeated 5 times! - by 5 different native speakers, or! - by 1 Mandarin-accented talker! (80 training trials, cf. 160 in B&B2008)!
Non−native 01 Non−native 02 Non−native 03 Non−native 04 Non−native 05 Non−native 06
−1.6% −2.4% +8.3% +11.4% +12.8% +16.2%0.4
0.6
0.8
1.0
test phase test phase test phase test phase test phase test phase
Prop
ortio
n ke
ywor
ds c
orre
ctly
tran
scrib
ed
Condition Control(training = 5 AmEng)
Talker−specifictraining
Training! Test!
Con
trol
!Ta
lker
-spe
cific
!
native (AmEng)
x 5!
non-native
speaker !
Conditions from Exp1!
Training! Test!
Sing
le-ta
lker
!M
ulti-
talk
er
!
non-native
x 5!
non-native
speaker !
different non-
native speaker
!
Conditions.!
Test materials. 16 sentences produced by a single Mandarin-accented talker.!- (identity of the test talker varied btw subjs;
6 different Mandarin-accented talkers total)!
Conditions.!
somebody stole the money
Stimulus w/o noise!
Stimulus in noise (+5 signal-to-noise ratio)!Noise added to avoid ceiling effects.!
Improved performance (adaptation)—or not—by test talker.!
For some talkers, performance was at ceiling regardless of exposure!
Exposure provided no reliable benefit, despite lack of ceiling effect. !
Talkers were relatively hard to understand, but easy to adapt to.!
Conclusions!We found evidence that generalization of accent adaptation is both more robust and less robust than previously claimed (cf. Bradlow & Bent, 2008). !
- More robust in that single talker training can be sufficient for cross-talker generalization (i.e., multi-talker training is not a necessary condition)!
- Less robust in that generalization was imperfect (i.e., cross-talker generalization was never as strong as talker-specific training)!
Our question!Under what conditions does adaptation to a foreign accent generalize to new talkers with that accent?!
*! n.s.!
Outcome variable: proportion of keywords correctly transcribed. !
Author's personal copy
A two-way repeated measures ANOVA with post-test (Chinese-accented vs. Slo-vakian-accented) as a within-subjects factor and training condition as a between-sub-jects factor (Conditions 1, 2, 3, 4 and 5 as shown in Table 2) showed significant maineffects of both post-test (F(l, 82) = 92.90, p < 0001) and training condition(F(4,82) = 14.036, p < 0001). The post-test · training condition interaction was alsosignificant (F(4,82) = 2.860, p < .03). Separate one-factor ANOVAs for each of thetwo post-tests showed significant effects of training condition (Chinese-accentedtalker: F(4,82) = 14.892, p < .0001); (Slovakian-accented talker: F(4,82) = 6.211,p < .001). For the post-test with the Chinese-accented talker (post-test 1), pair-wisecomparisons (1-tailed, t-tests), showed that performance following the no trainingcontrol condition (Condition 5) was significantly worse (at the p < .0001 level) thanperformance following all of the other conditions. Performance on post-test 1 fol-lowing the task control condition (Condition 4) did not differ significantly from per-formance following the single talker training condition (Condition 3), but waspoorer than performance following the talker-specific training condition (p < .02)and the multiple-talker training condition (p < .05) Similarly, performance onpost-test 1 following the single talker training condition (Condition 3) was poorerthan performance following the talker-specific training condition (p < .02) and themultiple-talker training condition (p < .05). Performance following the multipletalker training condition (Condition 1) and the talker-specific training condition(Condition 2) did not differ significantly from each other. For the post-test withthe Slovakian-accented talker, pair-wise comparisons (1 tailed, t-tests) showed signif-icantly worse performance (at the p < .02 level) in the no training control condition(Condition 5) relative to all of the other conditions. There were no significant differ-ences across any of the other training conditions (Conditions 1–4).
In summary, results of this training study support four major conclusions. First,perception of sentences produced by a foreign-accented talker was better after prac-tice with the task regardless of whether the training talker(s) came from the same lan-guage background as the test talker. This conclusion is supported by the finding thatperformance on the test with the Slovakian-accented talker (post-test 2) was better in
Fig. 3. Performance in RAU on the two post-tests in each of the training conditions in Experiment 2.Error bars represent the standard error of the mean. Training conditions are explained in Table 2 and inthe accompanying text.
720 A.R. Bradlow, T. Bent / Cognition 106 (2008) 707–729
Bradlow & Bent, 2008, Figure 3.!
*!*!
average training benefit
by talker!
0.00
0.25
0.50
0.75
1.00
Single talkertraining
Multi−talkertraining
Prop
ortio
n ke
ywor
ds c
orre
ctly
tran
scrib
ed
Item Repetitionduring Training
yes (Exp2) no (Exp3)
Accuracy at Test
Exp 3.!Replicated finding of cross-talker generalization in both the single-talker and multi-talker conditions, after removing item repetition during training.!
Results!- Cross-talker generalization following both single talker and multi-talker training!- Strength of generalization: no discernable benefit of multi-talker over single
talker training !- Imperfect generalization: multi-talker < talker-specific (cf. B&B, 2008)!
n.s.!
n.s.!**!
Effect of perceived accent familiarity on
test performance!
Participants. 95 Mechanical Turkers!
Control(training = 5 AmEng)
Single talkertraining
Multi−talkertraining
Talker−specifictraining
●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●
●● ●
●
0.6
0.7
0.8
0.9
daily weekly monthly yearly never daily weekly monthly yearly never daily weekly monthly yearly never daily weekly monthly yearly neverFrequency of exposure to the test talker's accent
(subjective report)
Prop
ortio
n of
key
word
sco
rrect
ly tr
ansc
ribed