subphonemic detail is used in spoken word recognition: temporal integration at two time scales
DESCRIPTION
Subphonemic detail is used in spoken word recognition: Temporal Integration at Two Time Scales Bob McMurray. Grateful Thanks to:. Advisors Dick Aslin Mike Tanenhaus. Collaborators Meghan Clayards David Gow. Saviors in the Lab Julie Markant Dana Subik. Committee Joyce McDonough - PowerPoint PPT PresentationTRANSCRIPT
Subphonemic detail is used in spoken word recognition:
Temporal Integration at Two Time Scales
Bob McMurray
Grateful Thanks to:
CommitteeJoyce McDonoughDavid KnillChristopher Brown
CollaboratorsMeghan ClayardsDavid GowSaviors in the LabJulie MarkantDana Subik
AdvisorsDick AslinMike Tanenhaus
People who put up with meKate Pirog Kathy Corser BetteAndrea Lathrop Jennifer Gillis McCormick
Scene Perception: build stable representation across multiple eye-movements, attention shifts.
Music: series of notes. Temporal properties (order and rhythm) are fundamental.
Meaningful stimuli are almost always temporal.
Temporal Integration fundamental to language, as it appears in the world.
Language as Temporal Integration
•Word: Ordered series of articulations.
•Sentence: Sequence of words.
•A Language: Series of utterances.
Phonology, syntax extracted from this series of utterances.
How are abstract representations formed?
Stimuli do not change arbitrarily.
At any point in time, subtle, perceptual cues tell the system something about the change itself.
Enable an active integration process.Anticipating future eventsRetain partial present representations.Resolve prior ambiguity.
Word recognition is an ideal arena:• Substantial perceptual information available.• Multiple timescales for integration.
?But:Early evidence suggested that this
perceptual information is not maintained.
1) Continuous perceptual variation affects word recognition.
Overview
6) Conclusions
5) The use of continuous detail during development.
4) Long-term temporal integration: development.
3) Integrating speech cues in online recognition.
2) A new framework for word recognition.
Speech and Word Recognition
Acoustic
Sublexical Units
/b//la//a/
/l/ /p//ip/
Speech Perception• Categorization of
acoustic input into sublexical units.
LexiconWord Recognition• Identification of target
word from active sublexical units.
bakeryba…
basic
barrierbarricade bait
baby
Xkery
bakery
XXX
X
Word Recognition as temporal ambiguity resolution
•Information arrives sequentially•At early points in time, signal is temporarily
ambiguous.
•Later arriving information disambiguates the word.
Current models of spoken word recognition
• Immediacy: Hypotheses formed from the earliest moments of input.
• Activation Based: Lexical candidates (words) receive activation to the degree they match the input.
• Parallel Processing: Multiple items are active in parallel.
• Competition: Items compete with each other for recognition.
timeInput: b... u… tt… e…
r
beach
bump putter
dog
butter
These processes have been well defined for a phonemic representation of the input.
But there may be considerably less ambiguity in the signal if we consider subphonemic information.
Example: subphonemic effects of motor processes.
Coarticulation
Sensitivity to these perceptual details might yield earlier disambiguation.
Example: CoarticulationMovements of articulators (lips, tongue…)
during speech reflect current, future and past events.
Yields subtle subphonemic variation in speech that reflects temporal organization.
n ne et c
k
Any action reflects future actions as it unfolds.
These processes have largely been ignored because of a history of evidence that perceptual variability gets discarded.
Example: Categorical Perception
Categorical Perception
B
P
Subphonemic variation in VOT is discarded in favor of a discrete symbol (phoneme).
•Sharp identification of tokens on a continuum.
VOT0
100
PB
% /p
/
ID (%/pa/) 0
100Discrimination
Discrimination
•Discrimination poor within a phonetic category.
Evidence against the strong form of Categorical Perception comes from a variety of psychophysical-type tasks:
Discrimination Tasks Pisoni and Tash (1974) Pisoni & Lazarus (1974)Carney, Widin & Viemeister (1977)
Training Samuel (1977)Pisoni, Aslin, Perey & Hennessy
(1982)Goodness Ratings Miller (1997)Massaro & Cohen (1983)
?Does within-category acoustic
detail systematically affect higher level language?
Is there a gradient effect of subphonemic detail on lexical
activation?
A gradient relationship would yield systematic effects of subphonemic information on lexical activation.
If this gradiency is useful for temporal integration, it must be preserved over time.
Need a design sensitive to both acoustic detail and detailed temporal dynamics of lexical activation.
McMurray, Aslin & Tanenhaus (2002)
Use a speech continuum—more steps yields a better picture acoustic mapping.
KlattWorks: generate synthetic continua from natural speech.
Acoustic Detail
9-step VOT continua (0-40 ms)
6 pairs of words.beach/peach bale/pale bear/pearbump/pump bomb/palm butter/putter
6 fillers.lamp leg lock ladder lip leafshark shell shoe ship sheep shirt
How do we tap on-line recognition?With an on-line task: Eye-movementsSubjects hear spoken language and
manipulate objects in a visual world.
Visual world includes set of objects with interesting linguistic properties.
a beach, a peach and some unrelated items.
Eye-movements to each object are monitored throughout the task.
Temporal Dynamics
Tanenhaus, Spivey-Knowlton, Eberhart & Sedivy, 1995
•Relatively natural task.•Eye-movements generated very fast (within
200ms of first bit of information).•Eye movements time-locked to speech.•Subjects aren’t aware of eye-movements.•Fixation probability maps onto lexical
activation..
Why use eye-movements and visual world paradigm?
A moment to view the items
Task
Task
Bear
Repeat 1080 times
By subject: 17.25 +/- 1.33ms By item: 17.24 +/- 1.24ms
High agreement across subjects and items for category boundary.
0 5 10 15 20 25 30 35 4000.10.20.30.40.50.60.70.80.9
1
VOT (ms)
prop
orti
on /p
/
B P
Identification Results
Task
Target = Bear
Competitor = Pear
Unrelated = Lamp, Ship
Time
200 ms
1
2
3
4
5
Trials
Task
00.10.20.30.40.50.60.70.80.9
0 400 800 1200 1600 0 400 800 1200 1600 2000
Time (ms)
More looks to competitor than unrelated items.
VOT=0 Response= VOT=40 Response=Fi
xati
on p
ropo
rtio
n
Task
Given that • the subject heard bear• clicked on “bear”…
How often was the subject looking at the “pear”?
Categorical Results Gradient Effect
target
competitortime
Fixa
tion
prop
ortio
n target
competitor competitorcompetitortime
Fixa
tion
prop
ortio
n target
Results
0 400 800 1200 16000
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0 ms5 ms10 ms15 ms
VOT
0 400 800 1200 1600 2000
20 ms25 ms30 ms35 ms40 ms
VOT
Com
petit
or F
ixat
ions
Time since word onset (ms)
Response= Response=
Long-lasting gradient effect: seen throughout the timecourse of processing.
0 5 10 15 20 25 30 35 400.02
0.03
0.04
0.05
0.06
0.07
0.08
VOT (ms)
CategoryBoundary
Response= Response=
Looks to
Looks to Co
mpe
tito
r Fi
xati
ons
B: p=.017* P: p<.001***Clear effects of VOTLinear TrendB: p=.023* P: p=.002***
Area under the curve:
0 5 10 15 20 25 30 35 400.02
0.03
0.04
0.05
0.06
0.07
0.08
VOT (ms)
Response= Response=
Looks to
Looks to
B: p=.014* P: p=.001***Clear effects of VOTLinear TrendB: p=.009** P: p=.007**
Unambiguous Stimuli Only
CategoryBoundary
Com
peti
tor
Fixa
tion
s
Summary
Subphonemic acoustic differences in VOT have gradient effect on lexical activation.
• Gradient effect of VOT on looks to the competitor.
• Seems to be long-lasting.• Effect holds even for unambiguous stimuli.
Consistent with growing body of work using priming (Andruski, Blumstein & Burton, 1994; Utman, Blumstein & Burton, 2000; Gow, 2001, 2002).
1) Word recognition is systematically sensitive to subphonemic acoustic detail.
The Proposed Framework
2) Acoustic detail is represented as gradations in activation across the lexicon.
3) This sensitivity enables the system to take advantage of subphonemic regularities for temporal integration.
4) This has fundamental consequences for development: learning phonological organization.
Sensitivity & Use
Lexical Sensitivity
1) Word recognition is systematically sensitive to subphonemic acoustic detail.
McMurray, Tanenhaus and Aslin (2002)
Other phonetic contrasts (exp. 1) Non minimal-pairs (exp. 2) During development (exps. 3 & 4)
Lexical Basis
2) Acoustic detail is represented as gradations in activation across the lexicon.
Lexicon forms a high dimensional basis vector for acoustic/phonetic space.
No unneeded dimensions (features) coded—represents only possible alternatives.
2) Acoustic detail is represented as gradations in activation across the lexicon.
timeInput: b... u… m…
p…
bun bumper
pump
dum
p
bump
bomb
3) This sensitivity enables the system to take advantage of subphonemic regularities for temporal integration.
Short term cue integration (exp 1):•Cues to phonetic distinctions are
spread out over time.•Lexical activation retains probabilistic
representation of input as information accumulates.
Longer term ambiguity resolution (exp 2):•Early, ambiguous material retained
until more information arrives.
Temporal Integration
4) Consequences for development: learning phonological organization.
Learning a language: •Integrating input across many utterances
to build long-term representation.
Sensitivity to subphonemic detail (exp 3 & 4).•Allows statistical learning of categories
(exp 5).
Development
Experiment 1
?1) Do lexical representations
serve as a locus for short-term temporal integration of acoustic cues?
2) Can we see sensitivity to subphonemic detail in additional phonetic contexts?
VOT Vowel Length
Phonetic Context
Asynchronous cues to voicing: VOT Vowel Length
Both covary with speaking rate: rate normalization
VOT Vowel LengthVOT Vowel Length
Phonetic Context
Asynchronous cues to voicing: VOT Vowel Length
Both covary with speaking rate: rate normalization
Manner of Articulation Formant Transition Slope (FTSlope): Temporal cue like VOT covaries with vowel length.
belt
welt
VOT precedes Vowel Length.Online processing: how are these cues integrated?
Alternative Models
Vowel Lengthtime
Model 1: Sublexical integration
VOT
The Lexicon
Sublex.Sublexical Rep. (phonemes)
VOT precedes Vowel Length.Online processing: how are these cues integrated?
VOT Vowel Lengthtime
Model 2: Lexical Integration (proposed framework)
The Lexicon
Partial representation retained...
More complete representation…
?Will the temporal pattern of fixations to lexical competitors
reveal when acoustic information contacts the
lexicon?
Eye-movements reveal lexical activation…
9-step VOT continua (0-40 ms) beach/peachbeak/peakbees/peas
9-step formant transition slopebench/wenchbelt/weltbell/well
2 Vowel Lengths x
Fillers•No effect of
vowel length
•Extend gradiency to new continua
9-step F3 onset (place)dune/goondew/goodeuce/goose
9-step F3 onset (laterality)lake/rakelei/railace/race
Task
Same task as McMurray et al (2002)
40 Subjects1080 Trials
Analysis
1) Validate methods with identification (mouse click) data.
2) Extend gradient effects of subphonemic detail to
• Multiple dimensions• New phonetic contrasts
3) Disambiguate integration models by examining when effects are seen.
Results: Stimulus Validation
1) Identification: Expected Results (from literature)
Long Short
B/P More /b/ More /p/
B/W More /b/ More /w/
R/L No difference
D/G No difference
/b/ /b//p/ /w/
B/P
B/W
00.10.20.30.40.50.60.70.80.9
1
0 5 10 15 20 25 30 35 40
VOT
% /p
/ res
pons
e
00.10.20.30.40.50.60.70.80.9
1
1 2 3 4 5 6 7 8 9
FTStep%
/w/ r
espo
nse
LongShort
LongShort
00.10.20.30.40.50.60.70.80.9
1
1 2 3 4 5 6 7 8 9
/l/
% /r
/ res
pons
e
LongShort
00.10.20.30.40.50.60.70.80.9
1
1 2 3 4 5 6 7 8 9
/d/%
/g/ r
espo
nse
/r/
LongShort
/g/
L/R
D/G
Stimulus Validation
Long Short
B/P More /b/ More /p/ B/W More /b/ More /w/ R/L No difference D/G No difference
Results: Gradiency
2) Eye-movements: Predicted Results
Extend gradiency to placeValidate methodsD/G
Replicate prior work 2D gradiencyB/P
Extend gradiency to manner 2D gradiencyB/W
Extend gradiency to lateralityValidate methodsR/L
Vowel FindingContinuum
F3 onsetB: p<.001P: p=.002
Vowel B: p=.006P: p=.061
InteractionB: p>.1P: p=.027
B/P
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
-25 -15 -5 5 15 25
Distance from Category Boundary
Fixa
tions
to C
ompe
titor Long
Short
Summary: Gradiency
Continuum Vowel Finding
B/P P=.0015 .006Replicate prior work 2D gradiency
B/W .001 .05Extend gradiency to FT Slope 2D gradiency
R/L .001 >.1Extend gradiency to F3Validate methods
D/G .017 >.1Extend gradiency to placeValidate methods
Across continua, looks to competitors validated gradient hypothesis.
?Results: Temporal Dynamics
When do effects occur?
VOT / FTStep effects cooccurs with vowel length.(Sublexical Integration)
VOT / FTStep precedes vowel length.(Lexical locus)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-30 -25 -20 -15 -10 -5 0
Distance from Boundary (VOT)
Com
petit
or F
ixat
ions
Y = M720x + B
•VOT / FTStep: Regression slope of competitor fixations as a function of VOT.
Compute 3 effect sizes at each 20 ms time slice.
Time (s)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0 500 1000 1500 2000
Com
petit
or F
ixat
ions
-25-20-15-10-5
VOT from Boundary
Time = 720 ms…
Time = 740 ms…
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-30 -25 -20 -15 -10 -5 0
Distance from Boundary (VOT)
Com
petit
or F
ixat
ions
Y = M740x + B
•VOT / FTStep: Regression slope of competitor fixations as a function of VOT.
Compute 3 effect sizes at each 20 ms time slice.
Time (s)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0 500 1000 1500 2000
Com
petit
or F
ixat
ions
-25-20-15-10-5
VOT from Boundary
Compute 3 effect sizes at each 20 ms time slice.
•Vowel Length: Difference (D) between fixations after hearing long vs. short vowel.
Time = 340 ms…
0.064
0.068
0.072
0.076
0.080
0.084
Long Short
Com
petit
or F
ixat
ions
L-S = D
•Repeat for each time slice, subject.
Compute 3 effect sizes at each 20 ms time slice.•Unrelated: Difference between looks to
target after a experimental vs. filler stimulus.
Information available from the earliest moments of processing: subjects should show early effect.
Does analysis have sufficient power?
Resulting dataset…
Subject Time Unrelated VOT (M) Vowel (D)1 20 0.02076 -0.0023 0.0094
40 0.02446 -0.0016 0.009560 0.02916 -0.0008 0.0108…2000 0.99871 0.06021 0.123
2 20 0.05642 0.0014 0.009140 0.07126 0.0018 0.008860 0.08926 0.0029 0.0104…2000 0.99261 0.0604 0.1223
…
Results: Temporal Dynamics
Model 1: Sublexical integration
Effect of VOT / FTStep appears at same time as Vowel Length
time
VOT Vowel Length
Sublexical Rep. (phonemes)
The Lexicon
time
VOT Vowel Length
Sublexical Rep. (phonemes)
The Lexicon
time
VOT Vowel Length
The Lexicon
Partial representation retained...
More complete representation…
time
VOT Vowel Length
The Lexicon
Partial representation retained...
More complete representation…
Model 2: Lexical Locus
Effect of VOT / FTStep precedes Vowel Length
Looks to competitor Combined (b/p).
B/P: Effects on looks to Competitor
-0.2
0
0.2
0.40.6
0.8
1
1.2
0 300 600 900 1200
Time (ms)
Effec
t Si
ze (
norm
aliz
ed)
VowelVOTUR
Little sequentiality—vowel length and VOT effects appear at same time.
fƒ
Looks to competitor (b/p)
Some sequentiality on voiced side
None on voiceless.
Time (ms)
-0.20
0.20.40.60.8
11.2
0 300 600 900 1200
Effec
t Si
ze (
norm
aliz
ed)
VowelVOTUR
B
-0.20
0.20.40.60.8
11.2
0 300 600 900 1200Time (ms)
Effec
t Si
ze (
norm
aliz
ed)
VowelVOTUR
P
fƒ
B/P Summary
Limited sequentiality of effects supports some kind of sublexical integration.
•Voiced: ~sequential effects.•Voiceless: effect of VOT simultaneous
with vowel length.
VOT requires at least some portion of the vowel for lexical interpretation.
•Voiceless sounds need “more”.•Consistent with prior measurement and
perceptual work.
Looks to competitor Combined (b/w).
Clearly sequential—FTStep effects appear before vowel length.
B/W: Effects on looks to Competitor
-0.4-0.2
00.20.40.60.8
11.2
0 300 600 900 1200Time (ms)
Effec
t Si
ze (
norm
aliz
ed)
VowelStepUR
fƒ
Looks to competitor (b/w)
Clear sequentiality on both sides.
Time (ms)
Effec
t Si
ze (
norm
aliz
ed)
-0.4-0.2
00.20.40.60.8
11.2
0 300 600 900 1200
B
Time (ms)
-0.4-0.2
00.20.40.60.8
11.2
0 300 600 900 1200
Effec
t Si
ze (
norm
aliz
ed)
W
fƒ
StepVowel
UR
B/W Summary
Manner of Articulation•Clear sequential effects on competitor.•Support lexical locus of temporal
integration.Formant transition slope may not work similarly to VOT.
•Is VOT the right cue for voicing?
•What was actually manipulated?FTSlope vs. Transition Duration
Experiment 1 Conclusions
•Additional phonetic dimensionsB/W: Manner of articulation R/L: LateralityD/G: Place of Articulation
•Multi-dimensional categoriesVOT & Vowel LengthFTStep & Vowel Length
Gradient effect on lexical activation extended to
•FTStep effect precedes vowel length.Supports lexical integration.
Temporal Integration:
•VOT effect precedes vowel length only for voiced sounds:
Some vowel required to interpret VOT.
Experiment 2
Lexical activation can play a role in integrating multiple phonemic cues.
?How long is the information available?
How is information at multiple levels integrated?
Competitor still active -- easy to activate it rest of the way.
Competitor completely inactive-- system will “garden-path”.
P ( misperception ) distance from boundary.
Gradient activation allows the system to hedge its bets.
What if a stimulus was misperceived?
Misperception
time
Input: …
parakeetbarricade
Categorical Lexicon
barricade vs. parakeet
parakeet
barricade
Gradient Sensitivity
// vs. /pit/
10 Pairs of b/p items.Voiced Voiceless OverlapBumpercar Pumpernickel 6Barricade Parakeet 5Bassinet Passenger 5Blanket Plankton 5Beachball Peachpit 4Billboard Pillbox 4Drain Pipes Train Tracks 4Dreadlocks Treadmill 4Delaware Telephone 4Delicatessen Television 4
Methods
10 Pairs of b/p items.• 0 – 35 ms VOT continua.
20 Filler items (lemonade, restaurant, saxophone…)
Option to click “X” (Mispronounced).
26 Subjects
1240 Trials over two days.
X
0.000.100.200.300.400.500.600.700.800.901.00
0 5 10 15 20 25 30 35
Barricade
Res
pons
e R
ate
VoicedVoicelessNW
Identification Results
Parricade
0.000.100.200.300.400.500.600.700.800.901.00
0 5 10 15 20 25 30 35
VoicedVoicelessNW
Barakeet Parakeet
Res
pons
e R
ate
Significant target responses even at extreme.
Graded effects of VOT on correct response rate.
05101520253035
0
0.2
0.4
0.6
0.8
1
300 600 900Time (ms)
Fixa
tions
to T
arge
t
VOTBarricade -> Parricade
Faster activation of target as VOTs approach lexical endpoint.
• Even within the non-word range.
fƒ
Eye Movement Results
Parakeet -> Barakeet
300 600 900 1200Time (ms)
“Garden-path” effect:Difference between looks to each
target (b vs. p) at same VOT.
VOT = 0 (/b/)
0
0.2
0.4
0.6
0.8
1
0 500 1000Time (ms)
Fixa
tion
s to
Tar
get
BarricadeParakeet
VOT = 35 (/p/)
0 500 1000 1500Time (ms)
Phonetic “Garden-Path”
-0.1
-0.05
0
0.05
0.1
0.15
0 5 10 15 20 25 30 35
VOT (ms)
Gar
den-
Path
Eff
ect
( Bar
rica
de -
Para
keet
)
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0 5 10 15 20 25 30 35
VOT (ms)
Gar
den-
Path
Eff
ect
( Bar
rica
de -
Para
keet
)
Target
Competitor
GP Effect:Gradient effect of VOT.
Target: p<.0001Competitor: p<.0001
fƒ
Gradient effect of within-category variation without minimal-pairs.
Experiment 2 Conclusions
Gradient effect long-lasting: mean POD = 240 ms.
Regressive ambiguity resolution:
•Subphonemic gradations maintained until more information arrives.
•Subphonemic gradation can improve (or hinder) recovery from garden path.
Lexical activation is exquisitely sensitive to within-category detail.
This sensitivity is useful to integrate material over time.
Adult Summary
Historically, work in speech perception has been linked to development.
Sensitivity to subphonemic detail must revise our view of development.
Development
Use: Infants face an additional problem of temporal integration:
Extracting a phonology from the series of utterances they hear.
Sensitivity to subphonemic detail:
For 30 years, virtually all attempts to address this question have yielded categorical discrimination.
Exception: Miller & Eimas (1996).•Only at extreme VOTs.•Only when habituated to non- prototypical token.
Nonetheless, infants possess abilities that would require within-category sensitivity.
•Infants can use allophonic differences at word boundaries for segmentation (Jusczyk, Hohne & Bauman, 1999; Hohne, & Jusczyk, 1994)
•Infants can learn phonetic categories from distributional statistics (Maye, Werker & Gerken, 2002; Maye & Weiss, 2004).
Use?
Speech production causes clustering along contrastive phonetic dimensions.
E.g. Voicing / Voice Onset TimeB: VOT ~ 0P: VOT ~ 40
Result: Bimodal distribution
Within a category, VOT forms Gaussian distribution.
VOT0ms 40ms
Statistical Category Learning
•Extract categories from the distribution.
+voice -voice
•Record frequencies of tokens at each value along a stimulus dimension.
VOT
frequ
ency
0ms 50ms
To statistically learn speech categories, infants must:
•This requires ability to track specific VOTs.
Why no demonstrations of sensitivity?
• HabituationDiscrimination not ID.Possible selective adaptation.Possible attenuation of sensitivity.
• Synthetic speechNot ideal for infants.
• Single exemplar/continuumNot necessarily a category representation
Experiment 3: Reassess issue with improved methods.
Experiment 3
Head-Turn Preference Procedure (Jusczyk & Aslin, 1995)
Infants exposed to a chunk of language:•Words in running speech.•Stream of continuous speech (ala
statistical learning paradigm).•Word list.
After exposure, memory for exposed items (or abstractions) is assessed by comparing listening time to consistent items with inconsistent items.
HTPP
Test trials start with all lights off.
Center Light blinks.
Brings infant’s attention to center.
One of the side-lights blinks.
When infant looks at side-light……he hears a word
Beach…
Beach…
Beach…
…as long as he keeps looking.
7.5 month old infants exposed to either 4 b-, or 4 p-words.
80 repetitions total.
Form a category of the exposed class of words. PeachBeach
PailBailPearBearPalmBomb
Measure listening time on…
VOT closer to boundaryCompetitors
Original words
Pear*Bear*BearPearPearBear
Methods
B* and P* were judged /b/ or /p/ at least 90% consistently by adult listeners.
B*: 97%P*: 96%
Stimuli constructed by cross-splicing naturally produced tokens of each end point.B: M= 3.6 ms VOTP: M= 40.7 ms VOT
B*: M=11.9 ms VOTP*: M=30.2 ms VOT
Novelty/Familiarity preference varies across infants and experiments.
1221P
1636B
FamiliarityNoveltyWithin each group will we see evidence for gradiency?
We’re only interested in the middle stimuli (b*, p*).
Infants were classified as novelty or familiarity preferring by performance on the endpoints.
Novelty or Familiarity?
CategoricalWhat about in between?
After being exposed to bear… beach… bail… bomb…
Infants who show a novelty effect……will look longer for pear than bear.
Gradient
Bear*Bear Pear
Liste
ning
Tim
e
4000
5000
6000
7000
8000
9000
10000
Target Target* Competitor
List
enin
g Ti
me
(ms)
BP
Exposed to:
Novelty infants (B: 36 P: 21)
Target vs. Target*:Competitor vs. Target*:
p<.001p=.017
Results
Familiarity infants (B: 16 P: 12)
Target vs. Target*:Competitor vs. Target*:
P=.003p=.012
4000
5000
6000
7000
8000
9000
10000
Target Target* Competitor
Lis
teni
ng T
ime
(ms) B
P
Exposed to:
NoveltyN=21
P P* B
.024*
.009**
P P* B
.024*
.009**
4000
5000
6000
7000
8000
9000
10000
List
enin
g Ti
me
(ms)
Infants exposed to /p/
P* B4000
5000
6000
7000
8000
9000
.018*
.028*
.018*
P
List
enin
g Ti
me
(ms) .028*
FamiliarityN=12
NoveltyN=36
<.001**>.1
<.001**>.2
4000
5000
6000
7000
8000
9000
10000
B B* P
List
enin
g Ti
me
(ms)
Infants exposed to /b/
FamiliarityN=16
4000
5000
6000
7000
8000
9000
10000
B B* P
List
enin
g Ti
me
(ms)
.06.15
7.5 month old infants show gradient sensitivity to subphonemic detail.
• Clear effect for /p/• Effect attenuated for /b/.
Contrary to all previous work:
Experiment 3 Conclusions
Reduced effect for /b/… But:
Bear Pear
Liste
ning
Tim
e
Bear*
Null Effect?
Bear Pear
Liste
ning
Tim
e
Bear*
Expected Result?
•Bear* PearBear Pear
Liste
ning
Tim
e
Bear*
Actual result.
•Category boundary lies between Bear & Bear*
• Between (3ms and 11 ms).•Will we see evidence for within-category
sensitivity with a different range?
Same design as experiment 3.VOTs shifted away from hypothesized
boundary Train
40.7 ms.Palm Pear Peach Pail
3.6 ms.Bomb* Bear* Beach* Bale*
-9.7 ms.Bomb Bear Beach Bale
Test:
Bomb Bear Beach Bale -9.7 ms.
Experiment 4
Familiarity infants (34 Infants)
4000
5000
6000
7000
8000
9000
B- B P
List
enin
g Ti
me
(ms) =.05*
=.01**
Novelty infants (25 Infants)
=.02*=.002**
4000
5000
6000
7000
8000
9000
B- B P
List
enin
g Ti
me
(ms)
•Within-category sensitivity in /b/ as well as /p/.
•Shifted category boundary in /b/: not consistent with adult boundary (or prior infant work). Why?
Experiment 4 Conclusions
/b/ results consistent with (at least) two mappings.
Cate
gory
Map
ping
Stre
ngth 1) Shifted boundary
•Inconsistent with prior literature.
•Why would infants have this boundary?
VOT
/b/ /p/
2) Sparse Categories/b/
VOT
Adult boundary
/p/
Cate
gory
Map
ping
Stre
ngth
unmappedspace
HTPP is a one-alternative task. Asks: B or not-B not: B or P
Hypothesis:Sparse categories: by-product of efficient learning.
Distributional learning model
1) Model distribution of tokens asa mixture of Gaussian distributions over phonetic dimension (e.g. VOT) .
2) After receiving an input, the Gaussian with the highest posterior probability is the “category”.
VOT
3) Each Gaussian has threeparameters:
/b/
VOT
Adult boundary
/p/
Categ
ory M
appi
ngStr
engt
h
unmappedspace/b/
VOT
Adult boundary
/p/
Categ
ory M
appi
ngStr
engt
h
unmappedspace
Computational Model
Statistical Category Learning
1) Start with a set of randomly selected Gaussians.
2) After each input, adjust each parameter to find best description of the input.
3) Start with more Gaussians than necessar--model doesn’t innately know how many categories.
-> 0 for unneeded categories.
VOT VOT
Overgeneralization • large • costly: lose phonetic distinctions…
Undergeneralization• small • not as costly: maintain distinctiveness.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 10 20 30 40 50 60
Starting
P(Su
cces
s)
2 Category Model
To increase likelihood of successful learning:• err on the side of caution.• start with small
39,900ModelsRun 3 Category Model
Sparseness coefficient: % of space not strongly mapped to any category.
00.050.1
0.150.2
0.250.3
0.350.4
0 2000 4000 6000 8000 10000 12000Training Epochs
Avg
Spa
rsen
ess
Coeffi
cien
t
Starting
VOT
Small
.5-1
Unmapped space
Start with large σ
00.050.1
0.150.2
0.250.3
0.350.4
0 2000 4000 6000 8000 10000 12000Training Epochs
Avg
Spa
rsit
y Co
effici
ent
20-40
Starting
VOT
.5-1
Intermediate starting σ
00.050.1
0.150.2
0.250.3
0.350.4
0 2000 4000 6000 8000 10000 12000Training Epochs
Avg
Spa
rsit
y Co
effici
ent
12-173-11
Starting
VOT
.5-1
20-40
1) Occasionally model leaves sparse regions at the end of learning.
• Competition/Choice framework:Additional competition or selection mechanisms during processing: categorization despite incomplete information.
Limitations
2) Multi-dimensional categories1-D: 3 parameters /
category2-D: 5 “ “3-D: 21 “ “
• Incorporating cue/model-reliability may reduce dimensionality.
•Similar properties in terms of starting and sparseness.
VOT
Categories•Competitive Hebbian Learning (Rumelhart & Zipser, 1986).•Not constrained by a particular equation—can fill space better.
Non-parametric approach?
Small or even medium starting ’s lead to sparse category structure during infancy—much of phonetic space is unmapped.
To avoid overgeneralization……better to start with small estimates
for
Sparse categories:Similar temporal integration to exp 2
Retain ambiguity (and partial representations) until more input is available.
Model Conclusions
Infants show graded sensitivity to subphonemic detail./b/-results: regions of unmapped phonetic space.
Statistical approach provides support for sparseness.
•Given current learning theories, sparseness results from optimal starting parameters.
Empirical test will require a two-alternative task.•AEM: train infants to make eye-movements
in response to stimulus identity.
Infant Summary
Conclusions
Infant and adult word learning are sensitive to subphonemic detail.
Sensitivity is important to adult and developing word recognition systems.
1) Short term cue integration.2) Long term phonology learning.
In both cases, partially ambiguous material is retained until more data arrives.
Change is the law of life. And those who look only to the past or present are certain to miss the future.
-- John F. Kennedy
The Future?
The Future?
Change is the law of life. And those [Word Recognition Systems] who look only to the past or present are certain to miss the future [Acoustic Material].
-- John F. Kennedy-[McMurray]
Subphonemic cues signal upcoming events.
Can the system use the information to prepare itself for future material?
Spoken language is defined by change.
But the information to cope with it is in the signal.
Within-category acoustic variation is signal, not noise.
The Last Word
Subphonemic detail is used in spoken word recognition:
Temporal Integration at Two Time Scales
Bob McMurray
• Infants make anticipatory eye-movements along predicted trajectory, in response to stimulus identity.
• Two alternatives allows us to distinguish between category boundary and unmapped space.