This material is based upon work supported by NIH under Grant No. F32MH68204-01A1.

1
The effect of experience on the perception and
representation of dialects
Meghan Sumner Stony Brook University University
of California, Berkeley
This material is based upon work supported by NIH
under Grant No. F32MH68204-01A1.
2
The basic questions

1. How do listeners accommodate the variation found
   in the speech signal?
2. What representations are used to perceive and
   understand speech?
3. What is the effect of experience on perception
   and representation?

3
The basic questions

• How do listeners accommodate the variation found
  in the speech signal?
• Is a surface pattern generalization made across
  dialectal variants?
• Are all variants equally able to map to meaning?
• What representations are used to perceive and
  understand speech?
• What is the effect of experience on perception
  and representation?

4
The basic questions

1. How do listeners accommodate the variation found
   in the speech signal?
2. What representations are used to perceive and
   understand speech?
3. What is the effect of experience on perception
   and representation?

5
The basic questions

• How do listeners accommodate the variation found
  in the speech signal?
• What representations are used to perceive and
  understand speech?
• In the long-term, are variants mapped to a single
  abstract representation, or to individual,
  exemplar representations?
• What is the effect of experience on perception
  and representation?

6
The basic questions

1. How do listeners accommodate the variation found
   in the speech signal?
2. What representations are used to perceive and
   understand speech?
3. What is the effect of experience on perception
   and representation?

7
The basic questions

• How do listeners accommodate the variation found
  in the speech signal?
• What representations are used to perceive and
  understand speech?
• What is the effect of experience on perception
  and representation?
• Does experience with a particular dialect affect
  perception and representation?

8
How can dialects help?

• Consider slender
• Variation
• Compare cross-dialectal variants in immediate and
  long-term tasks
• slend-er vs. slend-?
• Experience
• Control participant experience with a particular
  dialect
• Compare r-full forms of General American (GA)
  with r-less forms of New York City dialect (NYC)
  using psycholinguistic tasks
• GA participants who do not r-drop or exhibit
  other regionally-marked characteristics

9
What can priming tasks tell us?

• Form priming Do listeners generalize a pattern
  for cross-dialectal variants independent of
  dialect?
• How well does slend-er prime an identical item
  slend-er or a cross-dialect item slend-??
• Semantic priming Are cross-dialectal variants
  mapped onto by meaning by listeners independent
  of dialect?
• Is slend-? as effective in activating thin as
  slend-er?
• Long-term repetition priming Are
  cross-dialectal variants mapped onto a single
  abstract representation, or stored as individual
  concrete, or exemplar, representations?
• Does slend-er exhibit the same level of identity
  priming after 20 minutes as slend-??

Immediate processing
Long-term representation
10
The effect of experience Participants

• Post-experiment interview questionnaire
• Interview Productive r-dropping, yes or no?
• Questionnaire
• Born in what city, state, country (participants,
  parents, grandparents)
• Languages spoken (participants, parents,
  grandparents)
• Other locations lived (including dates)
• 3 listener populations
• Needed 144 participants, ran 207
• Some exclusion criteria
• Lived outside NYC area, bilingual, not 1st
  semester SB students

11
Three populations

• GA
• Never lived in the NYC/Tri-State area, or any
  other r-dropping region
• First-semester Stony Brook students (run in 1st
  month on campus)
• NYC R-less
• Lifelong NYC-area residents
• Prominent r-dropping
• NYC R-full
• Lifelong NYC-area residents
• No r-dropping

Whats the difference? 96 of R-less listeners
parents and grandparents are NYC natives
3 of R-full listeners parents and/or
grandparents are NYC natives
12
Familiar voices facilitate processing

• Stimuli created with FOUR speakers
• Primes produced by two females (one GA, one NYC)
• Targets produced by two males (one GA, one NYC)

13
Experiment 1 Form priming

• Do listeners generalize a pattern for cross
  dialectal variants independent of dialect?

Typical trial Auditory prime-target pairs
separated by a short interval Lexical decision
made for targets
Prime ISI Target Response (GA Voice
1) (GA Voice 2) runn-er runn-er
Word 100ms g?gi g?gi Pseudoword

14
Experiment 1 Form priming design

• Four conditions
• Condition Related Prime Target Control Prime
• GA GA runn-er runn-er bak-er
• NYC GA run? runn-er bak-?
• GA NYC runn-er run? bak-er
• NYC NYC run? run? bak-?
• 48 participants (16 for each population)
• 160 er final words
• No item or variant presented in more than 1 trial
• Counterbalanced lists
• Fillers avoid response bias

1. 2. 3. 4.
GA Identity NYC Identity
15
Experiment 1 Form priming predictions

• If pattern between variants is internalized
• Priming in all conditions, strongest in identity
  conditions

• If pattern between variants is internalized
• Priming in all conditions, strongest in identity
  conditions
• If pattern not internalized (arbitrary or new)
• Benefit for within-dialect variants

• If pattern between variants is internalized
• Priming in all conditions, strongest in identity
  conditions
• If pattern not internalized (arbitrary or new)
• Benefit for within-dialect variants
• If experience with a variation is critical
• Differences among three listener groups
• e.g., Difference between NYC R-Less and GA in
  facilitation of NYC targets

16
Experiment 1 Form priming results

• NYC R-Less
• Strong identity priming in both dialects
• Reduced variant priming

• NYC R-Full
• Similar to R-Less participants
• Exposure facilitates processing

• GA
• Priming only for GA targets
• Surface variation not internalized

Short-term mapping between all variants for NYC
participants No mapping to NYC variants for
GA Lack of experience has clear processing cost
17
Experiment 2 Semantic Priming

• Do all variants facilitate the recognition of a
  semantically-related target?

Typical trial Auditory prime-target pairs
separated by a short interval Lexical decision
made for targets
Prime ISI Target Response (GA Voice 1) (GA
Voice 2) slend-er thin Word 100ms bo??
? guggy Pseudoword
18
Experiment 2 Semantic priming design

• Four conditions
• Condition Related Prime Target Control
• 1. GA GA slend-er (GA voice 1) thin (GA
  voice 2) filt-er
• 2. NYC GA slend? thin filt?
• 3. GA NYC slend-er thin filt-er
• 4. NYC NYC slend? (NYC voice 1) thin (NYC
  voice 2) filt?
• 48 participants (16 for each population)
• 160 er final words and semantically-related
  targets
• Targets chosen via large-scale mass testing
  experiment
• No item or variant presented in more than 1 trial
• Fillers included to avoid response biases

19
Experiment 2 Semantic priming predictions
If non-dialect variants do not activate
semantically-related items

• Native dialect primes should facilitate
  recognition of related targets
• Non-dialect primes should not yield facilitation

• If lack of exposure prohibits lexical activation
  for variants
• Only GA primes should facilitate lexical
  activation for GA listeners (Cost)
• All variant primes should facilitate lexical
  activation for NYC listeners (Benefit)

20
Experiment 2 Semantic priming results

• NYC R-Less
• All variants equally effective in lexical
  activation
• Benefit for GA prime

• NYC R-Full
• Pattern similar to R-Less participants
• Strong semantic priming independent of variant
  dialect

• GA
• Priming only for GA primes
• NYC primes are unable to activate lexicon

Condition Prime Target GA -GA slend-er
thin NYC -GA slend? thin GA -NYC slend-er
thin NYC -NYC slend? thin
Exposure to two dialects results in equivalent
lexical activation across dialects NYC variant
inconsistent with representation resulting in
processing cost
21
How do listeners process cross-dialect variants?

• Listeners exposed to both dialects
• generalize surface pattern
• form processing
• lexical activation
• in the short-term

22
The issue of representation

• Striking similarity between R-Less and R-Full in
  the short term
• Listeners able to generalize surface pattern with
  experience
• Tasks do not inform us about nature of
  representations
• Issue of representation on backburner
• Surface generalizations derived from abstract
  representations
• Evidence that representations (and information
  they carry) have a role in speech perception
• Question What do listeners actually represent?

23
Experiment 3 Long-term repetition priming

• Do listeners store a single abstract
  representation,
• or multiple, concrete representations?

Typical trial Individual words presented
Lexical decision made for each word Primes and
targets presented in different blocks or
lists Effect based on long-term activation of
phonological representations
24
Experiment 3 Long-term repetition priming design

• Condition Block 1 Primes Block 2
  Targets
• GA GA repeated slend-er
• control filt-er filt-er
• 2. NYC GA repeated slend-er
• control filt-? filt-er
• GA NYC repeated slend-?
• control filt-er filt-?
• 4. NYC NYC repeated slend-?
• control filt-? filt-?

• Condition Block 1 Primes Block 2
  Targets
• GA GA repeated slend-er slend-er
• control filt-er
• 2. NYC GA repeated slend-? slend-er
• control filt-er
• GA NYC repeated slend-er slend-?
• control filt-?
• 4. NYC NYC repeated slend-? slend-?
• control filt-?

• 48 participants, 16 of each listener group
• 160 er final words
• No variant present in more than 1 trial
• Fillers used to eliminate response biases

25
Experiment 3 Long-term repetition priming
predictions

• Expect GA participants to store one
  representation
• Strong within-dialect identity priming
• If listeners store multiple, concrete
  representations
• Strong within- and cross-dialect identity
  priming
• If experience affects representation
• Differences among three listener groups
  (expected)

26
Long-term priming Results

• NYC R-Less
• Strong identity priming in both dialects
• Two variants stored

• NYC R-Full
• Similar to GA single abstract representation
• Representations differ from Overt-NYC

• GA
• Priming only for GA-GA condition
• Single representation

Type of experience plays role Language use ?
representation
27
Language use is not equivalent to representation

• Two critical distinctions can be made
• Production and representation
• Process and representation

• NYC R-Less Participants
• Produced one form in interview but store two
  representations
• Representations not necessarily equivalent

• NYC R-Full and GA Participants
• Same representation, different immediate
  ability to generalize variant pattern

28
To sum up

• Language use is not a predictor of representation
• Evident cost of lack of experience with surface
  variation
• Experience influences
• Immediate processing and representations
• Benefit for canonical form

29
Thank You!
30
Do these patterns generalize to within-dialect
phonetic variation?
31
Long-term repetition priming of phonetic variants
Are multiple phonetic variants stored as
independent forms?
Primes
Targets
flut flu?t? flu? flare ---
Basic Unreleased Glottal Repeated Control New
Control
flut flu?t? flu? flare flare
Basic Identity
Most frequent
Unreleased Identity
Glottal Identity
Repeated Identity
If multiple, concrete representations are stored
we should find identity priming for all
variants
32
Are multiple phonetic variants stored as
independent forms?

• Only Basic t identity as strong as repeated
  control
• Low frequency form encoded most efficiently
• Identity priming indicative of concrete
  representations

Prime Target Basic identity flut flut Unrel
eased identity flu?t? flu?t? Glottal
identity flu? flu? Control identity flare flar
e
Benefit for canonical form, even when it is not
the most frequent
33
Implications and new directions
34
Implications and important questions

• Benefit for canonical form
• Is something about t simply more memorable?
• Are certain acoustic cues more salient?
• Are abstract generalizations made over sounds or
  words?
• Specificity of representations?
• Rethink role and nature of representations
• At what level do multiple representations exist?
• What and where do listeners learn?
• Specific acoustic cues or robust generalizations?
• Is learning pre-lexical or lexical?

Learning and generalization of non-native
acoustic cues
35
Voicing contrast in native and non-native English

• Native English voicing contrast made with
  vowel-consonant duration ratio
• Consider beat and bead
• In Polish, final devoicing leads to near
  neutralization
• Contrast made in voicing into closure duration
• Pilot studies
• Native Polish speaker of English transfer this
  cue to English
• Native English listeners perceive beat and bead
  as beat

Bad Map
36
Learning and generalization of non-native cues

• How do listeners learn to move from reliance on
  a native acoustic cue to the use of a new
  non-native cue?

• Where does learning occur?
• Is a cue learned and remapped at the pre-lexical
  or lexical level?
• What is generalized?
• Once a cue is learned, does it generalize to
• New words? Once beat/bead contrast learned
• Generalize to seat/seed and to tack/tag?
• New speakers? Learn contrast for speakers A B
  
• Generalize to C D?
• New L1 with same cue? Contrast learned for
  Native Polish speakers of English Generalize
  to native Korean?

37
Conclusions

• Listeners
• Represent forms other than those they produce
• Process variants depending on prior experience
• Store multiple concrete representations
• Benefit from a canonical form (even if it is not
  the most frequent)
• A listeners knowledge is not autonomous, but
  bound by experience
• and through experience, multiple representations
  are promoted

38
Thank You!
39
(No Transcript)
40
Example illustration of priming
Long-term Representation
/?????-er/
/???/
Long-term repetition priming
/?????-?/
???
???
Phonemes or Phoneme Chunks
???
???
???
???
???
Feature Bundles
-er features
? features
Prime slend-er
Target slend-er
Speech Signal
Prime slend-?
Target slend-?
41
Comment on error rates
Introduction Preliminaries Dialect
Experiments Generalization to past results New
Directions Conclusion

• Error rates examined for form and long-term
  priming tasks
• Strong benefit for GA form across groups
• independent of native dialect

Form priming
Long-term priming
42
Error Rates show GA benefit
Error rates collapsed across form priming and
long-term priming experiments
Hearing a GA prime improves target recognition
accuracy even for Overt-NYC participants
43
The problem of variation

• Consider the word card
• Boston khad
• New York kh??d
• German-English kha?t

• CARD or
• cod?
• cawed?
• cart?

44

• Dialect atlas
• (Kurath, 1939 Labov et al., 2006)
• Social factors
• (Giles, 1970 Labov, 1972 Preston, 1986)
• Acoustic analysis
• Dialect identification
• (Clopper Pisoni, 2003)
• Dialect variation in OT
• (Anttila Cho, 1998)
• Perceptual learning speaker adaptation
• (Dahan Scarborough, 2005 Kraljic Samuel,
  2005 Norris, McQueen, Cutler, 2003)
• Perception of merged and unmerged dialects

45
Experiment 2 Semantic Priming Schema(Covert-NYC
perspective)
Semantic priming
/??????/
/???/
Lexical
?
???
???
Phoneme
???
???
???
???
???
??? features
??? features
Feature Bundles
Speech Signal
??????
???
??????
46
Coping with variation
White or dark meat? What was produced? d??k Wh
at did I hear? dog d??k da?k variation
not surprising
47
Crossing the threshold

• What changed?
• Do I now treat d??k as a systematic variant of
  /d??k/?
• Did I learn to map a variant (e.g., d??k) onto
  an existing representation, or did I create a new
  exemplar representation?

48
Examining linguistic issues with
psycholinguistic tasks Priming

• What is priming?
• Preparing a sound, chunk, word
• Benefit on the later recognition of a related
  stimulus
• What happens in a typical experiment
• Prime is presented (e.g., horse)
• Prime activates
• Sounds (h, o, etc.)
• Representation (/????/)
• Meaning horse, barn, cow, etc.
• Measure Reaction time to target
• Benefits of priming paradigm
• Sensitivity to subtle acoustic deviations
• Manipulate delay Immediate (processing) and
  long-term (representation) effects

Target presented Form-sharing course Identical
horse Meaning-related barn
49
Example illustration of priming
Long-term Representation
/?????-er/
/???/
???
???
Phonemes or Phoneme Chunks
???
???
???
???
???
Form priming
Feature Bundles
-er features
? features
Prime slend-er
Target slend-er
Speech Signal
Target slend-?
50
Example illustration of priming
Long-term Representation
/?????-er/
/???/
Semantic Priming
???
???
Phonemes or Phoneme Chunks
???
???
???
???
???
Feature Bundles
-er features
? features
Prime slend-er
Target thin
Speech Signal
51
Long-term Representation
/?????-er/
Long-term priming
???
???
Phonemes or Phoneme Chunks
???
???
???
???
???
Feature Bundles
-er features
? features
Prime slend-er
Speech Signal
52
Whats ahead

• Experiment design and results
• Generalization to within-dialect phonetic
  variation
• Discussion of models of representation
• Implications and future directions

53
What is the comparison?

• Comparison
• Reaction time to TARGET preceded by related
  PRIME to
• Reaction time to TARGET preceded by unrelated
  CONTROL PRIME
• Prime Target Reaction Time (Target)
• Control Trial runny runn-er 960
• Critical Trial runn-er runn-er 880
• Difference 80 msec Priming Effect
• Prime Target Reaction Time (Target)
• Control Trial runny runn-er 960
• Critical Trial run? runn-er 920
• Difference 40 msec Priming Effect

54
Experiment 1 Form priming error rates
Hearing a GA prime improves recognition of NYC
target across the board
GA NYC runn-er run?
NYC NYC run? run?
55
What can we learn through priming tasks?

• Form priming Do listeners generalize a pattern
  for cross-dialectal variants independent of
  dialect?
• Semantic priming Are cross-dialectal variants
  mapped onto by meaning by listeners independent
  of dialect?
• Long-term repetition priming Are
  cross-dialectal variants mapped onto a single
  abstract representation, or stored as individual
  concrete, or exemplar, representations?

Immediate processing
Long-term representation