A Lecture about

1
A Lecture aboutPhonetic Acquisition
Veronica Weiner May, 2006
2
What is phonetic acquisition?

• Learning what sounds are part of a language
• Learning to categorize sounds across different
  speakers and contexts
• Learning what groups of sounds constitute words
• Learning what prosodic cues (pitch, intonation,
  etc.) are part of a language
• There are developmental stages during the first
  year of life
• Phonetic acquisition continues until adulthood
• Available tools to language learners
• Statistical regularities in the auditory input
• Motherese, social interaction
• Multimodal statistics (lip reading)
• Learning by doing (babbling)
• Specialized neural apparatus?

3
Some research questions

• How different sound features (prosody, transition
  probabilities, multimodal correlations) interact
  at different stages of development
• The relationship between production and
  perception
• Brain activity related to phonetic acquisition
• Differences between L1 and L2
• and more

4
What I will cover (Agenda!)

• Part I Phacts about Phonemes
• Some properties of auditory input that I think
  are interesting / useful / good to know
• Part II Some key results in the field
• Infant timeline
• 5 important papers
• Part III Other interesting papers
• The role of social interaction
• Babbling in sign language
• Phonetic acquisition in blind children

5
Part I Phacts about Phonemesa.k.a.Phun
with Phonemes
6
The components of language

• Phonemes
• Phonetic units ?
• are combined into phonemic categories or just
  phonemes
• Example - /r/ and /l/ are phonetic units that
  compose
• different phonemes in English
• the same phoneme in Japanese.
• Each language uses about 40 phonemes

7
Properties of phonemes across languages

• General auditory perceptual abilities provided
  the basic cuts that influence the choice of
  sounds for the phonetic repertoire of the worlds
  languages
• (Kuhl, NRN, 2003)
• Infants are more sensitive to phonemic
  boundaries than other sounds at an equivalent
  distance, even in languages they have not heard.
• (Eimas, 1988)

Image source http//www.eskimo.com
8
Phonemes, phonetic units, who cares?

• To distinguish all the words in a language, a
  person must be able to distinguish all the
  phonemes of that language
• Why? Phonemes are defined as the minimum element
  of contrast between words
• rake vs. lake
• Coding the phonemes of speech data is a difficult
  problem for computers

9
Telling phonemes apart
Image source Wikipedia entry for Formant
10
Telling phonemes apartAre auditory properties
ambiguous?
Figure source Kuhl P. Nat Rev Neuro,
2004.Original figure Peterson Barney, 1952
11
Telling phonemes apartAuditory properties maybe
less ambiguous than we thought (at least for
vowels).
Hillenbrands group added sound duration and
formant contours.
Higher dimensional space of auditory features.
Figure source Hillenbrand et al 1995
12
Telling phonemes apartDistinguishing most
phonemes across speakers and contexts is still a
very difficult problem, though.
Male (Adler) Female,
high voice, a bit slower (Mary Pat)
Source Tony Ezzat, CBCL
13
Open research question

• What are the relevant features, then, for
  acquiring phonemes (or distinguishing them with a
  computer)?
• Auditory
• Articulatory (production-related)

14
Part IISome Key Results in Phonetic
Acquisitionsince 1983
15
1. Acquisition over the 1st year an overview
Figure source P Kuhl. Nat Rev Neurosci. 04.
16
2. Acquisition in older children

• Hazan and Barrett (2000) showed that phonetic
  acquisition occurs even after the age of 12.
• They used a minimal pair procedure

17
2. Acquisition in older childrenHazan and
Barrett (2000)

• A sound is synthesized that morphs in equal
  steps between Sue and Shoe.
• The cues that distinguish the sounds can be
  varied independently or together.
• Subjects are tested on their choice of Sue or
  Shoe at each step.

Sue
Bottom figure P Kuhl. NRN. 04.
18
2. Acquisition in older childrenHazan and
Barrett (2000)
19
3. Babies can distinguish fewer sounds as they
ageFrom Werker and Tees (1983)

• Thompson (Salish) is a Native Indian language
  spoken in British Columbia.
• Thompson phonemes are not distinguishable by
  English speaking adults.
• They are distinguishable by babies.

20
3. Babies can distinguish fewer sounds as they
age From Werker and Tees (1983)

• Experimental procedure Conditioned Head Turn
• Babies are conditioned to look at the speaker
  when there is a sound change by getting visual
  reinforcement for correct head turns
• Procedure works only in around 25 of babies

Photo source Werker lab website
21
3. Babies can distinguish fewer sounds as they
age From Werker and Tees (1983)

• Experiment 2
• Tested more ages
• Tried Salish and Hindi
• Cross sectional data
• Longitudinal data

Cross sectional data 10-12 subjects, MF approx
split, for each bar in figure. Longitudinal
data 3 M, 3 F
22
4. Babies sensitive to frequency distributions in
sounds From Maye, Werker, and Gerken (2002)

• Synthetic sounds ranging from da to ta.
• Training set 1 There are more sounds on the ends
  near da and ta (bimodal distribution).
• Training set 2 Most sounds are in the middle
  (unimodal distribution).

23
4. Babies sensitive to frequency distributions in
sounds From Maye, Werker, and Gerken (2002)

• Experimental procedure Looking time experiment
  comparing alternating and non-alternating stimuli
• Babies are 6-8 months old.
• Results
• Babies look longer at ta and da pairs after
  they have been trained on the bimodal
  distribution.

24
4. Babies sensitive to frequency distributions in
sounds From Maye, Werker, and Gerken (2002)

• Interesting corollary
• After hearing unimodal stimuli, babies actually
  discriminate ta and da worse than most
  infants their age
• Does this mimic what happens when infants hear a
  unimodal distribution in their ambient language?
• Only 6 blocks of 16 training sounds were used here

25
5. Using statistical information to learn word
boundaries From Saffran, Aslin, and Newport
(1996)

• One key problem in word segmentation
• Strategies
• - Statistical pretty baby transitional
  probabilities
• - Prosodic (word stress)

Image source P Kuhl. NRN. 04.
26
5. Using statistical information to learn word
boundaries From Saffran, Aslin, and Newport
(1996)

• Can 8 month old babies extract transitional
  probabilities from synthesized speech that
  contains no breaks, pauses, stress differences,
  or intonations?
• Experimental procedure a Looking Time experiment
• Babies hear 2 minutes of speech
• Looking at a light elicits a word, part word,
  or non word on repeat
• Looking time is measured

27
5. Using statistical information to learn word
boundaries From Saffran, Aslin, and Newport
(1996)
Training pabikugolatupabikudaropi Testing word
part-word pabiku kudaro
Figure source P Kuhl. NRN. 04.
28
6. Prosody vs. statistics to segregate words
From Johnson and Jusczyk (2001)

• Replicated Saffran et al.s results
• Added intonation and stress cues
• 90 of English words stress the first syllable
• When statistics says, isnt a word, but prosody
  says is a word, who wins??
• Experiment 1 Replication of Saffran et. al
• Experiment 2 Part-words had a stressed first
  syllable
• Experiment 3 Part-words were coarticulated
  (spoken together), while statistical words were
  not.
• Experiment 4 Control statistical words were
  coarticulated, part words were not.

29
6. Prosody vs. statistics to segregate words
From Johnson and Jusczyk (2001)
Experiment 1 Replication of Saffran et.
al Experiment 2 Part-words had a stressed first
syllable Experiment 3 Part-words were
coarticulated (spoken together), while
statistical words were not. Experiment 4
Control statistical words were coarticulated,
part words were not.
In 8 month olds, prosody wins.
30
6. Prosody vs. statistics to segregate words
From Johnson and Jusczyk (2001)

• To what extent do these results generalize to
• Different developmental stages?.
• Different input conditions?

31
Part III Odds and Ends(Other interesting
results)
32
Social effects on phonetic acquisition
Kuhl et al. (03) show that English speaking 9
month old babies can learn Mandarin phonetic
contrasts from a live person but not a video of
that person. Goldstein et al. (03) show that 8
month old babies receiving positive feedback from
their mothers vocalized more than yoked controls.
Figure source P Kuhl. NRN. 04.
33
Phonetic acquisition in the blind
Evidence of multimodal phonetic
acquisition. Rowland (83) shows that
blind children make fewer vocalizations but begin
babbling at the same time as sighted children.
Slide source N Pitchford lecturenotes,
Nottingham University
34
Babbling in sign language learners
Pettito et al. (91) show that deaf babies exposed
to sign language make ten times more babbling
hand gestures than controls.
35
Summary
36
What we have learned

1. Languages are distinguished in part by their sets
   of phonemes
2. Discriminating phonemes is a challenging
   computational problem
3. Phonetic learning and performance proceeds in
   stages, beginning in early infancy
4. Phonetic learning continues into adulthood
5. Babies can distinguish fewer sounds as they
   ageFrom Werker and Tees (1983)
6. Babies are sensitive to frequency distributions
   in sounds From Maye, Werker, and Gerken (2002)
7. Statistical information can be used to learn word
   boundaries From Saffran, Aslin, and Newport
   (1996)
8. Prosody can be compared with statistics to
   segregate words From Johnson and Jusczyk (2001)
9. Phonetic learning is affected by social context
10. Some facts about phonetic acquisition in blind
    and deaf learners

37
Thanks!