Humanities and social sciences

1. Language development
Phonology:Learning the sounds of language

2. Prenatal language development

3. Howearly is tooearly?
Prenatallanguage
development
 Hearing and speech are intimately connected.
 However, they do not develop simultaneously
 Hearing skill develops much earlier than speaking
 Hearing is a capability shared by all mammals
 There appears to be no very special developmental advantage for
human beings in hearing. Speaking is a different matter

4. Old views
 William James famously said, “the baby, assailed by eyes, ears,
nose, skin and entrails at once, feels it all as one great blooming,
buzzing confusion”.

 He was wrong: Babies start to make sense of their world, and
language, even before they are born.
 How?When? Evidence?
uhm, Sorry James
Uncle

5.  At birth, neonates are experiencing extra-uterine language for
the first time.
 However, in utero they have had the opportunity to learn
about some of the properties of language, at least.

6. Some findings
about the ‘early’
development of
hearing skill in
humans
 Physical and neural development of hearing occurs early in the life
of the human fetus.
 Physical developments:
 The vestibular system starts to develop early in the first months and is
finished after 4 1⁄2 months of uterine life. (Tomatis, 1987, p. 25).
 By just 20 weeks in utero, the auditory apparatus of a fetus is
structurally comparable to that of an adult (Chamberlain, 1987, p. 73).

7.  Neural development follows rapidly:
 At 5 1⁄2 months of prenatal life the cochlear nerve is
myelinized.
 The influence of the cochlear system expands quickly as
it takes in the vestibular system (Tomatis, 1987, p. 25).
 In one of the earliest studies, Eisenberg (1969) "found that
seventh-month prematures responded autonomically and
behaviorally to a number of acoustic variables."
 Thus, he concluded that ‘hearing system is operational before
the end of normal gestation’

8. Some initialresearch
findings
 Birnholz and Benacerraf (1983) used high resolution ultrasound
imaging to record eye-blinks of fetuses in response to "a vibro-
acoustic noise source" of 110-dB output intensity "applied firmly to
the maternal abdomen directly overlying a fetal ear."
 They found that "responses were first elicited between 24
and 25 weeks of gestational age and were present
consistently after 28 weeks" (p. 516).
 They concluded that "hearing is established as a functionally
interactive sensation by the start of the third trimester for the
specific stimulus used and with the restriction to short latency
craniofacial motor reactions“

9.  Mehler and Dupoux (1994) “the hearing of newborns is
excellent. They are whizzes at auditory perception,
recognition, and memorization. From their first moments and,
according to certain studies, even during the last weeks of
pregnancy, their auditory equipment is in perfect order" (p.
51).
 Human response to sound begins in the third trimester of life
and by birth reaches sophisticated levels . . . , especially with
respect to speech," DeCasper and Fifer (1980, p. 1174)

10. A landmark
study in 1986.
 DeCasper and Spencer carried out a study that involved asking
pregnant women to read ‘The cat in the hat’ to their unborn babies
for 2 and half months during their pregnancy.
 Three days after birth, researchers played tape recording of the
story to the infants.
 Two similar, but unfamiliar stories were also played with the target
story.
 The infants showed a distinct preference for the familiar story
 They concluded that the infants were tuned into the acoustic
properties of the story while in the womb.

11. Current
findings from
neonates’
preferences
 Recent evidence has suggested that the human brain is tuned to language
from the earliest stages of development.
 From birth, the human brain responds to speech.
 Only a few days afterbirth, neonates respond differently to language than
to non- linguistic sounds.
 Similar to adults, temporal and frontal areas of the brain are
activated in very young infants in response to spoken language,
 but not to non-linguistic signals such as scrambled speech,
sine-wave contours, tones, monkey calls, and backwards
speech

12.  Very young infants demonstrate a preference for listening to
speech over nonspeech (Vouloumanos andWerker, 2007).
 And are capable of discriminating languages from different
rhythmical classes (Mehler et al.,1988; Nazzi et al.,1998; Ramus et
al.,2000).
 New born infants born to monolingual mothers prefer to listen to
their native language over an unfamiliar language from a different
rhythmical class(Mehler et al.,1988; Moon et al.,1993).

13.  Stress patterns: French infants already understand that their
language is syllable-timed, English infants understand that their
language is stress-timed, and Japanese infants understand that
their language is mora-timed.
 Mehler and Christophe (1995) reported that at the age of 4 days,
infants can distinguish language from other sounds, and have
already begun to prefer not only their mother’s voice but also their
mother's language.
 These studies suggest that infants may have learned about the
properties of the native language while still in the womb.

14. What is the
tearing hurry
to listen?
 "probably subserves a variety of developmental functions such as
language acquisition . . . and mother-infant bonding” (DeCasper and
Fifer).

15. Research
methods for
prenatal stage

16. But this is just
the beginning
 Being tuned in to acoustic properties of language is just the
beginning.
 Speech is a continuous stream of sounds that combine into
syllables, words, phrases etc.
 First hurdle: learning phonemes of a language
 For example, /p/ and /m/ are two phonemes in English
 Now, there are few issues:
 1. lack of acoustic invariance
 2. cross linguistic differences in phonemes

17. Lack of
acoustic
invariance
 This means that acoustic signals vary:
 First, because no two utterances are exactly the same, even by the
same speaker
 Speech sounds produced by men and women have slightly
different properties
 The same sound can ‘sound’ different in different acoustic
environments (allophones).
 How do we identify the phonemes from this mess?
 In most of us this problem is solved by a feature called categorial
perception. (more on this little later)

18. Cross linguistic
differences in
sounds
 Ladefoged (2004; p.179): there are about 6oo different consonants
and 200 different vowels in the world’s languages.
 Each of us is exposed to only a small subset of this.
 Each subset (each language) will be different
 Thus, English speaking children have /r/ vs /l/ contrast in their
language, but he Japanese children will not.

19. So, the moot
question
 How do infants learn the sound system of their language given all
the complexities?
 And they DO learn, since as adults we are perfectly in tune with
the possible and impossible sounds and combination of sounds in
our language, even in a noisy environment.

20. The motor
theory of
speech
perception
(Lieberman)
 A fundamental feature: Categorial perception
 Human auditory system is made in such a way that it detects
easily when one phoneme changes into another
 i.e. differences that cross a phoneme boundary
 Differences within a phoneme boundary is less easy
 Now, this means, when a Hindi speaker hears the sound /k/, in any
acoustic environment, he will hear it as /k/
 Also, he will hear it as different from /g/ as /k/ to /g/ changes
phoneme boundary. Same for English.
 Similarly /k/ vs /kh/ distinction. However, this distinction does not
exist in English.
 This is called categorial perception and it is fundamental to our
ability to use language

21.  This theory posits that humans are born with universal inventory
of feature detectors
 This maps the speech sounds we hear onto the articulatory
gestures (features) we use to produce them (like movement of the
mouth, tongue etc).
 Thus, we possess an innate speech motor system that produces
and detects/understands speech.
 Support of this theory was first provided by Eimas et al 1971.

22. The study
 They studied one and four month old infant’s ability of category
perception.
 Method: high amplitude sucking technique using artificial teat.
Suckling rate of babies are higher when they hear anything novel
or interesting.
 Hypothesis: if the suckling rate differs between sounds, it will
mean that they distinguish between them

23. Stimuli using
VOT
manipulation
 Stimuli: Six artificially created sounds, that differed onVOT (Voice
onsetTime) (only plosives)
 VoT of each sound differed by only 20 ms from its neighbour.
 Thus they created sounds with 80, 60, 40, 20, 0 and -20 msecVoTs
 The difference between only two of these created a phoneme
boundary in real language (40 and 20 ms).

24. The question
 Due to our categorial perception, the adults English speakers
would hear three sounds above 40 ms VoT as /pa/.
 Similarly three sounds withVoT of 25 ms or less would sound /ba/
 Would infants hear the sounds as adults do? In other words, would
they react to phoneme boundaries ( and not within boundary
difference)?
 Results proved them right. none of the infants reacted to
differences within phoneme boundary (0 – 20 ms).

25.  This ability of infants to perceive sounds as per their properties
proved Lieberman’s theory that they are born with this innate
capacity.
 These kids have no experience of producing these speech sounds.
 And exposure to speech sounds (hearing) is also limited at this
point.
 All these mean that categorial perception is innate, not learnt

26. • Several similar studies have found the same.
• It is found that young infants can perceive phonemic differences in
sounds that don’t belong to their own language.
• Trehub (1976) recreated the experiment with Canadian English
infants of the same age group as before and made them listen to
vowel contrasts available in Polish, French: (/pa/ vs /pã/ and one
more contrast found in Czech: (/ řa/ vs /za/).
• None of these contrasts exist in English language, which the
children were exposed to.

27.  Findings like this are taken as proof that humans are born with innate
capacity of distinguishing between all possible sounds of the world’s
languages.
 However, there are issues with that claim.
 Categorial perception feature may not be dedicated to understanding
speech alone, as the theory says, i.e. this capacity is not speech- sound
specific.
 Humans can distinguish between musical tones ( Pisoni 1977)
 This means this is a general property of human auditory system

28. More
surprisingly
 Studies on non0human mammals have showed that they too can
respond categorically to speech stimuli.
 This means this property is not human specific either.
 Study on chinchillas showed this. (Kuhl & Miller 1975).They
switched responses at a change-over point closely resembling the
phonemic boundary of English speakers.
 Later studies by the same group on macaques with same result
(Kuhl & Padden 1982, 1983)

29. Our non-
human
competitors
/k/? /g/?
Nah re…/kh/

30. caveat
 Animals needed thousands of training trials before the actual
experiment.
 This is more than an average infant receives in utero.
 Also, for these non-human animals, the training was targeted,
unlike human infants.
 And,
 There are many phonemes that are perceived differently by
animals compared to humans.
 Nevertheless, the existence of similarities at some points is still
important.

31. The universal
theory
 This theory was proposed as an alternative to motor theory, by
Aslin and Pisoni, 1980.
 According to this theory, infants perceive speech sounds using
‘general auditory mechanism that is categorial in nature and
independent of experience with any native language’.
 Like the motor theory, they too subscribe to the idea that our
ability to perceive differences in speech sounds is innate.
 But the crucial difference they make is the claim that this ability is
derived from a domain general all purpose mechanism that we
share with other animals.

32.  In this theory, infants at birth are able to distinguish between
phonemes of all languages.
 Gradually, throughout the first year of life, they become tuned to
the phonemes of their own language.
 Thus, the phoneme distinctions of the ambient language is
maintained.
 And the features not present in this ambient language are lost.
 So, all children are born with /r/Vs /l/ distinction, but in English
children this is relevant and hence maintained, but for Japanese
this is not relevant, thus lost.

33. Werker &Tees,
1984
 This maintenance/loss of categorial perception ability was
demonstrated by a famous experiment by Werker andTees (1984).
 They tested whether English infants could distinguish between
two contrasts that are not phonemic in English but are so in Hindi
andThomson Salish( indigenous language from central British
Columbia. Indigenous name: Ntlakapamux).
 Hindi sound: /ta/ vs /ta/
 Thomson Salish: /k’i/ vs /q’i/
 English adults cannot distinguish between these.

34.  Subjects: infants aged 8-10 months, 10-12 months.
 Method: conditioned head turn paradigm.
 In this paradigm, infants are trained to turn their head with change of
sound.
 If she turns her head for ‘correct’ changes, there is a reinforcer. A
dancing bear or some other interesting toys appear in that side.
 Rest of the time, sounds play in the background, and the
experimenter keeps child’s attention focused straight in front with
another toy.
 After training the child thus, the actual experiment takes place.

35. Details of
method
 Hits: when baby moves head for changed sounds (reinforcement).
 Miss: when baby does not move head with change of sounds(no
reinforcement)
 False positives: turns head during no-change trial (no
reinforcement).
 Success: if baby correctly detects 8 out of 10 change trials, with no
more than 2 errors (either misses or false positives)

36. result
 8-10 month olds could distinguish between the two Hindi
contrasts and the twoThomson contrasts.
 But, 10-12 month olds could not.
 Meaning, younger infants were sensitive to contrasts not
belonging to their own language.
 But older infants lost that ability.
 Control groups of Hindi andThomson infants however, did not
show the same decline in the relevant contrasts.

37.  This was an important study in child language acquisition domain.
 One, because of the fascinating result.
 Secondly, for the method applied. Her methods are still used in
this domain of research because of the methodological finesse in
using controls.

38. Why this
remains an
important
study
 Some controls applied:
 Control group of infants.This ensured that failure was not due to
lack of attention.
 Sounds were recorded by several native speakers of those languages
 Multiple examples of each sound was used in the experiment, so
that they could check if the infant was capable to ignore speaker
differences (like we do in natural language processing [NLP] these
days).
 They also adapted the procedure to control for baseline infant
biases. Some babies might prefer to turn right, some left.

39.  Werker’s study remain influential till today.
 However, the universal theory has been challenged.
 EvenWerker later favored another theory.

40. Reasons why
universal theory
was later
challenged
 First, it was found in later studies that humans do not lose the
ability to distinguish non-native sounds completely.
 New findings suggesting that infants are not capable of
distinguishing all possible sound contrasts in all possible
languages.

41. Best et al, 1988
 Zulu clicks and English subjects: infants and adults
 Zulu, like many other African languages, have sounds called click
consonants. (listen to Miriam Makeba’s click song to understand
what this means. ‘Gods must be Crazy’ also has some examples).

42. The ability
does not
vanish with
age
 Best’s study reported that all the infant subjects (6-14 months)
could discriminate between different types of clicks.
 There was no loss of the capacity as they grew older.
 All adults could also distinguish between click contrasts.
 Many other subsequent studies also prove that sensitivity to non-
native sounds is not irretrievably lost (Flege,Takagi & Mann,
1995).
 Japanese adults can teach themselves to distinguish /r/ vs /l/ contrast

43. Even more
interesting
 Even when infants and adults are unable to detect difference in
sounds, their brains do register this difference.
 EEG study by Rivera-Gaxiola, Silva-Pereyra &Kuhl (2005) on
infants.
 7 and 11 months oldAmerican infants
 Native and non0native sound contrasts
 ERP data reveal sensitivity to non-native sounds even though they
did not behave as though they identified the same.
 Also, longitudinal data showed an increase in sensitivity to native
sound contrasts

44. The second
problem
 Some research findings suggest that infants are not really born
sensitive to all possible contrasts of all languages.
 Lasky and colleagues (1975) reported Guatemalan Spanish infants
could not distinguish between the Spanish phoneme /ba/ (VOT of -
20 ms) and /pa/ (VOT of +20 ms).
 Even though this contrast is meaningful in Spanish
 This ability develops as the child grows.

45.  Similar conflicting reports are also there in case of fricatives.
 Six to eight month olds could not differentiate between /f/ and
/ϴ/. (Eilers eta l, 1979)
 However, another group found opposite result in the same age
group
 Such contradicting findings suggest that there are possibly some
sounds that are more readily discriminated than others, by infants.
 This means children are not born with sensitivity to all sounds. And
universal theory may not be entirely tenable.

46. Thus,
 It is possible that we are born with sensitivity to some sounds
while others need to be created or fine-tuned by our experiences
of our native languages.
 This is called attunement.
 And this idea is at the roots of many current theories of speech
perception.

47. Attunement
theory
 This theory is based on two premises
 First, infants begin life with the ability of categorial perception.
 That is the ability to partition sounds into categories.
 But that is a ‘basic level cut’ and it roughly partition sounds.
 This is not the same as being able to segment all sounds of all
languages.

48.  The second premise is that experience with language tunes those
basic perceptual abilities.

 Experience means the fact that different sounds lead to different
meaning in words.
 This is not the same as maintenance/loss mentioned earlier.
 This has more to do with ‘perceptual reorganization’.
 This includes tuning, modification and reorganization of the
perceptual abilities by exposure to a particular language.

49.  There are a number of theories in this domain, all supporting the
concept of ‘attunement’
 Best: Perceptual Assimilation Model (1994/5)
 Werker: PRIMIR (2005)
 Kuhl: Native Language Magnet theory (2008)

50. NLM: main
components
 Infants start their life with ability to partition sounds roughly into
categories
 Language experience is necessary to fine tune this ability in order
to perceive the sounds of their own language
 the main focus of their theory: how this tuning process happens
 She proposes that exposure to a language changes the wiring of
the neural pathways.
 Through experience these pathways become committed to a
particular configuration.
 Ultimately, ‘language learning produces dedicated neural
networks that code the patterns of native language speech’

51.  One consequence of this neural commitment is ‘perceptual
magnet effect’
 Brain stores representations of highly frequent sounds as
prototypes and this acts as a ‘perceptual magnet’ for other
members of that sound category.
 As a result, phonemes that are similar to the prototype are made
to sound similar.
 This decreases our ability to tell the difference between members
of the same phoneme category.
 All variants of that phoneme ‘sounds’ the same to us, in spite of
possible differences owing to various factors discussed earlier.

52.  This process makes us less sensitive to the within-category
differences in our own language
 Also, between –category differences of other languages; because
we have not experienced those distinctions.
 It has also been claimed that this perceptual magnet effect may be
human/language specific.
 Unlike the other more general broad perceptual abilities that we
are born with.
 Kuhl (1991) demonstrated this with human infants, adults and
monkeys on a task involving variants of the sound /i/. Monkeys
failed the task.

53.  This theory also provides explanation as to how this tuning
happens.
 Each time we hear phonemes like /r/ and /l/ they are slightly
different from each other.
 So, both Japanese and English infants are likely to hear a range of
different sounds in this /r/ -/l/ spectrum.
 Some more like /l/ and some more like /r/ and some in the middle.

54.  How do they then put a category boundary on this spectrum?
 Kuhl’s solution is :
 there are different distributional properties in the speech signal of
different languages.
 And these properties help different prototypes to develop
 So, Japanese and English infants will hear different distributions of
the spectrum
 Thus, English infants develop 2 prototypes at separate points of the
spectrum, while Japanese develop an intermediate one

55. Maye (2002)
 This study aimed to check if this is true
 6 and 8 month old infants were made to hear 8 artificial syllables
varying along the /ta/-/da/ continuum, by changing the
distribution.
 Half the infants heard these syllables in bimodal distribution:
syllables near the ends of the continuum were presented more
frequently than those in the centre
 Other half of the group heard those iterations that were in the
middle of the continuum more
 In the test phase:
 they either heard alternating (da-ta-da-ta)
 Or non-alternating (da-da-da or ta-ta-ta)


56.  Infants from the bimodal group listened fro longer to the non-
alternating trials than alternating ones.
 This meant they could distinguish between the groups.
 The unimodal group did not show any difference in behavior,
indicating they treated both varieties of sounds as similar.
 Thus, simply by varying the frequency of input of the different
variants of the syllables, they could create different category
boundaries of sounds among the test subjects.
 Thus, Kuhl concluded, ;infants are able to use distributional
information in input speech to detect phonetic category structure’
(Maye et al 2002)

57. Criticisms:
Nittrouer
2001)
 Subjects:6 to 14 month old infants and 2-3 year old children
 Task: distinction between three types of contrast:
 Vowel contrast: /sa/VS /su/
 VOT contrast: /ta/VS /da/
 Fricative contrast:/su/VS /ʃu/
 Result: 65% infants could pass vowel distinction, 35%VOT
contrast.Among children only 56% passed vowel contrast.
 She thus said, ‘the data do not support that infants have clearly
defined phonetic boundary ‘.
 As per the results of the previous studies, she ascribed their
finding to infants excluded from the final data due to fussiness’.

58. Pierrehumbert
( 2003)
• Main point of criticism is that in experimental studies, infants are
made to hear repeated sound contrasts like /ta/ vs /da/.This is not
what the infant faces when she hears normal speech.
• In normal speech these sounds appear in variations, depending on
speaker, acoustic environment, emotional state of speaker etc.
• Hence, success of lab studies do not presuppose real life success

59.  There is no need to propose innate knowledge at all, as per
Pierrehumbert.
 Rather infants could be building up their knowledge bottom-up,
through experience, as Maye showed.
 This is relatively new claim and studies are going on (that infants
do not possess any innate knowledge of categorial perception of
sounds).

60. Now,somebiggerproblems

61. Segmenting
the speech
stream
 “fluent speech contains no known acoustic analogue of the blank
spaces between the printed words” (Brent and Siskind, 2001).
 Now, this is a real problem, something we become aware of when
listening to foreign languages. It’s a continuous flow of sounds we
do not know how to segment.
 This is important, because knowing the difference between /p/
and /b/ does not take us very far.
 So, our new question is: what cues do infants use to segment the
speech stream?

62. 1. Prosodic
cues
 Prosodic cues refer to rhythm, stress pattern and intonation of
speech.
 Languages tend to have a distinct rhythm and stress pattern
usually.
 Hence it was proposed that probably prosodic cues are helpful to
segment words from the speech stream.
 For example: trochaic stress pattern= strong-weak: DOCtor,
HEADache etc in English.
 Iambic: weak-strong: gi-RAFFE, tra-PEZE etc.

63. • Thus, if English infants assume strong-weak combination as a
word, they can accurately predict segmentation.
• This is known as Metrical Segmentation Strategy (Cutler & Norris
1988).
• Most famous evidence for this theory comes from a series of 15
experiments by Juszyk and colleagues (1999).
• Their studies showed that English learning infants are sensitive to
strong-weak stress pattern and can use it to segment words out of
the speech stream.

64.  One of the experiments conducted by Juszyk
 7 months old infants were familiarized with words containing
trochaic stress.These words were part of texts, not in isolation.
 In the test phase, they listened to words in isolation, both familiar
(from the previous phase) and unfamiliar.
 They listened to familiar words for longer than new words

65.  Also, they carried out familiarization with monosyllabic words and
tested them on disyllabic words containing the monosyllabic ones.
Like king, ham and then tested on kingdom, hamlet.
 But in this case, the infants did not show any preference for words
containing familiar monosyllabic words.
 The reason provided was: it is not the sound of the words, but the
stress pattern that children identified.

66.  However, this will not work in case of languages where stress
patterns do not provide reliable cues for word boundary.
 E.g. Inuktitut ( language of Eskimo-Aleut family) words may vary
on stress pattern in different conditions.
 Similarly there are a large number of languages that do not have
the word initial or word final stress pattern that can be reliably
used for word segmentation
 For those languages,Cutler suggest, infants might be using other
strategies.

67. 2. Phonotactic
regularities
 This refers to the system of which sounds/sound-combinations are
allowed in a language.
 E.g. In English, [vzg] is not allowed in any position (it is allowed in
Russian), [br] is used more often in word initial position and [nt]
more often in word final position.
 If infants are sensitive as to which combinations are allowed in
which position, then they can use it to segment words.
 Many studies have found evidence of this too

68. Mattys et al,
(1999)
 They tested whether infants are sensitive to phonotactic regularities by
presenting them with syllable sequences that varied in probability of their
occurrence within words.
 For this, they created two syllable nonwords. Like nongkuth, nongtuth
etc, by playing on the consonant clusters of different types.
 Ngk is more common within words, at syllable boundaries, but ‘ngt’ is
more common inter-words, at word boundaries.
 Hence, clusters with ‘ngk’ is expected to be single word, while clusters
with ‘ngt’ expected to be two words.
 Results showed 9 month olds displaying different behavior pattern with
respect to the two types of clusters, meaning they ‘understand’ the
difference.

69. 3.Allophonic
variations
 This variation refers to the different pronunciation of the same
phoneme in different acoustic environments (i.e. different
positions in the word)
 Like, /p/, /t/ etc sound different in English language in word initial
and word final positions
 This is a very subtle case of correspondence between
manifestation of a sound and its position
 Juczyk and colleagues had found that infants can differentiate
between aspirated and unaspirated sounds.
 They set out to find if infants can use this to locate word
boundaries.

70. Hohne and
Juczyk 1999
 They used conditioned head turn procedure as the paradigm
 Speech sounds play only when the infant turns head towards a
flashing red light above the speaker
 The familiarization phase: infants heard the word ‘nitrates’ or
‘night rates’ (not both) along with many filler words like doctor,
hamlet etc.
 In the test phase: they heard passages that included both the
words.
 E.g.
 businesses try to use night rates to send their packages
 Nitrates are something everyone needs
 My teacher taught me all about nitrates
 Night rates can help us save some money

71. result
 Infants are not expected to understand the sentences, hence
these are used to embed the target words.
 Infants who had been familiarized with ‘nitrates’ listened longer to
the same word in the passage.
 Same for ‘night rates’
 This effect was seen in case of 10 and half month old babies, but
not in nine month olds.
 Researchers concluded that sensitivity to allophonic cues develop
between nine and 10 and half month of age

72. So far
 We saw that infants are sensitive to prosodic cues and allophonic
cues.
 But being sensitive and using them to segment words are not the
same thing
 Also, over reliance on these cues can also lead to mis-
segmentation.
 E.g. aspirated /t/ occurs not only in word initial position , but also
word internally, like ‘atomic’.
 Hence, babies need to also know where the cue is useful for
segmentation and where it is not.
 Whether they possess that knowledge is not known for sure.

73. Also
 Pattern of cues differ across languages; no ‘one size fits all’ here.
 In order to know the relevant cues of a specific language, infants
need to learn a bit of words in that language and to know the
words, they need to learn the word boundaries of that language.
 This leads to a catch 22 situation.
 For example, in order to use trochaic stress pattern to segment
English words, English infants need to know that trochaic pattern
is specific to English.

74.  One possible answer to this problem is that they start by using
different cues which help them learn a few words.
 Once a few words are there in vocabulary, they use those to learn
the language specific cues like stress patterns etc.
 Then those cues are used for learning more words.
 Based on this idea ‘isolated words’ theory was proposed

75. 4. Isolated
words
 Perhaps young infants start by learning words in isolation
 Brent and Siskind (2001) explored this idea by studying what
(English learning) children actually hear
 For this they analysed transcripts of mothers talking to their
children.
 Mothers with their first born infants under 9 month old were
chosen
 This duo was visited at home every 2 weeks while the infant was
between 9 and 15 months of age
 Total 14 visits. Mother’s speech with the child was recorded as she
went about her normal life.

76.  Sessions lasted for 90-120 min. Middle 75 min from 3 or 4 sessions
for each child were used
 Infants age before 12 month’s sessions were called ‘early
transcripts’ and after 14 months were called ‘late transcripts’
 The early transcripts were used only to assess maternal speech to
infants.
 whereas the late transcripts were used only to assess the infants'
productive vocabulary.

77.  The early transcripts were analyzed to determine the frequency
and diversity of isolated words in maternal speech.
 An isolated word is one that has a minimum of 300 ms gap before
and after the utterance.
 They reported 9% of isolated words among all the utterances.
 This means that there are plenty of ‘isolated words’ that the
infants hear at that stage of life.

78.  It was also found that those words are the first words learnt by the
infants.
 Their result also showed that a child’s ability to produce words at
12 months was corelated with the frequency of isolated words
produced by the mother.
 Hence they concluded that it is possible that infants begin to learn
words from the isolated words that they hear (input).

79.  However, there are disagreements on this theory (Brent &
Cartwright, 1996).
 One of the problems pointed out by them is the difference
between isolated words and isolated phrases: ‘mummy’ and
‘mummy hen’.
 Can infants distinguish between words and phrases?
 If not, then they will learn a lot of phrases as words.
 This in itself is not a big problem, since it is well known that infants
treat phrases as words often.

80.  But if they use stress cues to segment, they will end up learning
wrong word stress information.
 This is important because, remember words prosodic cues
learning more words.
 Phrases have different stress pattern compared to words.

81. Transitional
probabilities
 This is another possibility.
 One that thinks infants are clever statisticians, who keep track of
probabilities of sounds’ co-occurrence/
 For example, if one hears the syllable ‘pre’ [pri], what are the
probabilities that the next syllable will be /ti/ ?
 In this case probability is high because in English the word ‘pretty’
is common and high frequency.

82.  But the probability of /ty/ being followed by /bab/ is very low as
there is no word like ‘tybab’ in English.
 Unless it occurs in a phrase like ‘pretty baby’
 So, the transition between /ty/ and /bab/ happens only between
words, not within.
 This is what is meant by transitional probability
 If infants are capable of understanding this, it can be used as a cue
to segment words.

83.  Though this seems improbable, researchers actually found proof
of infants doing as predicted by this theory.
 (Saffran,Aslin and Newport 1996): 8 month old English infants
could identify transitional probabilities between syllables in
continuous stream of nonsense speech sounds.

84. Multiple cues
approach
 Till now we have seen multiple theories as to how infants segment
continuous speech into words
 Each theory has data to back up
 Also there are criticism backed up by another set of data
 Thus, a new approach (Ambridge & Lieven 2011) suggests that
perhaps infants need multiple cues to help them sort the issue at
hand.
 They suggest that when infants hear a sound pattern with an
object they know well, they tend to learn it as a word, with
meaning. E.g. ‘mummy’.
 This way, meaning helps learn initial words, then stress pattern
related cues help them learn more words
 And so on.

85. Speech production

86. Why newborns
cannot
produce
speech
sounds?(even
if they want to)
 Their vocal tract is wrong shape. Newborn baby has a shorter
vocal tract than adults, with larynx placed higher.
 Larynx starts to descend in the first year
 Also, their tongue is larger, with under developed tongue muscles
 So, they are capable of vocalizing but not producing speech
sounds.
 Physiological changes throughout the first year of life finally
makes the child start speaking.
 Also, they go through prelinguistic vocal developments during this
time, the stages of which are largely similar across languages.

87. While they are
at it
 Transition from babbling to words (Barbara Lust, 2006)
 Sounds are omitted: broke = bok
 Sounds are substituted for each other: rabbit =wabbit
 Sounds within a word gets assimilated: doggy =goggy
 Repeat syllable within words: tummy=tum tum; bottle =ba ba
 The question is why do they make these errors when their ability
to perceive difference in sounds is comparatively sophisticated at
a very young age?

88. Possible
answers
 ‘Mushy mouth-mushy ear’ hypothesis (Lust 2006)
 Mispronunciation due to misperception
 This means that infants often pronounce words wrong because
they do not perceive it correctly.
 Now, this is contradictory to everything we saw till now
 Here is the thing: identifying sounds in isolation and identifying
them in words in continuous speech AND connecting it to
meaning is aTASK for children that small.

89.  Bih-dih minimal pair study
 Infants less than 17 months old
 Nonsense words paired with pictures (meaning)
 Bih Dih
 In the familiarisation phase, they hear the sounds and see the
corresponding picture.
 In the test phase, the combination is reversed, to check if they
identify the change.
 Dependent variable: longer time to look at the picture.

90. findings
 At 14 months, infants could distinguish between sounds that were
totally different, like ‘lif’VS ‘neem’ (non minimal pairs)
 But not ‘bih’ vs ‘dih’
 After 17 months of age, they could differentiate the minimal pairs
better.
 This suggests that misperception could be one reason behind
mispronunciation.

91. Fiss
phenomenon
 However, some other examples do not support this
 Child: there is a fis in the pail
 Adult: there is a fis in the pail?
 Child: No, not fis, FIS.
 Again,
 Adult: this is your fis?
 Child : no, my fis. (she continued to reject the adult’s
pronunciation till he said the following)
 Adult: this is your fish
 Child: yes, my fis.
 The mispronunciation does not mislead her to misperceive.

92.  Examples like these show that they perceive the difference, but
are not able to produce sounds correctly.
 So, it could be a matter of articulatory constraints.

93. But then, we
have this
 Puzzle-puddle phenomenon (Smith 1973)
 In this case, the child deforms ‘puddle’ but produces ‘puzzle’ as
‘puddle’
 Puddle puzzle target
 Puggle puddle child
 So, the child can produce ‘puddle’
 What’s the problem , then?

94.  Regression is also reported.(Leopold’s Hildegard)
 Hildegard first pronounced the word ‘pretty’ without deformation.
Later deformed it and subsequently went on to learn it correctly.
 Target: Pretty
 10-16 month: [priti]
 18 months: [bidi]
 18+ month: [priti]

95.  Other theories to rescue:
 Jacobson’s implicational universals
 Template theory etc.
 However, no single theory seems to answer the entire
phenomenon of infant learning to speak in the initial stages.
 More on this in the next segment.