4 Learning Sound Patterns

T
o get a sense of what’s really involved in learning lan-
guage, cast your mind back to what it was like before
you knew any words at all of your native tongue. Well,
wait … since you obviously can’t do that, the best you can do
is to recall any experiences you may have had learning a second
language at an age old enough to remember what the experience
was like (or a third or fourth language, if you were lucky enough
to learn more than one language as a tot). If these memories
involve learning a language in a classroom setting, they turn out
to be a useful point of departure for our purposes, especially
to highlight the striking difference between how you learned
language in the classroom, and how you learned it as a newborn
initiated into your native language.
In a foreign language classroom, it’s usual for the pro-
cess to kick off with a teacher (or textbook) translating a list
of vocabulary items from the new language into your native
language. You then use a small but growing vocabulary to
build up your knowledge of the language. You begin to insert words into prefab-
ricated sentence frames, for example, and eventually you build sentences from
scratch. This is simply not an approach that was available to you as an infant
because then you had no words in any language that could be used as the basis
of translation. Worse, you didn’t even know what words were, or where words
began or ended in the stream of speech you were listening to. You were basically
swimming in a sea of sound, and there wasn’t a whole lot anyone could do in
the way of teaching that would have guided you through it.
If you were to have the unusual experience of learning a second language
by simply showing up in a foreign country and plunging yourself into the
language as best you could, without the benefit of language courses or tourist
phrase books, that would be a bit closer to what you faced as an infant.

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
106  Chapter 4

But still, you have advantages as an adult that you didn’t have as an infant.
You are much more sophisticated in your knowledge of the world, so you’re
not faced with learning how to describe the world using language while you’re
trying to figure out what the world is like. And your intellect allows you to be
much more strategic in how you go about getting language samples from the
speakers of that language; you can, for example, figure out ways to ask speakers
about subtle distinctions—like whether there are different words for the con-
cepts of cat and kitten, or how to interpret the difference in similar expressions,
such as screw up and screw off. Not to mention that your motor skills allow you
to pantomime or point to objects as a way to request native speakers to produce
the correct words for you.
Deprived of many of the possible learning strategies that older people use,
it might make sense that babies would postpone language learning until they
develop in other areas that would help support this difficult task. And, given
the fact that most babies don’t start producing recognizable words until they’re
about a year old, and that they take quite a bit longer than that to string sen-
tences together, it might seem that that’s exactly what does happen. But in fact,
babies begin learning their native language from the day they’re born, or even
earlier; it turns out that French babies tested within 4 days of birth could tell
the difference between French and Russian, and sucked more enthusiastically
on a pacifier when hearing French (Mehler et al., 1988). On the other hand,
infants born into other linguistic households, such as Arabic, German, or Chi-
nese, did not seem to be able to tell the difference between French and Russian
speech, nor did French household babies seem to notice the difference between
English and Italian. This study indicates that babies in utero can begin to learn
something about their native language. Obviously, this can’t be the result of
recognizing actual words and their meanings, since in utero babies have no ex-
perience of the meanings that language communicates. Rather, it suggests that
even through the walls of the womb and immersed in amniotic fluid, babies
learn something about the patterns of sounds in the language they hear.
Humans are unlike honeybees and certain species of songbirds, which
are genetically programmed for a specific type of bee dance or birdsong. The
speech and accent of a child born of French parents but raised from infancy in
the United States will usually be indistinguishable from that of a child born to
U.S. parents and raised in that country. This linguistic flexibility reflects the
fact that humans are powerful learning machines.
In this chapter, we’ll look at what young children need to learn about the
sounds of their language—and the sound system of any language is an intri-
cate, delicately patterned thing. Not only does it have its own unique collec-
tion of sounds, but it has different “rules” for how these sounds can be com-
bined into words. For example, even though English contains all the individual
sounds of a word like ptak, it would never allow them to be strung in this order,
though Czech speakers do so without batting an eye. And no word in Czech
can ever end in a sound like “g,” even though that consonant appears in abun-
dance at the beginnings and in the middles of Czech words. English speakers,
on the other hand, have no inhibitions about uttering a word like dog.
In fact, the sound pattern of a language is a complex code that infants man-
age to crack and mostly master within the first couple of years of life. A mag-
nificent amount of learning happens within the first few months of birth. Long
before they begin to produce words (or even really show that they understand
their meanings), babies can:
1. Differentiate their native language from other languages.
2. Have a sense of how streams of sound are carved up into words.

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 Learning Sound Patterns 107

3. Give special attention to distinctions in sounds that will be especially

useful for signaling different meanings (for example, the distinction
between “p” and “b” sounds; switch the “p” in pat to a “b” sound, and
you get a word with a different meaning).
4. Figure out how sounds can be “legally” combined into words in their
language.
Babies also develop many other nifty skills. Perhaps the only people who
deserve as much admiration as these tiny, pre-verbal human beings are the
scientists who study the whole process. Unlike foreign language teachers who
can test students’ mastery of a language via multiple choice exams and writ-
ing samples, language researchers have to rely on truly ingenious methods for
probing the infant mind. In this chapter, you’ll get a sense of the accomplish-
ments of both groups: the very young who crack the sound code, and the scien-
tists who study their feats.

4.1 Where Are the Words?

A stunningly ineffective way to learn language would be to simply memorize
the meaning of every complete sentence you’ve ever heard, never bother-
ing to break the sentence down into its component parts. In fact, if you took
this route (even if you could actually memorize thousands upon thousands of
complete sentences in the form of long strings of sound), you wouldn’t really
be learning language at all. No matter how many sentences you’d accumu-
lated in your memory stash, you’d constantly find yourself in situations where
you were required—but unable—to produce and understand sentences that
you’d never encountered before. And without analyzing language in terms of
its component, reusable parts, learning a sentence like My mother’s sister has
diabetes would be no help at all in understanding the very similar sentence My
father’s dog has diabetes. You’d treat the relationship in sound between these
as purely coincidental, much as you would the relationship between the sim-
ilar-sounding words battle and cattle; each one would have to be memorized
completely independently of the other. As you saw in Chapter 2, the aspect of
language that lets us combine meaningful units (like words) to produce larger
meaningful units (like phrases or sentences) is one of the universal properties
of human language, and one that gives it enormous expressive power.
Fundamentally, learning language involves figuring out which sounds
clump together to form basic units, and learning how these units in turn can
be combined with other units—which is why foreign language instruction for
beginners puts so much emphasis on learning lists of words. One of the in-
fant’s earliest tasks, then, is to figure out which strings of sounds form these
basic units—no trivial accomplishment. In talking to their babies, parents
are not nearly as accommodating as Spanish or German textbooks, and they
rarely speak to their children in single-word utterances (about 10 percent of
the time). This means that babies are confronted with
speech in which multiple words are sewn seamlessly
together, and they have to figure out all on their own
WEB AC TIVIT Y 4.1
where the edges of words are. And unlike written lan-
guage, where words are clearly separated by spaces (at Finding word boundaries In this
least in most writing systems), spoken language doesn’t activity, you’ll hear speech in several
present convenient breaks in sound to isolate words. To different languages, and you’ll be asked
get an intuitive feel for what this task might feel like to to guess where the word boundaries might be.
a baby, see how you fare in Web Activity 4.1.

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
108 Chapter 4

METHOD 4.1
The head-turn preference paradigm

T he head-turn preference paradigm is an invaluable

tool that’s been used in hundreds of studies of
infant cognition. It can be used with babies as young
of headphones; this prevents the adult from hearing the
experimental stimuli and either purposely or inadvertently
giving cues to the baby. A video camera is set up to face
as about 4.5 months, and up to about 18 months. After the child, recording the baby’s responses, and an unseen
this age, toddlers often become too fidgety to reliably sit observer monitors the experiment by watching the baby
still through this experimental task. Applied to speech on video. (The observer can’t hear the stimulus sounds
perception, the technique is based on two simple and is usually not aware of which experimental condition
principles: the child has been assigned to, though the observer does
1. Babies turn their heads to orient to sounds. control the sequence of events that occur.) A flashing light
mounted next to the video camera and straight ahead in the
2. Babies spend more time orienting to sounds that child’s view can be activated to draw the child’s attention
they find interesting. to a neutral point before any of the experimental stimuli
In a typical experiment, the baby sits on the lap of a are played. The stimuli of interest are then played on two
parent or caregiver who is listening to music over a set speakers, mounted on the left and right walls (see Figure 4.1).
Each experiment usually consists
Observers of a familiarization phase and a test
phase. There are three goals for the
familiarization phase. The first is
Viewing window
simply to have the infant become
familiar with the sound stimuli. In
some cases, if the purpose of the
Green light Video camera study is to find out whether babies
will learn something about new
sounds, the sounds played during the
familiarization phase might consist
of the new stimuli to be learned.
The second goal is to train the baby
Red light
to expect that sounds can come
from the speaker on either the left
or the right wall. The third goal of
the familiarization phase is to tightly
lock together the head-turn behavior
to the infant’s auditory attention.
Babies tend to look in the direction
Speaker of a sound that holds their attention
anyway, but this connection can
be strengthened by flashing a light
continued on next page
Infant
Figure 4.1 A testing booth set up for the
head-turn preference paradigm. The baby sits
Parent with on the caregiver’s lap, facing the central panel.
headphones The observer looks through a small window or
one-way mirror to note the baby’s head turns.
(Adapted from Nelson et al., 1995.)

uncorrected pageQuestions:
proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
The text does not mention the red lights.
 Learning Sound Patterns 109

M E T H O D 4 . 1 (continued)

in the location of the speaker before each sound, and listen to sounds of their own language over an unknown
by making sure that sounds during the familiarization language? Other times, the researchers are only interested
phase are played only for as long as the baby looks in in whether the infants discriminate between two
the direction of the sound. This signals to the child that categories of stimuli, and it doesn’t matter which category
if she wants to continue hearing a sound, she needs to is preferred. For instance, in a learning experiment, any
be looking in its direction. After all these goals have been distinction in looking times for familiar versus new stimuli
achieved, the baby is ready for the test phase. should be an indication of learning, regardless of whether
During the test phase, the sounds of interest are played babies prefer to listen longer to the new or the familiar
on either the left or the right speaker, and the baby’s head- sounds. And in fact, it turns out that there isn’t a clear
turn behavior is recorded by the video camera for later preference for either new or familiar sounds—at times,
coding. Researchers then measure how long the baby babies show more interest in sounds that they recognize,
spends looking in the direction of each sound, and these and at other times, they show more interest in completely
responses are averaged over stimulus type. novel ones. The preferences seem to depend somewhat
Sometimes researchers are interested in which sounds on the age of the child and just how often they’ve heard
the infants prefer to listen to—do they prefer a female the familiar sounds (overly familiar sounds might cause a
voice to a male voice, for example, or do they prefer to baby to become bored with them).

Probing infants’ knowledge of words

Babies begin to produce their first words at about a year or so, but they start
to identify word breaks at a much younger age than that. In fact, the whole
process is under way by 6 or 7 months. Scientists can’t exactly plop a transcript
of speech in front of a baby, hand them a pencil, and ask them to mark down
where the word breaks are. So how is it possible to peer into infants’ minds and
determine whether they are breaking sentences down into their component
parts?
In studying the cognitive processes of infants, researchers have to content
themselves with a fairly narrow range of infant behaviors as a way to mea-
sure hidden psychological mechanisms. So, when they decide to study infants
of a particular age, they need to have a clear sense of what babies can do at
that point in their development—and more specifically, which behaviors reflect
meaningful cognitive activity. It turns out that a great deal of what we now
know about infant cognition rests on one simple observation: when babies are
faced with new sounds or images, they devote their attention to them differ-
ently than when they hear or see old, familiar sounds or images. And at the age
of 6 or 7 months, one easy way to tell if a baby is paying attention to something
is if she swivels her head in its direction to stare at it; the longer she keeps her
gaze oriented in its direction, the longer she’s paying attention to that stimu-
lus. New sounds and sights tend to draw attention differently than familiar
ones, and babies will usually orient to novel versus familiar stimuli for different
lengths of time—sometimes they’re more interested in something that’s famil-
iar to them (a sort of “Hey, I know what that is!” response), but sometimes they
prefer the novelty of the new stimulus.
These simple observations about the habits of babies gave birth to a tech-
nique that psycholinguists now commonly use, called the head-turn prefer- head-turn preference paradigm An
ence paradigm (see Method 4.1). This technique compares how long babies experimental framework in which infants’
keep their heads turned toward different stimuli, taking this as a measure of speech preference or learning is measured
their attention. (If the target stimulus is a sound, it’s usually coupled with a by the length of time they turn their heads
visual stimulus such as a light or a dancing puppet in order to best elicit the in the direction of a sound.

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
110  Chapter 4

familiarization phase A preparation head-turn response.) But what makes the method really powerful is that it can
period during which subjects are exposed leverage the measure of looking time as a way to test whether or not the babies
to stimuli that will serve as the basis for the taking part in the study have learned a particular stimulus. For instance, let’s
test phase to follow. say we give babies a new word to listen to during a familiarization phase. At
test phase The period in which subjects’ some later time, during a test phase, we can see whether their looking times
responses to the critical experimental when hearing this word are different from those for a word they’ve never heard
stimuli is tested following a familiarization before. If babies spend either more or less time looking at the previously pre-
phase. sented word than they do at a completely new word, this suggests that they’ve
learned something about the first word and now treat it as “familiar.” On the
other hand, if they devote equal amounts of looking time to both words, it
suggests that they haven’t learned enough about the previously heard word to
differentiate it from a completely novel word.
The head-turn preference paradigm has been used—for example, by Peter
Jusczyk and Richard Aslin (1995)—to tackle the question of whether babies
have learned where word breaks occur. Here’s how: During the familiariza-
tion phase of that study, the baby participants heard a series of sentences that
contained a target word, say, bike, in various different positions in the sentence:
His bike had big black wheels.
The girl rode her big bike.
Her bike could go very fast.
The bell on the bike was really loud.
The boy had a new red bike.
Your bike always stays in the garage.
During the test phase, the researchers measured how long the infants were in-
terested in listening to repetitions of the target word bike, compared with a word
(say, dog) that they hadn’t heard during the familiarization phase. To first get
the baby’s attention before the test word was played, a flashing light appeared
above the loudspeaker the word was to come from. Once the baby looked in
this direction, the test word repetitions began to play. When the baby’s interest
flagged, causing him to look away from the loudspeaker, this was noted by a
researcher, and the baby was scored for the amount of time spent looking in the
direction of the loudspeaker. Jusczyk and Aslin found that overall, 7.5-month-
old babies spent more time turning to the speaker when it played a familiar
word (bike) than when it played an unfamiliar word (dog). This might not seem
like a tremendous feat to you, but keep in mind that the babies must somehow
have separated the unit bike from the other sounds in the sentences during the
familiarization phase in order to be able to match that string of sounds with the
repeated word during the test phase. Six-month-old babies didn’t seem to have
this ability yet.
This study shows that by the tender age of 7.5 months, babies seem to be
equipped with some ability to separate or segment words from the speech
stream—but it doesn’t tell us how they manage to come by these skills, or what
information they rely on to decide where the words are. Since Jusczyk and
Aslin’s initial study, dozens of published articles have explored the question of
how babies pull this off. We’ll investigate several ways that they might begin
to crack the problem.

Familiar words break apart the speech stream

Here’s one possibility. Remember that babies hear single-word utterances only
about 10% of the time. That’s not a lot, but it might be enough to use as a start-
ing point for eventually breaking full sentences apart into individual words. It

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 Learning Sound Patterns   111

may be that babies can use those few words they do hear in isolation as a way
to build up a small collection of known word units. These familiar words can
then serve as anchoring points for breaking up the speech stream into more
manageable chunks. For example, imagine hearing this procession of sounds
in a fictional language:
bankiritubendudifin
Any guesses about what the word units are? The chances of getting it right are
not very high. But suppose there are two words in this stream that you’ve heard
repeatedly because they happen to be your name (“Kiri”) and the name for
your father (“Dudi”). You may have learned them because these are among the
few words that are likely to be uttered as single words quite often, so they’re es-
pecially easy to recognize. Perceptually, they’ll leap out at you. If you’ve learned
a foreign language, the experience of hearing sentences containing a few famil-
iar words may be similar to the very early stages of learning a new language
as a baby; you might have been able to easily pull out just one or two familiar
words from an otherwise incomprehensible sentence. With this in mind, imag-
ine hearing:
ban-kiri-tuben-dudi-fin
Now when you hear the names kiri and dudi, their familiarity allows you to
pull them out of the speech stream—but it might also provide a way to identify
other strings of sound as word units. It seems pretty likely that ban and fin are
word units too, because they appear at the beginning and end of the utterance
and are the only syllables that are left over after you’ve identified kiri and dudi
as word units. So now, you can pull out four stand-alone units from the speech
stream: kiri, dudi, ban, fin. You don’t know what these last two mean, but once
they’re firmly enough fixed in your memory, they might in turn serve as clues
for identifying other new words. So tamfinatbankirisan can now be pulled apart
into:
tam-fin-at-ban-kiri-san
The residue from this new segmentation yields the probable units tam, at, and
san, which can be applied in other sentences in which these units are combined
with entirely new ones.
In principle, by starting with a very few highly familiar words and generating
“hypotheses” about which adjacent clumps of sound correspond to units, an in-
fant might begin to break down streams of continuous sound into smaller piec-
es. In fact, experimental evidence shows that babies are able to segment words
that appear next to familiar names when they’re as young as 6 months old, sug-
gesting that this strategy might be especially useful in the very earliest stages of
speech segmentation. (Remember that the Jusczyk and Aslin study showed that
6-month-olds did not yet show evidence of generally solid segmentation skills.)
This was demonstrated in an interesting study led by Heather Bortfeld (2005).
Using the head-turn preference paradigm, the researchers showed that even
6-month-olds could learn to segment words such as bike or feet out of sentenc-
es—but only if they appeared right next to their own names (in this example, the
baby subject’s name is Maggie) or the very familiar word Mommy:
Maggie’s bike had big, black wheels.
The girl laughed at Mommy’s feet.
That is, when babies heard these sentences during the familiarization phase,
they later spent more time looking at loudspeakers that emitted the target word
bike or feet than at speakers that played an entirely new word (cup or dog)—this

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
112  Chapter 4

shows they were treating bike and feet as familiar units. But when the sentences
in the familiarization phase had the target words bike and feet right next to
names that were not familiar to the child (for example, The girl laughed at Tom-
my’s feet), the babies showed no greater interest in the target words than they
did in the new words cup or dog. In other words, there’s no evidence that the
infants had managed to pull these words out of the stream of speech when they
sat next to unknown words. At this age, then, it seems that babies can’t yet seg-
ment words out of just any stream of speech, but that they can segment words
that appear next to words that are already very familiar.

Discovering what words sound like

Relying on familiar words to bust their way into a stream of sound is just one
of the tricks that babies have in their word-segmentation bag. Another trick
has to do with developing some intuitions about which sounds or sequences
of sounds are allowed at the beginnings and ends of words, and using these
intuitions as a way to guess where word boundaries are likely to be.
To get a more concrete feel for how this might work, try pronouncing the
following nonsense sentence in as “English-like” a way as you can, and take a
stab at marking the word boundaries with a pencil (hint: there are six words in
the sentence):
Banriptangbowpkesternladfloop.
If you compare your answer with those of your classmates, you might see some
discrepancies, but you’ll also find there are many similarities. For example,
chances are, no one proposed a segmentation like this:
Ba-nri-ptangbow-pkester-nladfl-oop.
This has to do with what we think of as a “possible” word in English, in terms
of the sequence of sounds it’s allowed to contain. Just because your language
includes a particular sound in its inventory doesn’t mean that sound is allowed
to pop up just anywhere.
Languages have patterns that correspond to what are considered “good”
words as opposed to words that look like the linguistic equivalent of a patched-
together Frankenstein creature. For example, suppose you’re a marketing ex-
pert charged with creating a brand new word for a line of clothing, and you’ve
decided to write a computer program to randomly generate words to kick-start
the whole process. Looking at what the computer spat out, you could easily sort
out the following list of words into those that are possible English-sounding
words, and those that are not:
ptangb sastashak
roffo lululeming
spimton ndela
skrs srbridl

What counts as a well-behaved English word has little to do with what’s

actually pronounceable—you might think that it’s impossible to pronounce se-
quences of consonants like pt, nd, and dl, but actually you do it all the time in
words like riptide, bandage, bed linen—it’s just that in English, these consonants
have to straddle word boundaries or even just syllable boundaries. (Remember,
there are no actual pauses between these consonants just by virtue of their be-
longing to different syllables or words). You reject the alien words like ptangb
or ndela in your computer-generated list, not because it takes acrobatic feats of

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 Learning Sound Patterns 113

your mouth to pronounce them, but because you have ingrained word tem-
plates in your mind that you’ve implicitly learned, and these words don’t match
those mental templates. These templates differ from one language to another
and are known as phonotactic constraints.
Using English phonotactic constraints to segment another language, though,
could easily get you into trouble, especially if you try to import them into a lan-
guage that’s more lenient in allowing exotic consonant clusters inside its words.
For instance, skrs and ndela are perfectly well-formed pronunciations of words
in Czech and Swahili, respectively. So, trying to segment speech by relying on
your English word templates would give you non-optimal results. You can see
this if you try to use English templates to segment the following three-word
stream of Swahili words:
nipemkatenzuri
You might be tempted to do either of the following:
nipem-katen-zuri
nip-emkat-enzuri WEB AC TIVIT Y 4.2
but the correct segmentation is: Segmenting speech through pho-
nipe-mkate-nzuri
notactic constraints In this activity,
you’ll hear sound files of nonsense
And while languages like Czech and Swahili are quite words that conform to the phonotactic constraints
permissive when it comes to creating consonant clus- of English, as well as clips from foreign languages
ters that would be banned by the rules of English, other that have different phonotactic constraints. You’ll
languages have even tighter restrictions on clusters than get a sense for how much easier it is to segment
English does. For instance, “sp” is not a legal word- unknown speech when you can use the phonotactic
initial cluster in Spanish (which is why speakers of that templates you’ve already learned for your language,
language often say “espanish”). (You can see some more even when none of the words are familiar.
examples of how languages apply different phonotactic
constraints in Box 4.1.)
It turns out that by 9 months of age, babies have some knowledge of the
templates for proper words in their language. Using the head-turn preference
paradigm, researchers led by Peter Jusczyk (1993) have shown that American
babies orient longer toward strings of sounds that are legal words in English
(for example, cubeb, dudgeon) than they do to sequences that are legal Dutch
words but illegal words in English (zampljes, vlatke). Dutch 9-month-olds show
exactly the opposite pattern. This suggests that they’re aware of what a “good”
word of their language sounds like. And, just as neither you nor your class-
mates suggested that speech should be segmented in a way that allows bizarre
words like ptangb or nladf, at 9 months of age, babies can use their phonotactic
templates to segment units out of speech (Mattys & Jusczyk, 2001).
You might have noticed in Web Activity 4.2 that there was another bit of
information that might have helped you segment words, in addition to clues
about phonotactic constraints. In that exercise, English-like stress patterns
were also present. In English, stress tends to alternate, so within a word, phonotactic constraints Language-
specific constraints that determine how
you usually get an unstressed syllable sitting next to a stressed one: reTURN,
the sounds of a given language may be
BLACKmail, inVIgoRATE. (In some other languages, such as French, syllables
combined to form words or syllables.
are more or less evenly stressed.) English words can follow either a trochaic
stress pattern, in which the first syllable is stressed (as in BLACKmail), or an trochaic stress pattern Syllable em-
iambic stress pattern, in which the first syllable is unstressed (as in reTURN). phasis pattern in which the first syllable is
But as it turns out, it’s not an equal-opportunity distribution, and trochaic stressed, as in BLACKmail.
words far outnumber iambic words (on the order of 9 to 1 by some estimates). iambic stress pattern Syllable empha-
Chances are, you subconsciously made use of this knowledge in your segmen- sis pattern in which the first syllable is
tation answers in Web Activity 4.2. If babies have caught on to this pattern in unstressed, as in reTURN.

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
114 Chapter 4

BOX 4.1
Phonotactic constraints across languages

A s languages go, English is reasonably loose in allowing

a wide range of phonotactic templates. English
allows consonants to gather together in sizable packs at
but others are highly arbitrary. The table gives a few
examples of possible and impossible clusters that can
occur at the beginnings of words and syllables in several
the edges of words and syllables. For example, the single- languages.
syllable word splints has the structure CCCVCCC (where
C = consonant and V = vowel). Many other languages are To get a feel for how speech segmentation might be
far more restrictive. To illustrate, the Hebrew, Hawaiian, affected by language-specific phonotactic constraints, try
and Indonesian languages allow only the following syllable listing all the possible ways to break down the following
structures: stream of sounds into word units that are legal, depending
Hebrew Hawaiian Indonesian on whether your language is English, German, French, or
Italian:
CV V V
bakniskweriavrosbamanuesbivriknat
CVC CV VC
CVCC CV
Allow words and syllables to start with
CVC
Language /kn/ /skw/ /sb/ /vr/
In addition to broadly specifying the consonant-vowel
structure of syllables, languages have more stringent rules English no yes no no
about which consonants can occur where. For example, German yes no no no
in English, /rp/ is allowed at the end of the word, but not French no no no yes
at the beginning; the reverse is true for the cluster /pr/.
Italian no no yes no
Some constraints tend to recur across many languages,

the language, they might also be able to use this to make guesses about how
words are segmented from speech. And indeed, they do: by 7.5 months, babies
have no trouble slicing words with a trochaic pattern (like DOCtor) from the
speech stream—but when they hear an iambic word like guiTAR embedded in
running speech, they don’t recognize it. In fact, if they’ve heard guitar followed
by the word is, they behave as if they’ve segmented TARis as a word (Jusczyk
et al., 1999).
But if you’ve been paying very close attention, you might have noticed a
paradox: in order for babies to be able to use templates of permissible words to
segment speech, they already have to have some notion of a word—or at the
very least, they have to have some notion of a language unit that’s made up of
a stable collection of sounds that go together and can be separated from other
collections of sounds. Otherwise, how can they possibly have learned that “ft”
can occur as a sequence of sounds in the middle or at the end of a word unit, but
not at its beginning? It’s the same thing with stress patterns: How can babies
rely on generalizations about the most likely stress patterns for words in their
language unless they’ve already analyzed a bunch of words?
To get at generalizations like these, babies must already have segmented some
word units, held them in memory, and “noticed” (unconsciously, of course) that
they follow certain patterns. All this from speech that rarely has any words that
stand alone without the added confusion of adjacent speech sounds.
Earlier we speculated that maybe isolated familiar words like the baby’s
name serve as the very first units; these first words then act as a wedge for
segmenting out other words, allowing the baby to build up an early collection

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 Learning Sound Patterns   115

of word units. Once this set gets large enough, the baby can learn some useful
generalizations that can then accelerate the whole process of extracting ad-
ditional new words from running speech. In principle, this is plausible, given
that babies can use familiar words like names to figure out that neighbor-
ing bunches of sounds also form word units. But this puts quite a big burden
on those very first few words. Presumably, the baby has managed to identify
them because they’ve been pronounced as single-word utterances. But as it
happens, parents are quite variable in how many words they produce in isola-
tion—some produce quite a few, but others are more verbose and rarely pro-
vide their children with utterances of single words. If this were the crucial
starting point for breaking into streams of speech, we might expect babies to
show a lot more variability in their ability to segment speech than researchers
typically find, with some lagging much further behind others in their segmen-
tation abilities.
Luckily, youngsters aren’t limited to using familiar, isolated words as a de-
parture point for segmentation—they have other, more flexible and powerful
tricks up their sleeves. Researchers have discovered that babies can segment
streams of sounds from a completely unfamiliar language after as little as two
minutes of exposure, without hearing a single word on its own, and without
the benefit of any information about phonotactic constraints or stress patterns.
How they manage this accomplishment is the topic of the next section.

4.2 Infant Statisticians

Tracking transitional probabilities:
The information is out there
In a seminal study, Jenny Saffran and colleagues (1996) familiarized 8-month-
old infants with unbroken 2-minute strings of flatly intoned, computer-gener-
ated speech. The stream of speech contained snippets such as:
bidakupadotigolabubidaku
Notice that the sounds are sequenced so that they follow a repeating consonant-
vowel structure. Because English allows any of the consonants in this string to
appear either at the beginnings or the ends of syllables and words, nothing
about the phonotactic constraints of English offers any clues at all about how
the words are segmented, other than that the consonants need to be grouped
with at least one vowel (since English has no words that consist of single con-
sonants). For example, the word bidakupa could easily have any of the following
segmentations, plus a few more:
bi-dak-upa bid-aku-pa bid-ak-u-pa
bi-da-kup-a bid-ak-up-a bidaku-pa
bid-akupa bida-kupa bi-dakup-a
bi-daku-pa bid-akup-a bidak-upa

If this seems like a lot, keep in mind that these are the segmentation possibili-
ties of a speech snippet that involves just four syllables; imagine the challenges
involved in segmenting a two-minute-long continuous stream of speech. This
is precisely the task that Saffran and colleagues inflicted on the babies they
studied.
In the Saffran et al. study, though, the stream of sound that the babies lis-
tened to during their familiarization phase was more than just a concatenation
of consonant-vowel sequences. The stimuli were created in a way that repre-

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 116 Chapter 4

Experimenters create an artificial “language” of four “words”:

Speaker bidaku, golabu, padoti, dutaba

Familiarization phase
Infant hears each “word” repeated 45 times in random order, in an unbroken
2-minute synthesized speech stream:
bidaku-golabu-dutaba-golabu-padoti-bidaku-dutaba-padoti-golabu-dutaba-
padoti-dutaba-golabu-bidaku-dutaba-bidaku-padoti-bidaku-padoti-golabu etc.
Infant
Test phase
Figure 4.2 In this study, Saffran and colleagues
Loudspeakers present infant either a “real” word: prepared stimuli that amount to a miniature
bidaku, golabu artificial language of four “words,” each word
or a sequence of syllables with parts of two words: consisting of three consonant-vowel syllables.
dakugo, buduta Infants then heard an uninterrupted, 2-minute
stream of random combinations of the four
Results
words. The researchers noted how much atten-
Mean looking times for 8-month-olds tion the babies paid to the four “words” from
“Real” words 6.77 seconds the familiarization phase and compared it with
Part-words 7.60 seconds the attention the babies paid to three-syllable
Speaker
sequences that also occurred in the speech but
that straddled “word” boundaries (part-words).
(Adapted from Saffran et al., 1996.)

artificial language A “language” that sents a miniature artificial language. That is, the string of sounds corresponded
is constructed to have certain specific to concatenations of “word” units combining with each other. In this particular
properties for the purpose of testing an “language,” each “word” consisted of three consonant-vowel syllables (Figure
experimental hypothesis: strings of sounds 4.2). For example, bidaku in the above stream might form a word. The unin-
correspond to “words,” which may or may terrupted two-minute sound stream consisted of only four such “words” ran-
not have meaning, and whose combina-
domly combined to form a sequence of 180 “words” in total, which meant that
tion may or may not be constrained by
each “word” appeared quite a few times during the sequence. (The fact that the
syntactic rules.
words were randomly combined is obviously unrealistic when it comes to how
real, natural languages work. In real languages, there’s a whole layer of syn-
tactic structure that constrains how words can be combined. However, for this
study, the researchers were basically only interested in how infants might use
very limited information from sound sequences to isolate words.)
Later, during the test phase, the researchers noted how much attention the
babies paid to actual “words” they heard in the familiarization phase and com-
pared it with the attention the babies paid to three-syllable sequences that also
occurred in the speech but that straddled “word” boundaries—for example,
e Sedivy dakupa (see Figure 4.2). The infants showed a distinction
between “words” and “part-word” sequences. In this
WEB AC TIVIT Y 4.3 case, they were more riveted by the “part-words,” listen-
12-02-13
Segmenting “words”: You be the ing to them longer than the “words”—possibly because
baby! In this activity, you’ll hear 2 min- they were already bored by the frequent repetition of the
utes of stimuli from an artificial language “word” units.
very similar to that used by Saffran et al. (1996). You’ll How did the 8-month-old infants Saffran et al. studied
be asked to distinguish “words” from “non-words” manage to do this? If the sound stimuli were stripped of
to see if you, too, can manage to segment speech. all helpful features such as already-familiar words, stress
(Ideally, you should attempt this exercise before you patterns, and phonotactic cues, what information could
read any further.) the babies possibly have been using in order to pull the
“words” out of the 2-minute flow of sound they’d heard?
To see how you’d fare in such a task, try Web Activity 4.3.
The answer is that there’s a wealth of information in the speech stream
waiting to be mined, and it’s there just by virtue of the fact that the stream is
composed of word-like units that turn up multiple times. Saffran and her col-

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
 Learning Sound Patterns   117

leagues suggested that babies were acting like miniature statisticians analyzing transitional probability (TP) The prob-
the speech stream, and were keeping track of transitional probabilities (TPs) ability that a particular syllable will occur,
between syllables—this refers to the likelihood that one particular syllable will given the previous occurrence of another
be followed by another specific syllable. Here’s how such information would particular syllable.
help to define likely word units: Think of any two syllables, say, a syllable like
ti and a syllable like bay. Let’s say you hear ti in a stream of normal English
speech. What are the chances that the very next syllable you hear will be bay?
It’s not all that likely; you might hear it in a sequence like drafty basement or
pretty baby, but ti could just as easily occur in sequences that are followed by
different syllables, as in T-bone steak, teasing Amy, teenage wasteland, Fawlty Tow-
ers, and many, many others.
But notice that ti and a bay that follows it don’t make up an English word. It
turns out that when a word boundary sits between two syllables, the likelihood
of predicting the second syllable on the basis of the first is vanishingly small.
But the situation for predicting the second syllable based on the first looks very
different when the two syllables occur together within a word. For example,
take the sequence of syllables pre and ti, as in pretty. If you hear pre, as pro-
nounced in this word, what are the chances that you’ll hear ti? They’re much
higher now—in this case, you’d never hear the syllable pre at the end of a word,
so that leaves only a handful of words that contain it, dramatically constrain-
ing the number of options for a following syllable. Generally, the transitional
probabilities of syllable sequences are much higher for pairs of syllables that fall
within a word than for syllables that belong to different words. This is simply
because of the obvious fact that words are units in which sounds and syllables
clump together to form a fairly indivisible whole. Since there’s a finite number
of words in the language that tend to get used over and over again, it stands
to reason that the TPs of syllable sequences within a word will be much higher
than the TPs of syllable pairs coming from different words.
How does all this help babies to segment speech? Well, if the little tykes can
somehow manage to figure out that the likelihood of hearing ti after pre is quite
high, whereas the likelihood of hearing bay after ti is low, they might be able to
respond to this difference in transitional probabilities by “chunking” pre and ti
together into a word-like unit, but avoid clumping ti and bay together.
Here’s the math: The transitional probability can be quantified as P(Y|X),
that is, the probability that a syllable Y will occur given that the syllable X has
just occurred. This is done by looking at a sample of a language and dividing
the frequency of the syllable sequence XY by the frequency of the syllable X
combined with any syllable:
frequency (XY)
TP = P(Y|X) =
frequency (X)
In the study by Saffran and her colleagues, the only cues to word boundaries
were the transitional probabilities between syllable pairs: within words, the
transitional probability of syllable pairs (e.g., bida) was always 1.0, while the
transitional probability for syllable pairs across word boundaries (e.g., kupa)
was always 0.33.
That babies can extract such information might seem like a preposterous
claim. It seems to be attributing a whole lot of sophistication to tiny babies. You
might have even more trouble swallowing this claim if you, a reasonably intel-
ligent adult, had trouble figuring out that transitional probabilities were the rel-
evant source of information needed to segment the speech in Web Activity 4.3
(and you wouldn’t have been alone in failing to come up with an explanation
for how the speech might be segmented). How could infants possibly manage
to home in on precisely these useful statistical patterns when you failed to see

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.
118  Chapter 4

them, even after studying the speech sample and possibly thinking quite hard
about it?
Though it might seem counterintuitive, there’s a growing stack of evidence
that a great deal of language learning “in the wild”—as opposed to in the class-
room—actually does involve extracting patterns like these, and that babies and
adults alike are very good at pulling statistical regularities out of speech sam-
ples, even though they may be lousy at actually manipulating math equations.
As we’ll see later, sensitivity to statistical information applies not just to seg-
menting words from an unfamiliar language, but also to learning how sounds
make patterns in a language, how words can be combined with other words,
or how to resolve ambiguities in the speech stream. The reason it doesn’t feel
intuitive is that all of this knowledge is implicit and can be hard to access at a
conscious level. For example, you may have done reasonably well in identifying
the “words” in the listening portion of Web Activity 4.3, even if you had trouble
consciously identifying what information you were using in the analysis task.
(Similarly, you may have had easy and quick intuitions about the phonotactic
constraints of English but worked hard to articulate them.) That is, you may have
had trouble identifying what it is you knew and how you learned it, even though
you did seem to know it. It turns out that the vast majority of our knowledge of
language has this character—throughout this book, you’ll be seeing many more
examples of a seeming disconnect between your explicit, conscious knowledge
of language and your implicit, unconsciously learned linguistic prowess.
Being able to track transitional probabilities gives infants a powerful device
for starting to make sense of a running river of speech sounds. It frees them up
from the need to hear individual words in isolation in order to learn them, and
it solves the problem of how they might build up enough of a stock of words
to serve as the basis for more powerful generalizations about words—for ex-
ample, that in English, words are more likely to have a trochaic stress pattern
than an iambic one, or that the consonant cluster “ft” can’t occur at the begin-
ning of a word, though it can occur at its end. Ultimately, these generalizations,
once in place, may turn out to be more robust and useful for word segmentation
than transitional probabilities are. At times, such generalizations might even
conflict with the information provided by transitional probabilities. Eventually,
infants will need to learn how to attend to multiple levels of information, and
to weight each one appropriately.

Is statistical learning a specialized human skill for language?

We’ve now seen some of the learning mechanisms that babies can use to pull
word-like units out from the flow of speech they hear, including keeping track
of various kinds of statistical regularities. Now let’s step back and spend a bit of
time thinking about how these learning mechanisms might connect with some
of the bigger questions laid out in Chapters 2 and 3.
Much of Chapter 2 focused on the questions of whether language is unique
to humans, and on whether certain skills have evolved purely because they
make efficient language use possible. In that chapter, I emphasized that it was
impossible to think of language as a monolithic thing; learning and using lan-
guage involve an eclectic collection of skills and processes. Since we’ve now
begun to isolate what some of those skills might look like, we can ask a much
more precise question: Do non-humans have the ability to segment speech by
keeping track of statistical regularities among sounds?
If you found yourself surprised and impressed at the capacity of babies to
statistically analyze a stream of speech, you may find it all the more intrigu-
ing to learn that as a species, we’re not alone in this ability. In a 2001 study,

uncorrected page proofs © 2014 Sinauer Associates, Inc. This material cannot be copied, reproduced,
manufactured or disseminated in any form without express written permission from the publisher.