Learning Phonemes Without Minimal Pairs

Learning Phonemes Without Minimal Pairs

Jessica Maye and LouAnn Gerken
University of Arizona


1. Introduction


The question addressed by this research is how humans acquire the
internalized, mental categories that reflect the phonemes1 of their language. That
humans have such categories is evidenced by the fact that, for example, English
speakers immediately recognize pear and bear as being different words. This
immediate realization is independent of the fact that these two words have
different meanings, since English speakers will also report that nonsense words
like bove and pove are different words – despite the fact that neither word means
anything.

Phonemic categories, once acquired, have an influence on the perception of
speech sounds. Although adults can easily perceive acoustic distinctions that
differentiate the categories of their own language, they often have difficulty
perceiving non-native phoneme contrasts. For example, the English distinction
between /r/ and /l/ is difficult for native Japanese speakers to perceive
(Miyawaki et al. 1975), and the Hindi distinction between /t/ and /ţ/ (the latter
sound is retroflex, pronounced with the tongue curled backwards) is difficult for
native English speakers to perceive (Werker et al. 1981).

Infants are born with the ability to categorize speech sounds (Eimas et al.
1971), but these categories are initially universal, not based on the phonemes of
a particular language. For instance, Japanese-learning infants can initially
discriminate English /r/ and /l/ (Tsushima et al. 1994), and English-learning
infants can initially discriminate Hindi /t/ and /ţ/ (Werker et al. 1981). Over the
course of the first year of life, though, the infant gains experience with the native
language, and begins to only discriminate contrasts that represent a phonemic
distinction in the native language (Werker & Tees 1984). Through our research,
we would like to understand how these phoneme-related effects on speech
perception arise in the language learner.

2.0 The Acquisition of Phoneme Categories


Two types of hypotheses have been proposed to explain how these
phoneme-based categories are acquired, which can be roughly characterized as
minimal pair-based learning versus distribution-based learning. The minimal
pair-based hypothesis is that infants begin to attend to native language
phonemes when they learn that a phonetic distinction can differentiate the


© 2000 Jessica Maye and LouAnn Gerken. BUCLD 24 Proceedings, ed. S.
Catherine Howell et al., 522-533. Somerville, MA: Cascadilla Press.

---

523
meanings of two words. For example, if an infant learns that the words pear and
bear consistently refer to different items, they will begin to attend to the /p/~/b/
voicing distinction, since it can differentiate between two meanings. Under this
hypothesis, it is crucial for an infant to actually know the meanings of words
that form minimal pairs, in order to acquire phoneme categories.
The distribution-based hypothesis is that infant speech perception is shaped
by the native language sounds directly, on the basis of the distribution of
phonetic exemplars. For example, in a language in which /p/ and /b/ are
different phonemes, there will be considerable phonetic variation in the actual
exemplars of the two phonemes, with some overlap between the two categories.
However, exemplars of a particular phoneme will presumably cluster together
along one or more acoustic dimensions, and these clusters can be used to
differentiate the categories which are used contrastively in a language. Under
this hypothesis, speech perception abilities are automatically shaped by the
statistical distribution of exemplars that an infant is exposed to, and word-
learning is not a necessary prerequisite.

2.1 The Minimal Pair Hypothesis

In response to the early studies on infant speech perception, MacKain
(1982) argued that in order to be shaped by the ambient language, an infant must
experience members of different phonetic categories as representing a contrast.
That is, the infant must be aware that they are experiencing sounds that are
contrastive in the language, before their perceptual system will be influenced by
the contrast in question. This argument amounts to a requirement that infants
know whether a given phonetic distinction can result in a meaning distinction in
their language; in essence, it proposes minimal pair-based learning.
A similar suggestion was made by Werker & Tees (1984), when they found
that the shift from language-general to language-specific speech perception
occurs during the latter half of the first year. Since it is also during the latter half
of the first year that infants begin to understand and produce words, these
researchers suggested that the shift might occur because of the infant’s
developing lexicon.
Despite the appeal of the minimal pair-based hypothesis, there is emerging
evidence that it cannot be true. One piece of evidence comes from the timeline
of word learning. If the reorganization of speech perception is based on the
knowledge of minimally contrastive words, no reorganization should be evident
before a child knows any minimal pairs. Since the perceptual reorganization has
been shown to begin between 8-10 months for consonants (Werker & Tees
1984), and as early as 6 months for vowels (Kuhl 1991, Polka & Werker 1994),
the minimal pair-based hypothesis predicts that infants have learned minimal
pairs that contrast vowels by 6 months, and pairs that contrast consonants by 8-
10 months. However, infants have not been shown to possess a large receptive
vocabulary before the age of 12 months, and what words they do know may not
include minimal pairs. Caselli et al. (1995) performed an extensive cross-

---

524
linguistic study on the content of the expressive and receptive lexicons of
English-learning and Italian-learning infants between 8 and 16 months of age.
According to their findings, the average 8-month-old can understand around 36
words. However, the 50 words most likely to be in the early receptive lexicon of
English-learning babies do not include a single minimal pair. For Italian-
learning babies, they found a single minimal triplet (nonno “grandpa”, nonna
“grandma”, nanna “sleep/bedtime”), which differed minimally with respect to
vowels, not consonants.
A second argument against the minimal pair-based hypothesis comes from
recent experimental findings on word-learning and its interaction with speech
perception. Stager & Werker (1997) conducted a study to test infants’ ability to
discriminate minimal pairs of nonsense words (e.g. bih and dih). They found
that it is not until 18 months of age that infants can discriminate minimal pairs,
when those words have semantic referents. That is, when infants are beginning
to learn word meanings, it interferes with the ability to discriminate fine
phonetic detail. Only older infants, who are more adept at word-learning, are
able to discriminate minimal pairs of words.

2.2 The Distribution-Based Hypothesis

In support of a distribution-based account of phoneme learning, previous
research has demonstrated that infants utilize distributional information in other
areas of language learning. One such study was conducted by Jusczyk, Luce,
and Charles-Luce (1994), who showed that infants are aware of the relative
frequency with which various phonotactic patterns occur in their language. They
presented English-learning 9-month-old infants with nonsense words having
either common English phonotactics (e.g. mubb), or legal but less frequent
patterns (e.g. jurth). The infants preferred to listen to the words with frequently
occurring phonotactics, indicating that they recognized the difference between
the frequent and infrequent patterns.
Another study showing infants’ use of distributional information in
language learning was conducted by Saffran, Aslin, and Newport (1996), who
showed that 8-month-old infants can utilize statistical information about syllable
co-occurrence, in order to segment a continuous stream of speech into words.
This study presented infants with continuous synthetic speech composed of 3-
syllable nonsense words (e.g. tupiro, golabu, bidaku, padoti), pronounced
without stress or any information about word boundaries. The researchers
hypothesized that infants might be able to learn which syllables went together to
form words, on the basis of how often syllables occurred next to each other.
Syllables within the same word would often occur next to each other, in a
particular order (e.g. golabu#padoti), while syllables from different words
would only occur next to each other if their words happened to be adjacent (e.g.
tupiro#bidaku). After listening to this continuous speech for 2 minutes, infants
showed a preference for the experimental nonsense words presented in isolation,
over non-words (syllables presented in random order: e.g. tidoku) and part-

---

525
words (sequences of syllables that had occurred during training, but crossed a
word boundary: e.g. padoti#bidaku).
The findings of the above studies indicate that infants have access to
detailed information about the relative frequency with which elements of their
language occur. This type of information could potentially also be used for
learning a language’s phoneme categories. Guenther & Gjaja (1996) proposed a
computational model of speech perception, which demonstrates how a neural
correlate of phoneme categories might be acquired, on the basis of the statistical
distribution of exemplars in a language. In their model, exposure to a particular
language leads to nonuniformities in the distribution of the firing preferences of
neural cells in the auditory system. Over time, proportionately more neurons
become devoted to firing in response to the most frequently heard sounds.2 This
results in clusters of neurons that reflect the phoneme categories of the language.
These clusters, in turn, give rise to the phoneme-based effects on speech
perception.

3.0 Experiment

Guenther & Gjaja’s model elegantly exemplifies the distribution-based
hypothesis of phoneme acquisition, and argues that such a model is neurally
plausible. However, what remains to be shown is that humans are actually
capable of forming categories on the basis of the distribution of phonetic
exemplars. Our experiment was designed to test this hypothesis by presenting
adult participants with phonetic exemplars from an artificial language, and
giving them no information about word meaning.
We designed the experiment according to the following reasoning. In real
speech, speakers produce sounds with a large degree of phonetic variation.
When a language has two, contrastive phoneme categories (like English /b/ and

Figure 1: Monomodal vs. Bimodal Distributions




Frequency
of

Occurrence
b
p




(short VOT)
e.g., Voice Onset Time
(long VOT)


= Monomodal distribution
= Bimodal distribution

---

526
/p/), although the categories are pronounced with much variation and even
overlap, the most frequently heard tokens will fall into two clusters, forming a
bimodal distribution, as shown by the broken line in Figure 1. When a language
has a single phoneme category along some acoustic dimension (e.g. voice onset
time), its tokens will fall into a single cluster, forming a monomodal
distribution, as shown by the solid line in Figure 1.
In our experiment, we presented two groups of participants with the same
stimuli, but varied the frequency with which they heard each token, such that
one group was exposed to a monomodal distribution, and the other group a
bimodal distribution. We then tested whether each group treated the stimuli as
corresponding to one or two phoneme categories.

3.1 Stimuli

Because our intent was to test adult English speakers, we needed to find a
contrast that English speakers are actually able to perceive, but one that they do
not discriminate as a phonemic contrast. The reason for this is that we wanted to
bias the two groups of participants in different directions: we wanted the
bimodal group to attend to the acoustic differences, and treat them as phonemic,
while the monomodal group should ignore the acoustic differences, treating
them as non-phonemic phonetic variation. A study by Pegg & Werker (1997)
provided us with an appropriate contrast: between English voiced /d/ (as in day)
and voiceless unaspirated /t/ (as in stay). Although both of these sounds come
from English, they do not constitute a phonemic contrast in English, since they
never occur in the same environment (i.e. unaspirated /t/ only occurs after /s/ in
English, while voiced /d/ never does). English speakers perceive both of these
sounds as members of the /d/ category, when occurring in syllable initial
position. For example, if the /s/ is removed from the word stay, English
speakers will perceive it as the word day. However, Pegg & Werker showed
that in a discrimination task, English speakers can hear the acoustic differences
between voiced /d/ and unaspirated /t/.
It is also important to point out that, although the distinction between voiced
/d/ and voiceless unaspirated /t/ is phonemic in many languages (e.g. Spanish,
French, Japanese), the particular sounds used in this experiment (and in Pegg &
Werker 1997) are taken from English syllables beginning with /d/ and /st/. The
distinction between /d/ and /t/ is generally characterized as a “voicing”
distinction; however, in English, actual tokens of /d/ and /st/ do not differ in
terms of voice onset time. Instead, in our own measurements (cf. measurements
reported by Pegg & Werker 1997), the differences between English /d/ and /t/
are in the formant onset frequencies for the following vowel, with /d/ having a
more extreme transition from formant frequencies at the onset of the following
vowel to the center of the vowel. Also, although prevoicing is not reliably
produced in English, it was included in /d/ tokens, in order to ensure that the two
sounds were different enough to be distinguished by participants. For our
experiment, the cues that distinguish these two sounds are of little consequence,

---

527
so long as the stimuli satisfy two criteria: that they be not readily distinguished,
yet still be discriminable to native English speakers.
To create our stimuli, we began with natural English productions of the
syllables /da/, /sta/, /dæ/, /stæ/, /dr/, and /str/. We then removed the /s/ from the
/s/-initial syllables, resulting in the syllables /da/, /ta/, /dæ/, /tæ/, /dr/, and /tr/.
These six syllables were then re-synthesized into three continua, running from
/d/ to /t/ (in each of the three vowel contexts) in eight equal steps.
Filler stimuli were syllables beginning with the consonants /m/ and /l/ (/ma/,
/mæ/, /mr/, /la/, /læ/, and /lr/). There were four tokens (different utterances of
each filler syllable), each presented twice per block of training, for a total of 24
filler stimuli per block.
During the test phase, participants were presented with only the endpoint /d/
and /t/ stimuli. These stimuli were paired with themselves (e.g. /da/~/da/) on
“same” trials, paired with each other (e.g. /da/~/ta/) on experimental “different”
trials, or paired with filler items (e.g. /da/~/ma/) on filler “different” trials. Filler
pairs consisted of pairs of identical filler stimuli (e.g. the same utterance of /ma/
repeated twice), pairs of nonidentical filler stimuli (e.g. two different utterances
of /ma/), and pairs of different filler stimuli (e.g. /ma/~/la/).

3.2 Participants

Participants were 32 native English speakers (23 female, 9 male) enrolled in
courses at the University of Arizona who received course credit for their
participation, or who volunteered to participate. Their ages ranged from 18 to 41
years, with a mean age of 23. All had normal hearing and no language
impairment.
Participants were randomly assigned to one of two groups for training. The
two groups differed with respect to the distributional frequency of stimuli
presented during training. One group was presented with a monomodal
distribution of each continuum, in which the tokens from the center of each
continuum were presented four times as often as the endpoint tokens, as shown
by the solid line in Figure 2. The other group was presented with a bimodal
distribution, in which the tokens from the near the endpoints of each continuum
were presented four times as often as the center tokens, as shown by the broken
line in Figure 2. In this way, both groups of participants were presented with all
tokens along each continuum, the only difference being the frequency of
stimulus presentation. Both groups of participants heard 16 experimental stimuli
in each of the three vowel contexts, for a total of 48 experimental stimuli per
block of training. Also, both groups heard tokens 1 and 8 from each continuum
only once per block. This was because tokens 1 and 8 were used as the
contrasting stimuli during the test phase (see Procedure section), and we wanted
to ensure that both groups had heard the test stimuli the same number of times.

---

528
Figure 2: Stimuli Presentation Frequency during Acquisition Phase



4
Number of
Presentations 3
per Block of

2
Training


1


1
2 3 4 5 6 7
8



/da/
Token Number
/ta/

= Monomodal group

= Bimodal group


3.3 Procedure

Participants were informed that in this experiment they would be listening
to a language they had never heard before, with the purpose of learning about
the sounds of the language. After listening to words from the language, they
would be given a task in which they would hear pairs of similar-sounding words,
and would have to decide whether each pair was the same word repeated twice,
or two different words in this language.
Practice Task: Participants were first given a practice task, during which
they heard 10 pairs of English words, half of which were “same” pairs (two
utterances of the same word), and half of which were “different” pairs (two
English words differing only in a single consonant). Participants were instructed
to mark either “same” or “different” on a response sheet to indicate their
answers. The inter-stimulus interval within each pair was 500 msec, and
between trials there was a 2 second pause, during which participants recorded
their responses.
Acquisition Phase: During the acquisition phase, participants were
presented with words from the artificial language presented in list form, with an
inter-stimulus interval of 1 second. No information was provided regarding the
meaning of these words in the language. The words included all of the stimuli
from the three experimental continua (da, ta, dæ, tæ, dr, and tr) as well as all of
the filler stimuli (ma, mæ, mr, la, læ, and lr), presented with the frequency
distributions appropriate for each participant group, as discussed in the Stimuli
section, above. The entire block of experimental and filler stimuli was repeated
four times, for a total listening time of 9 minutes. In order to help the
participants maintain their attention to the stimuli, they were given a check-sheet
with 384 empty boxes on it (one for every stimulus presented during

---

529
acquisition), and instructed to check a box every time they heard a word.
Participants were told that their task during this phase of the experiment was
simply to listen carefully to the words of this language, and the way that the
words sounded; and that the purpose of the check-sheet was simply to help them
pay attention to the words.
Test Phase: After completing the acquisition phase, participants were given
a decision task, in which they were presented with pairs of stimuli and asked to
indicate on a response sheet whether the two stimuli in each pair were the same
word repeated twice, or two different words in this language. Participants were
reminded that this task was the same one they initially performed using English
words. Participants were instructed to listen carefully to each pair since the
items in each pair would sound very similar to each other; and that if they were
unsure of their response, they should make their best guess and then be prepared
to listen to the next pair.

During the test phase, the inter-stimulus interval was 500 msec, with 2
seconds between each pair. The items of interest were the pairs of words like
/da/~/ta/, in which one word began with /d/ (the /d/ endpoint of one of the
continua), and the other word began with /t/ (the /t/ endpoint of the same
continuum). We predicted that the two participant groups should differ on their
responses to these pairs, and these pairs only. According to our hypothesis, the
participant group that had been trained on the monomodal distribution should
believe that in this language there is only one phoneme represented by the /d/~/t/
stimuli; therefore, these participants were predicted to respond “same” to these
items. The participant group trained on a bimodal distribution, however, should
believe that these stimuli represent two phonemes in this language; therefore,
these participants were predicted to respond “different” to these items.

3.4 Results

To calculate each participant’s performance on the test, responses were
scored as either “correct” or “incorrect.” For the filler pairs, pairs of words that
were either pairs of identical stimuli, or pairs that were two utterances of the
same filler word (e.g. /ma/~/ma/) were scored as correct if the subject responded
“SAME”. Filler pairs that consisted of two different filler words (e.g. /ma/~/la/)
were scored as correct if the participant responded “DIFFERENT.” And for the
experimental pairs (e.g. /da/~/ta/), responses were scored as correct if the
participant responded “DIFFERENT.” Because of this last scoring choice,
participants from the bimodal training group, who were expected to distinguish
the experimental contrast more often, were expected to receive higher scores on
the test.
We performed a 2 Training Group x 5 Test Pair Type anova on the number-
correct data. There was a significant effect of Group (F = 5.35, p < .05), with
participants from the monomodal training group scoring lower than those from
the bimodal group. There was also a significant effect of Test Type (F = 97.58, p
< .0001), with scores on the /d/~/t/ contrast pairs significantly lower than scores

---

530
for the four types of filler pairs, which were at ceiling. Importantly, there was
also a significant interaction between Group and Test Type (F = 7.65, p <
.0001). Follow-up comparison revealed the only significant effect of Group was
for the /d/~/t/ pairs, with participants from the monomodal training group
scoring lower on the /d/~/t/ contrast pairs than did participants from the bimodal
training group (t = 2.19, p < .05).
The results for /d/~/t/ pairs, the experimental contrast, are illustrated in
Figure 3. Participants from the bimodal training group were more likely to
respond “DIFFERENT” to the experimental /d/~/t/ pairs, indicating that these
sounds represent a phonemic contrast in this language. This finding confirms our
hypothesis that humans can utilize distributional information in order to form
phoneme categories. The group that was trained on a two-cluster distribution
was more likely to indicate that the stimuli corresponded to two categories, than
was the group trained on a one-cluster distribution.

Figure 3: Results, Experimental Contrast Pairs
100%
80%
Percent of
60%
"different"
responses to
40%
/d/~/t/ pairs
20%
0%
Monomodal
Bimodal

4. Conclusion


The results of this experiment support a distribution-based model of
phoneme learning. These results cannot be interpreted as arising from factors of
the participants’ native language, since participants from both groups were
native English speakers. The only difference between the two groups of
participants was the statistical distribution of sounds they were exposed to
during training. What is especially interesting, is that participants learned to
discriminate minimal pairs in this language, although they were not trained on
minimal pairs, and were not given any information about the meanings of words
in the language.

The findings of this experiment highlight the role of particular exemplars,
and their frequency of occurrence, for learning phoneme categories. These
findings suggest that humans maintain some sort of mental histogram for
acoustic patterns they encounter in their language. We remain agnostic as to

---

531
how speech sounds are represented neurally, as this is an issue of much
contention in the field of speech perception; however, Guenther & Gjaja (1996)
and Guenther et al. (1999) present algorithms for how speech sounds could
result in histogram-like organization of the auditory cortex.

There are several directions in which this research can be extended. First,
we would like to know whether the categories formed are specific to training
items, or whether they are generalizable to new tokens, for example, spoken by a
new voice. In this experiment, participants were tested on their categorization of
stimuli that they had heard during training (specifically, the endpoint /d/ and /t/
tokens). In reality, though, listeners rarely encounter the same exemplars more
than once. To test generalization, we plan to test participants’ categorization of
new tokens not presented during training.

Another direction for extending this research is to test the language-
specificity of category learning. Do learners have expectations about speech
sound categories that are particular to language? Or is this type of category
formation performed by a more general learning mechanism? To test this, we
plan to test participants on their categorization of a new contrast that was not
presented during training (e.g. /g/~/k/). Since languages tend to have multiple
analogous contrasts, such that a language with a /d/~/t/ contrast might also have
a /g/~/k/ contrast, learners might expect to encounter multiple, analogous
contrasts.

In addition, this type of generalization will enable us to test whether
learning is constrained by linguistic markedness. Linguistic markedness refers to
cross-linguistic regularities in linguistic inventories. For example, all languages
have coronal sounds (pronounced with the tip of the tongue, like /d/), but not all
languages have dorsal sounds (pronounces with the body of the tongue, like /g/).
The commonly-occurring sound (/d/) is said to be “unmarked,” while the
relatively less common sound (/g/) is said to be “marked.” It will always be the
case that a language with the sound /g/ also has the sound /d/, but the reverse is
not always true (having /d/ does not imply the presence of /g/). If the learning
mechanism for phoneme categories is constrained by markedness implications, a
learner who has heard the sound /g/ might assume the sound /d/ will also occur
in the language; while a learner who has heard the sound /d/ will not have
evidence for the presence of /g/. If markedness implications have this effect on
phoneme learning, it would be evidence that the learning mechanism is
constrained by factors that are specific to language.

And finally, because our hypothesis initially arose from findings regarding
infant language development, an important extension of this research is to test
whether infants are also capable of utilizing distributional information for the
purposes of learning phoneme categories. The previous research showing
infants’ use of statistical information in other areas of language learning,
suggests that infants will utilize this type of information for phoneme learning as
well. Also, since young infants do not have well-established native language
phoneme categories, they may learn new categories even better than adults do,

---