Infant sensitivity to distributional information can affect phonetic

discrimination

Jessica Maye, Janet F. Werker, LouAnn Gerken

A brief article from Cognition 2001

The goal of the study

It is a well known fact that babies, in addition to being really cute, are very good at distinguishing sounds.

This study aimed to see if babies are influenced by statistical distribution of phonetic variation.

The assumption is that babies somehow “track” stochastic information in order to determine the linguistic relevance of acoustic properties.

This is going to be investigated by familiarizing babies with sounds presented under different statistical distributions, then testing their discrimination of the sounds later.

Let's get stochastic

Stochastic refers to certain type almost of randomness. A stochastic process in one that is not deterministic. It is a process, so it follows some path, but you can't determine from one state what the next will be.

VOT is an example of one place to use “stochastic” when talking about linguistics. You could also just throw it into conversation if want to sound pretentious.

Any two [d]s are probably not acoustically similar, but they are more acoustically similar to each other than any given [t].

What babies can do when

12 months –

6 months –

9 months –

10 months –

16 months –

So it would seem babies are sensitive to certain distributional patterns. Let's test that, shall we?

can also distinguish native and non native vowels

can distinguish between frequently occurring speech sounds and rare speech sounds

can distinguish native and non-native consonants

sound discrimination is similar to adults in their community

stops spitting up on your best clothes

The test

The experimenters created an 8-point continuum of speech sounds using voiced vs. voiceless unaspirated stops, since babies have been shown to be sensitive to this.

Babies were then divided into two groups: the Bimodal group and Unimodal group.

Bimodal? That sounds kinky...

Not really. It refers to a type of statistical distribution.

The Bimodal group was familiarized with a bimodal distribution frequency, so that sounds occurring at either end of the continuum were heard more often.

The Unimodal group, not surprisingly, was familiarized with a unimodal distribution frequency, so that sounds from the center of the continuum were heard more frequently.

This will be demonstrated in Fig 1 on the next slide

Fig 1

Sorry, I mean Figure 1

The hypothesis

Babies exposed to a bimodal distribution would form two categories and those exposed to a unimodal distribution would form one category.

Also to note, they would be able to do this without any clues as to whether these sounds represented different or similar meanings (in other words it is all about acoustics and phonology).

If this first predictions holds, then after a familiarization phase wherein the stimuli had these patterns, the bimodal babies would be better able to discriminate sounds.

Where did they get these babies from?

There were 48 babies that underwent the experiment, 24 of them 6 months old and 24 of them 8 months old.

The ages were chosen because 8 month old babies have been shown to be sensitive to statistical information, but 6 months old babies have not.

All babies were native speakers of English (or rather, will become native speakers of English)

There was an even mixture of gender within each group.

What could possibly go wrong?Sadly, not all of the babies worked out. Here are some of the things that go wrong with testing, and how many babies' results were excluded because of it:

Failure to meet language requirement (2 babies)

Crying (2 crybabies)

Inattention to visual stimulus (2 babies)

Equipment failure (1 baby)

Baby failure (0 babies)

Experimenter error (1 baby)

Parental interference (1 baby)

Not a baby but actually an alien (0 babies)

In all, 12 babies had to be excluded

The procedure and stimulus

Babies heard one of four sounds during the first phase: [ta], [da], [la], or [ma]. Those second two were there to de-emphasize the acoustically relevant information, and make the task more like language acquisition.

They heard unaspirated [ta] and [da], recorded by an American English speaker. The [ta] was excised from stop to avoid aspiration. These were digitally edited and placed along an 8 point continuum. Sounds were about half a second in length

During the second phase, babies heard either [ta] or [da], and their looking time was recorded.

The test consisted of two phases, a familiarization phase, and the real testing phase.

The familiarization phase

In this phase, babies heard 6 blocks of 24 syllables. In each block, there were 16 [ta]s or [da]s falling along the continuum in accordance with the group the baby was in.

Babies also heard 8 filler syllables of either [la] or [ma]

Syllables were presented in a random order, and the whole ordeal lasted 2.3 minutes.

The Experiment

Experimenters used paradigm developed in 1998 by Best and Jones.

In this, there are two types of trials: Alternating and Non-Alternating.

As you might guess, in Alternating trials babies heard two different stimuli, and in Non-Alternating trials, one stimulus was repeated.

There were 8 tests, and half of the test trials were Alternating, and half were Non-Alternating, switching between the two.

On half of the Non-Alternating trials, token 3 was heard and on the others token 6 was heard. In the Alternating trials tokens 1 and 8 were heard. There was a total of 8 tokens played each trial

Each sound was about .5 sec and ISI was 1 sec, so each trial lasted 11 seconds. Babies' looking times were recorded as a measure of how well they discriminated the sounds.

Results

Babies in the Bimodal group generally had longer looking times, and specifically looked longer on Non-Alternating tests, and babies in the Unimodal group showed no real preference.

This is assumed to mean that only those familiarized with bimodal stimuli were able to discriminate the sounds.

What does it all mean?

The researchers draw the conclusion attention to statistical distribution in speech may be a factor in learning to discriminate speech sounds. They suggest that given the results, it may be the case that younger infants perform similarly on these tasks.

Since the Unimodal group showed nothing significant in their looking times, even those age groups have been demonstrated to be receptive to statistical distribution, it is concluded that Unimodal input may be detrimental to learning.

This could also explain why infants learn to discriminate vowels earlier than consonants. There are always fewer Vs than Cs in a language, and the Vs make up the majority of the input.