LING 2120 chapter 5 speech segmentation

Segmentation

children must segment the speech stream into linguistic units (words, phrases, clause)

---

is required for successful acquisition, as the production of words requires the identification of words in the speech stream

Challenges for speech segmentation (2)

Speech is continuous so id units within the speech stream is challenging

Infants are generally not taught words in isolation; they must build a lexicon, in bottom-up fashion, through pairing strings of segments with meaning

How does speech segmentation happen?

chat- is the process by which listeners (especially infants) figure out where one word ends and the next begins in a continuous stream of speech — because spoken language doesn’t come with neat pauses between words.

---

Through phonological bootstrapping- the use of phonological properties of the speech stream to identify linguistic units and the boundaries between them. bootstrapping babies can pick themselves up by their bootstraps by noticing patterns in the prosody of their language, such as the location of pauses in speech.

Rhythms built out of syllables

Finding boundaries relies on…

Prosodic cues- pauses, changes in frequency, lengthening segments in certain positions

Prosodic information is what is contained in the lower frequencies of the acoustic signal, and it is this information that is most available to an infant in utero.

What also plays a role along with prosodic cues?

Transitional probabilities and phonetics

Phonotactics

the set of constraints on how sound segments can be arranged

Segmental phonology- concerned with how sounds interact when they are close to eachother

Prosody

deals with suprasegmental features such as stress and intonation

Phonotactic constraints

the ways that sounds can be combined in a language, the knowledge behind judgements of what is acceptable or not

Different languages have different restrictions on possible syllables

Ex: eng;ish has complex onsets ( less than 3 consonants) complex codas

Transitional probability TP

refers to the probability of a segment or syllable occurring, given an adjacent segment or syllable

Are TPs sensitive to phonotactic constraints?

Yes! Constraints on which segments can appear adjacent to which others within across syllables in a given language

Phontotactically well formed sequences are more likely to…

Occur next to each other within a syllable or word than phonotactically ill formed sequences

Ex: tr is common across word boundaries (train, attract)

tl is uncommon with words, BUT occurs across word boundaries

Train, atlas

Experiment

Saffran, aslin and newport- experiment with 8M old infants

Exposure phase- They listened to a string of CV syllables for 2 minutes

Some of the syllables co-occured more frequently; syllables with a higher rate of co-occurance are equivalent to words

Experimental phase- they were presented with “words” based on previously established transitional probabilities

Did infants listen to part words more than words?

Yes, they were interested in it because it sounded new. This suggests that infants were surprised by and more interested in the novel words

Statistical learning

the difference in level of interest is only possible if there is learning based on occurrences that the infants detect in the input

Findings from the experiment

8-month-old infants can use this information and prefer new “words” created from word parts, such as kugola, over old words, such as bidaku

Finding Boundaries: prosodic cues

Pasek, nelson, juszczyk, cassidy, cruss and kennedy

experiment where 7-20 M were presented with two versions on a story ( using IDS)

Version 1: pauses appear in natural places

Version 2: pauses appear in unnatural places

Results

infants listened longer to the coincident version (V1, pauses line up with natural linguistic boundaries)

Christophe, dupoux, bertoncini, mehler

another prosodic cue exper

experiment where french exposed infants at 3 days old are sensitive to prosodic differences involved presence or absence of a word boundary

Infants can discriminate a word-internal bisyllabic string from a string where the two syllables are different words

But how are infants so young sensi5ve to prosodic cues to word boundaries (such as syllable duration cues

Prosodic information is available in …

utero

T/F: if a given language (rela9vely) consistently exhibits stress in the same loca9on in a word, children will learn that the word boundary aligns with or is close to the stressed syllable

true

85% of words in English…

have primary stress on first syllable

are infants sensitive to initial stress in english?

f ya

7.5 month old english exposed infants are sensitive to initial stress in english

They can segment words with initial stress (ex: doctor), but can't with final stress (ex: guitar)

experiment

7.5 month old english exposed infants are sensitive to initial stress in english

They can segment words with initial stress (ex: doctor), but can't with final stress (ex: guitar)

Familiarization: Infants familiarized on pairs of strong/weak or weak/strong words (across experiments)

Testing: Infants tested on passages that either contained or did not contain the familiarized target words

results from the exper

Results:

7.5-month-olds detected the familiarized target words in the passages only when

the words were strong-weak like dóctor: they listened significantly longer to the passages containing these words

important:In passages with weak-strong (where stress was on the second syllable) words like guitár, infants mis-segmented táris as a word in a string like guitár is

Function words

express the grammatical relationships that hold between other lexical words in a sentence, they are a close class

Articles

Determiners

Pronouns

conjunctions

Function words are a _____ class

Closed. We don't add new function words to a language

Lexical words

express the core information or content of a sentence

lexical words are an _____ class

open

new lexical words can be added to a language at any time

nouns

verbs

adjectives

Function words are good cues to…

segmentation

• Because they are a closed class (not that many to learn) and frequent

• Because the type of function word tells you about the type of lexical (open class) word that will follow

English- determiners are followed by a nominal element ( adjective/noun) not a…

Verb

Ex: the book, *the eat

French- determiners are followed by a nominal element as well ( almost always a noun, sometimes an adjective)

e livre bleu ‘the blue book’ (Det-Noun-Adj)

Segmentation in french

many determiners in french are cv in shape (optimal shape for syllables)

Children segment det +noun strings in the input by id the cv shape of function words.. Except,

this doesn't always work.

Segmentation errors in french

Children's segmentation errors suggest they are sensitive to and seek out Cv shaped determiners followed by c initial nouns

The most common error pattern- extra z due to many determiners being pronounced with z before v initial nouns

Summary

Infants use phonological properties to segment speech ( known as phonological bootstrapping)- breaking into the speech stream

Cues to speech segmentation include prosodic properties (stress, lengthening, pauses)