Speech segmentation

Question 1

Jusczyk and Aslin (1995) 7.5 months repetitions of sentences containing 2 target words

Answer 1

Tested on individual words. Infants listened longer to words embedded in familiarisation sentences. By 7.5 months, at least some rudimentary ability to detect words when they occur in fluent speech contexts.

Question 2

6 possible ways to help segment speech stream

Answer 2

1. Prosodic cues.
2. Phonotactic regularities.
3. Allophonic variation.
4. Isolated words.
5. Transitional probabilities.
6. Multiple cues approach.

Question 3

What are prosodic cues?

Answer 3

Rhythm, stress pattern, and intonation of speech. English has trochaic patten (strong-weak stress) so segmentation through high probability that strong syllable belongs to onset of new word.

Question 4

Jusczyk et al. (1999) prosodic cues familiarisation to individual strong-weak words

Answer 4

7.5 month infants listened longer to passages containing words they were exposed to (suggests segmentation). Effect not seen for weak-strong words (but is at 10.5 months).

Question 5

2 critiques of prosodic cues

Answer 5

1. Not a perfect cue (is segmentation seem before 7.5 months?).
2. Not relevant for all languages.

Question 6

Phonotactic regularities

Answer 6

There are patterns as to which sounds can co-occur within a language (eg. [br] at beginning of word, [nt] at end, [vzg] not legal in Englsih). Sensitivity could provide reasonable cue to word boundaries.

Question 7

When do we see evidence of sensitivity to phonotactic regularities?

Answer 7

About 9 months.

Question 8

Allophonic variation

Answer 8

A phoneme might be pronounced differently at the beginning/ends of words (eg. aspirated vs. non-aspirated).

Question 9

Jusczyk et al. (1994) allophonic variation ‘nitrates’ vs. ‘night rates’

Answer 9

Infants heard examples of ‘nitrates’ vs. ‘night rates’ embedded in continuous speech. Infants listened longer to sentence containing the word they were familiarised to (seem at 10.5 months but not 9 months).

Question 10

Big problem in speech segmentation

Answer 10

Need words to segment and need to segment to get words. Infants need to learn what’s relevant for their language before they can use it for segmentation.

Question 11

Isolated words

Answer 11

Might help with early segmentation. Around 9% of words in IDS occur in isolation and tend to be words learnt first. But this method also identifies phrases as words.

Question 12

Statistical learning

Answer 12

Discovery of patterns or structure in the input. Learning is implicit (without instruction or rules). Infants highly sensitive to distributional info in input.

Question 13

Transitional probabilities

Answer 13

Conditional probability of Y given X in sequence XY.

Question 14

Saffran et al. (1996) familiarised infants to continuous speech of trisyllabic words strung together in random order

Answer 14

8-month infants. Test word vs. non-word vs. part-word (syllable sometimes follows). Looked longer to near novel stimulus (non-word or part-word). Suggests babies are sensitive to transitional probabilities.

Question 15

Factors influencing direction of preference for novelty vs. familiarisation

Answer 15

1. Length of familiarisation.
2. Meaningfulness of stimuli.
   Need pre-registration.

Question 16

Gambel and Yang (2004) computer model of English utterances and use TP to identify word boundaries

Answer 16

Only 41% were real English words. 77% of words were missed. TP not a sufficient cue to identify words from continuous speech on its own.

Question 17

Multiple cues

Answer 17

Use all possible cues (prosody, phonotactic, allophonic, isolated words, TPs). Once some words identified, they can pick up common patterns that mark beginning and ends of words.

Question 18

Thiessen and Saffran (2003) artificial speech stream with colliding cues

Answer 18

7 months-olds: more attention to TPs to segment speech.
2. 9 month-olds: prioritised stress cues over TPs.

Developmental change.

Question 19

Another form of conditional probability

Answer 19

Meaning predicting form. Infant regularly hears ‘mummy’ in speech stream and mother is generally present at the same time. Link the 2 together (correlations).

Question 20

Saffran (2020) statistical learning process in language acquisition

Answer 20

Ability to track probabilistic relationships between different aspects of language. Input-driven approach.

Question 21

2 challenges for statistical learning

Answer 21

1. Do infants actually use it?
2. Rules? Could abstractions about patterns be generalised to novel elements. Language is structured, not just statistical.

Question 22

Graf Estes et al. (2007)

Answer 22

17 month infants. Exposed to continuous speech stream with TP as only clue to word boundaries. Then, habituated to object-label association (objects paired with words and nonwords).

Test: switch or same object-label pair. Switch condition led to longer looking time for words but not nonwords or partwords. Suggests that words treated as linguistic units that can be mapped to meanings.

Question 23

Marcus et al. (1997) what else is needed in addition to statistical learning?

Answer 23

Need an abstract rule learning mechanism. Necessary to generalise to new exemplars, which is critical to language learning.

Question 24

Marcus et al. (1999) familiarised to different patterns ABA or ABB and then tested on preferential listening for same or novel structure

Answer 24

Infants show a looking-time preference for novel sentences. Claimed that this is evidence for a rule-learning system that operates alongside statistical learning.

Question 25

2 challenges to language rule-learning system

Answer 25

1. Lack of specificity (not language- or species-specific). A general purpose system instead. 2. No need for a separate system - perhaps set of experience enough for generalisation.

Question 26

Gerken (2006) diagonal AAB (abstract rule) vs. just having the same ending (low-level regularity) to test the basis of generalisation

Answer 26

9 months. Diagonal condition generalised to AAB but column condition only generalised to 'di'. Challenges idea of a dedicated abstract rule system. Instead, learning is an interaction between perceptual salience and consistency of contextual cues (determine generalisation). Single mechanism. Learning accomplished without instruction through structured input exposure.

Question 27

Geambasu et al. replication of Marcus et al. (1999)

Answer 27

Failed to replicate. No difference in looking time between test patterns. Not support hypothesis that infants are able to learn abstract 'algebraic' rules and apply them to novel input.