L2 - An introduction to computational models of speech perception 8th October

Question 1

What is bottom up processing?

Answer 1

Sensory input –> building up to semantic understanding

Question 2

What is top down processing?

Answer 2

Semantic knowledge –> sensory input

Question 3

What happens when people speak to us?

Answer 3

We access it in the mental lexicon

Question 4

What are the five challenges to lexical access?

Answer 4

Continuous speech stream
Homonyms and homophones
Co-articulation
Different accents
Invariance problem

Question 5

What is the continuous speech stream?

Answer 5

We don’t have staggered speech despite having individuals words - we make one long continuous production of sound - this is one of the challenges of speech as to how we separate

Question 6

What is co-articulation?

Answer 6

When speech production is influenced by sounds that proceed and follow a phoneme - ie thin book requires a different pronunciation to thin carpet as the ‘n’ sounds different

Question 7

What are the three ways to disambiguate the speech stream?

Answer 7

Categorical perception
Perceptual learning
Top-down processing

Question 8

What is categorical perception?

Answer 8

The ability to distinguish between sounds on a continuum based on Voice Onset Times eg Va vs Fa

Question 9

What is perceptual learning?

Answer 9

We become accustomed to the different VOTs of our language

Question 10

What is top down processing?

Answer 10

We use what we should be expecting to hear to disambiguate the speech stream

Question 11

What does spreading activation do?

Answer 11

Facilitates predictions of what may be coming up next via activation of items related to the acoustic input

Question 12

What are the three lexical characteristics that affect speed of lexical access?

Answer 12

Word length
Neighbourhood density
Frequency

Question 13

Describe how word length affects speed of lexical access

Answer 13

Long words are slower to process and therefore longer to access

Question 14

Describe how neighbourhood density affects lexical access

Answer 14

Fast access when less than ten neighbours (spring fruit choir), and slower when more

Question 15

Describe how frequency affects speed of lexical access

Answer 15

The more frequently a word is accessed in the lexicon the quicker you can access it

Question 16

What are the two models of speech perception?

Answer 16

Marslen-Wilson - The Cohot Model
Elman and McClelland - The TRACE model

Question 17

What does the Cohort Model predict?

Answer 17

That we access words in the lexicon via activation of all words sharing initial features and gradually de-activate words that stop matching the features

Question 18

What does the TRACE model predict ?

Answer 18

The TRACE model suggests that as we hear sounds (features), they gradually activate related speech sounds (phonemes) in the brain. These phonemes then activate possible words that match them. Over time, the word that best matches all the sounds becomes the most strongly activated and is “chosen” as the recognized word.

Question 19

What is the uniqueness point?

Answer 19

A point in lexical access when all the other words that do not match the input have been deactivated - crucial component of the COHORT model

Question 20

What type of processing does the Cohort model use?

Answer 20

Bottom up processing - we start of with sensory input and go from there,

Question 21

Describe the word prickly when trying to activate the word apricot

Answer 21

Items that do not match the onset of the word are not activated

Question 22

What is the neighbourhood effect?

Answer 22

Words that match the acoustic input compete for activation - they compete for recognition. Ie, if Aprikol was a word, it would become a competitor for Apricot so they compete for recognition so need more information before it can be deactivated - learning aprikol will slow down recognition of the word apricot and slows down the uniqueness point.

Question 23

Describe the relationship between activation and frequency

Answer 23

Words with high frequency require less activation to be recognised
Ie Aprikol lacks frequency, more difficult to access

Question 24

Describe Warren&Marslen-Wilson’s Gating experiments

Answer 24

Participants are presented with fragments of words that gradually reveal the whole word and asked to guess what the word is after each presentation

Evidence in favour of a Cohort
Gating experiments suggests that
- recognition of a word is a gradual process that starts from word onset and continues until the end of the word
- Candidate words that no longer fit the acoustic input are eliminated

Question 25

Describe the diagram of the Cohort model

Answer 25

Lexical item - Words
Facilitatory signals are sent to words that match the speech input and Inhibitory signals are sent to words that do not match the speech input
Speech input (signals sent from here)
All with Bottom-up processing

Question 26

Describe sentence context and the cohort model

Answer 26

Sentence context does not influence the process of lexical access

Question 27

What is lexical selection based on in the Cohort model?

Answer 27

Activation of phonology and semantic information

Question 28

What did early iterations of the Cohort model suggest?

Answer 28

Context constrained the cohort

Question 29

What are the 3 stages to word recognition in the Cohort model?

Answer 29

1) Access - acoustic-phonetic information is mapped onto lexical items
2) Selection - candidate words that mismatch the acoustic output are de-selected - candidate word is chosen
3) Integration - Semantic and syntactic properties of the word are integrated and checked against the sentence

Question 30

Describe the priming paradigm with doctor and nurse

Answer 30

In a lexical decision task
If the word doctor was shown with a bunch of words and non words, and u were shown the word nurse, you are more likely, due to having a prime with the word doctor, which is a related semantic concept, to access the word quickly and select that this is a word.

Question 31

Cross modal priming

Answer 31

Receiving prime word through speech (through headphones) and a visual target word
More likely to access ship once heard captain. Related prime-target pair

Question 32

Cross modal priming - word fragment

Answer 32

Hearing the word cap..
Could be captain or capital
Under the cohort model - cap should still prime you for ship through spreading activation

The word capital should prime you to respond to money - should still be in the initial cohort.

Question 33

Describe what Zwisterlood found when measuring word fragment reaction times with ship, money, and wicket

Answer 33

Zwitserlood 1989 measured reaction times to respond to money, ship, wicket.
They are faster to respond to the word ship and money but not wicket when the word cap is presented. Indicates that the cohort is active and semantic activation is active.

Question 34

What is the impact of context biassing on priming word fragments? COHORT

Answer 34

The participants were presented with an auditory presentation of the sentence ‘The men around the grave mourned the loss of their cap___’
Then presented either the word ship, money, or wicket
Because of the contextual biassing constraining the cohort, only one word should make sense.
What we should see is quicker lexical access for ship - context should bias.
But what we actually see is that quicker lexical access for both ship and money - context does not constrain the cohort. Both concepts are activated. Contextual biassing has not worked.
THE ONLY TIME biassing works is when the WHOLE word is presented - captain.

Question 35

How are word recognised in the TRACE model?

Answer 35

In TRACE, the words are recognised incrementally by slowly ramping up the activation of the correct units at the phoneme and word levels

Slowly build up the words that actively match - we do not start with a big group and deselect

Question 36

Give examples of gradual activation of item that matches the input

Answer 36

AP -
apex
april
apricot
apart
apple
shape
apricot
tape

Question 37

What are the key elements of the TRACE model?

Answer 37

Features - voicing, manner of production
Phonemes
Words

Question 38

Describe a structural diagram of the TRACE model

Answer 38

Lexical item - words
facilitatory connection
Phonemes
facilitatory connection
Features - acoustic phonetic patterns
BOTTOM UP PROCESSING

Activation spreads up from features to lexical items
Each level is connected via facilitatory connections travelling up AND travelling down
Connection between nodes within each level are inhibitory to stop wrong information being passed between levels

Question 39

Describe what would happen if we were trying to perceive the word van

Answer 39

We would start with feature information for example telling us that the first sound is voiced
This would activate phonemes that correspond to that feature and are voiced such as v b and d
Selection of V would send an inhibitory signal to the other phonemes
The remaining phonemes in the word would be activated and in turn activate words that matched the input
Increased activation to van as the matching item would send inhibitory signals to the other words that had become activated
Top down processing increases activation of phonemes and features

Question 40

What does the radical activation model suggest?

Answer 40

Any consistency between input and representation may result in some degree of activation

Question 41

Describe Allopenna et al’s study for activation of words in the lexicon

Answer 41

Using an eye tracking study demonstrated that words with overlapping phonology that do not start with the same onset as the speech input (rhyme competitors) are activated in speech perception

Participants are presented with a grid that contains images of items such as this grid with a beaker, a beetle, a speaker, and a pram
Participants are asked to click on the beaker and place it under the triangle (multiple shapes on the grid
PPs eye movements are monitored whilst they complete the task

Participants looked at the beaker, the beetle and the speaker - would not be seen in the cohort model - if words are related to beaker are active in the lexicon participants will look towards those items

Evidence from Allopenna ea and any other words suggest that words that rhyme with sounds in any part of a word may become activated.

Question 42

What did Mirman et al research and find?

Answer 42

Participants were asked to detect a t or k in words and non words
Faster identification of t and k in words
Demonstrating the effect of top-down processing
Evidence for the TRACE model