CATEGORICAL VS. EPISODIC MEMORY FOR PITCH ACCENTS IN ENGLISH by - - PowerPoint PPT Presentation

CATEGORICAL VS. EPISODIC MEMORY FOR PITCH ACCENTS IN ENGLISH

Presented by Ben Posner for the Seminar Exemplar Theorie by Prof. Dr. Bernd Möbius

by Amelia E. Kimball, Jennifer Cole, Gery Dell & Stefanie Shattuck-Hufnagel

Categorical Memory vs Episodic Memory

Categorical Features

• Stemming from Formal Phonology
• Set of abstract representations
• Governed by rules and constraints
• Models variable error-laden input and output
• Remembering is encoding and storing abstract representation
• Phonological abstraction

What is Stored in Memory

• Each phonetic instantiation of a sound unit is mapped to an abstract

category

• Only the category of a unit is committed to memory
• Different sounds of the same category are indistinguishable in memory

Predictions?

• Two distinct sounds mapped to the same category will be reported as the

same

• Only category is retrievable

Episodic Memory

• Exemplar Models
• Encode and remember subcategorical phonetic detail

What is Stored in Memory

• Episodic memories with all details
• All perceived acoustic information (to some extend)
• Within Episodes even 'irrelevant' information is stored
• Irrelevant to language understanding

Prediction

• Within category differences will be remembered
• Even subtle differences can be perceived and later recalled

Evidence

• Phonological Abstraction
• Exemplar Models

Evidence

• Phonological Abstraction
• Exemplar Models

Phonological Abstraction

1: Goldinger, 2007

• Perception task
• Categorical perception of phonemes

Method

Goldinger, 2007

• Two sounds that may or may not vary in voice onset time
• Participants often unable to detect differences when both sounds are

within category

• Short-lag memory task instead of perception task(Remembering the first

sound for a brief amount of time)

• Perception task vs. Short lag memory task

Results

Goldinger, 2007

• Participants quite bad at the task
• Evidence that only category is encoded and retrievable

2: Stress deafness

Dupoux, Knaus, Orzechowska, Weise (2008)

• Speakers of languages with fixed lexical stress
• Participants do not hear a difference in unfamiliar words with differing

stress patterns

• In EEG studies differences are perceived but can't be recalled
• Evidence that listeners do not remember irrelevant acoustic detail even

though it is perceived

Evidence

• Phonological Abstraction
• Exemplar Models

Exemplar Models

(1) McMurrey et al., 2002

• Eye-Fixation Study
• Visual World Paradigm

Visual World Paradigm

• What
• Relies on Cooper (1974), Tanenhaus et al. (1995)
• Gaze toward objects in the real world or a visual representation is measured during speech

production/perception

• How
• Eye Tracking in a controlled visual workspace
• Why
• Natural, Unintrusive
• Can be used on people who cannot write/read
• Can measure online processing

(1) McMurrey et al., 2002

McMurrey et al., 2002

• Input Signal: Beach vs Peach
• Responses: Categorical
• But: Eye Fixations driven by subcategorical variation
• Evidence that listeners are sensitive to within-category differences

Toscano et al., 2010

"At perceptual levels, acoustic information is encoded continuously, independent of phonological information"

Exemplar Models

2: Goldinger, 1996 & Pufal et al., 2014

• Word recognition memory
• Up to one week between tests
• Task: Identifying whether a word was heard before
• Easier with identical background noise (even unrelated noise, e.g. barking

dogs)

• Evidence for episodic memories of speech that include acoustic detail of

within-category variation and linguistically irrelevant variation

Contradictory Evidence

• For categorical Perception
• Short lag memory/Voice onset differences
• Stress deafness
• For subcategorical Phonetic Details
• Recognition Memory Tasks
• Priming Tasks
• Do listeners encode subcategorical detail for speech?
• → This study

Intonational Pitch Accent

• What is that?
• Creating accents in words by shifting the pitch instead of other stress

patterns (length, loudness, etc.)

• In English pitch is part of stress patterns
• In Japanese pitch is the sole form of stressing and distinguishes

meaning

• Sake ↓↗ (酒) = Alcoholic Beverage vs. Sake ↑↘︎ (鮭) = Salmon

Why Pitch Accents?

3 Reasons

Why Pitch Accents?

3 Reasons

• Don't mark lexical contrasts
• Lexical contrasts are categorical
• Mark Information status distinctions related to focus and accessibility
• Focus and Accessibility are gradient

Why Pitch Accents?

3 Reasons

• Might not be categorically perceived
• Has been looked into but there is no strong evidence

Why Pitch Accents?

3 Reasons

• Correlation between Intensity, duration and/or f0
• But: No evidence which of these properties listeners pay attention in

perceiving and interpreting pitch accents

2 Types of Variation

• Phonological Variation (Presence vs. absence)
• Variation in Phonetic Cues to Pitch accent (Duration, F0 peak values)
• If pitch accent is encoded and remembered listeners will be sensitive to

variation in status variation and phonetic cues

• If not only accent status will matter

Study

• 6 Experiments
• 2 sets of 3 Experiments
• Listeners hearing two speech samples after delay or interference
• Listeners have to judge whether both samples were identical
• Different samples can differ in accent categorically (status) or with sub-

categorical change in accent cues

Stimuli

• American English
• Mostly voiced
• "Beavers love building"
• Twelve nouns from six sentences (E.g. Beavers and building from above)
• Each sentence recorded with four accent patterns
• 1&2, 1&¬2, ¬1&2, ¬1&¬2
• Accented words were synthesised to create a bigger difference (still

within original category)

Stimuli

• Pitch corrected in Praat
• Pitch peak corrected 25 Hz up or down
• Duration of word corrected by up to 10%

(using PSOLA in Praat)

• Differences are detected at the same rate as

presence/absence of pitch accent

Participants

• 193 total participants
• native English speakers from the USA
• Age 19-59 (mean of 31, standard deviation of 8.4)
• Excluded participants who did not finish or were bilingual
• 30 participants per experiment

Procedure

• Amazon Mechanical Turk
• 6 different Experiments

Experiment 1, 2 and 3

AX Tasks

• Two words with one second of silence between
• Asked to answer whether recordings were "the exact same recording or

different recordings"

• Recordings were either the same (1/2 of trials) or differed in one of three

ways

Experiment 1

• Variation in Accent Status
• Naturally produces accented recording and
• Naturally produced unaccented recording of the same word

Experiment 2

• Shortened and Lengthened Version of the Same Recording
• Artificially shortened version against artificially lengthened version
• Both resynthesised
• No distinction between a natural vs. a resynthesised sample

Experiment 3

• Lowered pitch peak and raised pitch peak
• Artificially lowered pitch peak against raised pitch peak
• Both resynthesised
• No distinction between a natural vs. a resynthesised sample

Experiment 4, 5 & 6

• Same as 1, 2 & 3 with delay and interference instead of silence
• Four different words (exposure phase)
• Followed by tone
• Word from exposure phase (Test token)
• Question whether test token was exactly the same as the exposure version
• More difficult
• Interference from other words, time delay, and Increased working

memory load

Results

• For AX task:
• Well above chance for all three contrasts
• Experiment 1 = 77%
• Experiment 2 = 85%
• Experiment 3 = 75%
• Comparable performance between status difference and duration/pitch

changes

• All three differences are equally easy to differentiate

Results

• With delay and interference
• Still above chance
• Comparable for Experiment 1 and 4 (Status)
• Worse between 2, 3 and 5, 6 (duration and pitch)

Results

Accent Duration Pitch AX (Exp 1, 2, 3) 77 % 85 % 75 % Delay (Exp 4, 5, 6) 83 % 67 % 54 %

Worse

What does that mean?

In Summary

• Accent difference (status) is remembered after time lag and in presence
• f interference
• Duration and pitch differences are detectable above chance, less

accurately remembered

What does that mean?

• Group effects hold when analysed with mixed effect logit model with

random slopes and intercepts to accound for individual variability

• Individual performance in AX pitch task significantly higher than SD of

scored in AX duration task

• → more variation from listener to listener in pitch task than in duration or

accent task

• holds despite excellent discrimiation (of pure tone differences of the

same magnitude pitch in post-tests)

Discussion

• Pitch accent status and phonetic cues for pitch accents are perceived,

encoded and available for immediate access

• But: After delay or with interference some pitch and duration information

is lost.

• Consistent with hypothesis that detailed instances of pitch accents are

encoded, but phonetic detail related to pitch accent quickly becomes less accessible in memory compared to categorial accent distinction

Discussion

• Listeners vary in performance especially for pitch manipulation
• Surprising, tested differences well above "just noticeable"
• Evidence that listeners at a group are bad at the task but some are good
• No explanation why some are better

Conclusion/Summary

• Listeners encode both categorical distinctions and phonetic detail in

memory

• But: Categorical distinctions are easier accessible for retrieval in explicit

judgement tasks and listeners may vary in the degree to which they remember/can access prosodic detail

• Evidence suggest that both premises can exist together
• All acoustic information is encoded (episodic/exemplar) but not all

information is retained or accessible later (categorical/abstraction)

Thank you for listening!

Sources

• https://likan.info/en/paradigm/language/visual-world-paradigm/
• https://www.youtube.com/watch?v=dyQs27Rycaw