Speech Perception & Comprehension Flashcards
1
Q
SPEECH = VARIABLE
A
- every word takes dif acoustic shape each time it’s uttered; due to:
1) speaker (vocal track size/regional accent/socio-economical tier)
2) articulation rate (4/5 syllables/sec in sentences)
3) prosody (music of speech ie. rhythm/melody/amplitude)
4) mode (voiced/whispered/creaky)
5) coarticulation (individual phonemes influenced by preceding/upcoming segments ie. regressive/progressive assimilation)
2
Q
VISUALISING SOUND
A
- 2 main ways:
1) WAVEFORM - y-axis represents amplitude (w/0 on horizon); x-axis represents time
2) SPECTROGRAM - derived from Fourier transform to represent time on x-axis
- y-axis = frequency/energy (ie. amplitude)
- colour = 3rd dimension (aka. brighter = stronger)
3
Q
SPEECH = QUASI-CONTINUOUS
A
- no unique/systematic way to flag word boundaries aka. rarely silence between 2 words
- short silences (100ms) typically correspond to vocal tract closing to produce so-called plosive/STOP consonant in “pocket”
4
Q
SPEECH = LEXICALLY AMBIGUOUS
A
- words = made of limited number of sounds/syllables aka. embedded words = everywhere inside other words
- ie. captain -> cap
- ambiguity also arises due to straddling words as soon as we put 2 words together
- ie. clean ocean -> notion
5
Q
SPEECH = AUDIOVISUAL
A
- visual info given by lips/adjacent facial areas about articulation = integral to speech perception when available
6
Q
MCGURK & MCDONAL’S ILLUSION (1976)
A
- visual signal should be weakly constraining for it to work aka. visual /ga/ = ^ ambiguous > visual /ba/
- /ga/ = don’t actually see if speaker is closing glottis
- so visual cues = also compatible w/ /da/
- visual /ba/ = unambiguous as you see lips closing preventing illusion from occurring
- visual signal must be compatible w/both back/medial closure of vocal track (/ga/ VS /da/); conflict w/front closure implied by auditory /ba/ attracts perception towards mid-point between front/back of mouth (/da/)
FUSION - /ga/ (vision) + /ba/ (audition) = /da/ (perception)
7
Q
INFO FOR IDENTIFYING WORDS
A
PHONEMES
SUPRA-PHONEMIC INFO
8
Q
PHONEMES
A
- building blocks of vocab
- smallest units in signal allowing meaning distinction (ie. bat/mat have 3 phonemes & differ by 1st one)
- limited number so words are created by combining them in unlimited ways specific to language
- English = 20 vowels & 24 consonants
9
Q
SUPRA-PHOENEMIC INFO
A
- prosody/music of speech (ie. rhythm/melody/energy) ie:
1) lexical stress/accentuation (ADmiral/admiRAtion)
2) tones (same strong of phonemes can have dif meanings depending on pitch contour in some languages ie. ma in Mandarin (horse/mother/scold)
10
Q
SUPRA-PHOENMIC INFO: DAHAN ET AL. (2001)
A
- carried by larger chunks > phonemes ie. syllables
- languages vary in terms of importance of supra-phonemic info for recognising words (ie. French < English < Mandarin)
- phonemic/prosodic info is needed for lexical distinctions BUT word recognition = also sensitive to subtle articulatory details ie. co-articulation cues
- the way in which vowel is pronounced/sounds depends on identity of following consonant
11
Q
SPEECH = MENTAL CATEGORIES
A
- when presented w/exemplars along continuum of syllables between 2 end-points (ie. gi-ki) we perceive whole continuum section as 1 category (ie. gi) while the other is a separate category (ie. ki) despite physical changes in category
- aka. step-like shift indicating category boundary at some point in continuum
- we experience stimulus as either 1 or other BUT not as in-between aka. categorical perception
- most obvious in consonants (ie. rapid acoustic changes) > vowels/tonal info (steadier/continuous)
12
Q
CATEGORICAL PERCEPTION IN DISCRIMINATION TASKS
A
- can also occur in discrimination tasks
- hearing dif between 2 adjacent exemplars in continuum is maximal at category boundary (ie. across categories) BUT at chance within category
- category boundary lies roughly at location of continuum for all speakers of given language
13
Q
CATEGORICAL PERCEPTION IN CONSTNANT CONTRASTS
A
- cannot be easily demonstrated on all contrasts as you need to identify key parameters involved in contrast & latter must be easily manipulated
- ie. voicing distinction (pa/ba; ga/ka) = regulated by 1 acoustical parameter aka. Voice Onset Time (VOT) corresponding to noisy segment from consonant explosion release up to start of periodicity in vowel
- aka. voiced consonants (b/d/g) = shorter VOT > voiceless counterparts (p/t/k) in English
14
Q
VOICE ONSET TIME (VOT)
A
- can be manipulated to create continuum from voiced consonant to voiceless counterpart (ie. gi VS ki) & see if perception follows progression along continuum linearly VS showing mental categories
- pps asked if 2 stimuli adjacent on continuum = same/dif acoustically -> maximal discrimination occurs at perceptual boundary & would be at chance for all other adjacent comparisons
15
Q
WERKER & TEES (1984)
A
- examined ability of English infants to discriminate non-native (ie. Hindi/Salish) contrasts during 1st year of life
- cross-sectional/longitudinal approaches using conditioned head-turn paradigm
- newborns come to life equipped to deal w/any possible phonetic contrast
- non-native contrasts disappear w/exposure to language BUT native contrasts = maintained
- aka. infants transform language-general phonetic skills -> language-specific phonological abilities via “winnowing” (aka. narrowing down) initial set of “innate” discrimination abilities
16
Q
SPEECH -> WORD MAPPING
A
DIRECTIONALITY OF LEXICAL ACCESS
ACTIVE COMPETITION BETWEEN WORDS
INTERACTIVITY FROM MEMORY -> PERCEPTION (VICE VERSA)
17
Q
DIRECTIONALITY OF LEXICAL ACCESS
A
- auditory memories for words “open up” only if initial sound = perceived
- left -> right processing aka. first few sounds carry most of info weight (word endings = easily “guessed”)
- contrasts w/parallel processing in visual/orthographic modality
- importance of Uniqueness Point
18
Q
ALLOPENNA ET AL. (1998)
A
- used eye fixations to determine which words was being evoked by incoming speech (“and now click on the beaker”)
- more fixations of onset-overlapping words (ie. beetle) > rhyme-overlapping words (ie. speaker)
- so word memories = more easily evoked by initial sounds > endings (ie. directionality)
19
Q
MARSLEN-WILSON & WELSH (1978)
A
THE COHORT MODEL
STEP 1) Activation
- first sound of word activates all words in memory beginning w/said sound aka. cohort
STEP 2) DE-ACTIVATION
- words no longer match signal as it unfolds = progressively rejected from cohort
- initial cohort gets smaller as more info arrives
STEP 3) UNIQUENESS POINT
- aka. word identification just occurred
- info afterwards barely has role in word recognition; system already committed itself
20
Q
COMPETITON BETWEEN LEXICAL CANDIDATES
A
- each phoneme in input can logically only belong to 1 word at a time
- so sound-overlapping memories (ie. succeed/seed) = necessarily enemies
21
Q
INTERACTIVITY
A
- 2 processes (A/B) said to interact if B receives A output as input & able to send result of computations back to A before A = completed
- similar to “Larsen” acoustical effect
- 2 conditions must be met:
1) info must be allowed to travel both ways
2) info should be passed to next lvl immediately pre current lvl finishes own computation (ie. cascading NOT seriality)
22
Q
THE GANONG EFFECT (1980)
A
- lexical content dictating how preceding ambiguous phoneme should be heard = compatible w/notion that perception is no autonomous process but 1 that LT lexical memory influence
- Ganong effect motivated inclusion of top-down connections in TRACE (McClelland & Elman (1986))
- BUT could be that Ganong doesn’t reflect influence of memory on perception per se but simply combination of perception + memory influencing conscious decision 1 needs to take to complete labelling task
23
Q
SPEECH STREAM -> WORD SEGMENTATION
A
LEXICAL SOLUTIONS
PRELEXICAL CUES
24
Q
LEXICAL SOLUTIONS
A
1) word offset anticipation
2) lateral inhibition between word memories
- solution proposed in Cohort model
- segmentation = by-product of word recognition
- lexical boundaries perceived as consequence of recognising words in speech
25
Q
LS: WORD OFFSET ANTICIPATION
A
- useful for words that have their uniqueness point per/on last sound (ie. catheDRALrenovated)
- BUT many cases of short words embedded at start of longer words aka. clearly not ideal (ie. CATerp…/ etic VS illar?)
26
Q
LS: LATERAL INHIBITION BETWEEN WORD MEMORIES
A
- if overlapping words = enemies & adjacent words = friends -> segmentation outcome = optimal:
1) elected words don’t overlap by any of their sound (aka. 1-to-1 mapping only)
2) no phonetic segment is left unaccounted for (aka. exhaustivity) - “shipinquiry” = ONLY “ship” + “inquiry” despite other words fully compatible w/portions of signal (ie. in/ink/shipping)
- BUT ship/inquiry = only 2 words w/which conditions 1 & 2 are met
- some memories could have reaction lvl pushed below baseline via lateral inhibition from overlapping competitors; contrasts w/machine parsing “recognise speech”
27
Q
PRELEXICAL CUES (TRUBETZJOY (1939))
A
- listeners learn that certain properties of language = associated w/presence/absence of word boundaries
- these cues modulate lexical competition to favour certain segmentation outcomes
- cues can be proximal (located at word boundary)/distal (further away)
- 2 types:
1) allophonic cues
2) rhythmic cues
3) phonotactic cues
28
Q
ALLOPHONIC CUES
A
- phonemes take on particular shape/quality depending on position relative to word/syllable boundaries
IE) ENGLISH - voiceless stop consonants = aspirated at syllable/word onset
- vowels can be preceded by glottal stop at word onset (ie. grey t(h)ape VS great (?)ape)
- speakers universally lengthen word-initial sounds (ie. great aaaape VS great tttttape)
- syllables tend to be shorter when part of multisyllabic word compared to being a monosyllabic word in own right
29
Q
RHYTHMIC CUES (CUTLER & BUTTERFIELD (1992))
A
- relate to beat of language & relative syllable weight
IE) ENGLISH - listeners take STRONG syllables as word onsets
- reflects that STRONG-weak = most common stress pattern in said language
30
Q
PHONOTACTIC CUES
A
- each language has own rules on sequencing sounds inside words/syllables & on which sound occupies which position; need not be all or nothing (aka. restrictions) but based on probabilities (ie. position specific frequencies)
IE) ENGLISH - “th” (ie. that) = always followed by vowel aka. must be word boundary right after “th” when consonant follows (ie. “bathe more”)
IE) FINNISH - vowels = either all of same category (all back OR front) or neutral; changing back -> front (vice versa) in running speech signifies word boundary
31
Q
SUMMARY
A
-word recognition = strongly directional; follows unfolding of signal
- lexical competitors = words matching signal from OWN onset; don’t have to be aligned w/word onset
- segmentation can be solved by:
1) reorganising words in signal (lexical solution)
2) learning that some linguistic events correlate w/presence/absence of word boundaries (pre-lexical solution)
- phoneme perception = influenced by words we know (interactivity)
