From sounds and experiments to models
Psycholinguistic experiments
Experiment types
Lexical decision tasks
spoken or written
word played/displayed
yes/no button
'is this your language?'
accuracy and speed measured
Priming
a word seen or heard second time recognised faster
semantic/conceptual priming
a related word seen or heard recognised faster
(nurse primes doctor, etc.)
syntactical / morphological
endings of words
sentence structures
Cross-modal priming
read first, heard second
less effective than single modality
cross-linguistic priming
1st language, 2nd language
less than single language
Gating
gradually increase duration of a word
determine when word is recognised
hear 1st fragment
make guess/state confidence
hear 2nd fragment
- we don't need to hear whole words
- reverse gating shows start more important than end
- neighbourhood effects - similar words near, more info needed
Shadowing
listen and repeat as fast as possible
- 25% of women close shadowers
250-300ms - 75% women + all men distant shadowers
500ms - time taken to produce syllable = 150ms
- subjects typically correct errors
- shows don't just repeat but process
- used with dichotic listening - repeat only one stream
understanding barrier to copying
Click detection
- played speech with click inserted
- clicks mask speech sounds
- asked where heard click
- typically report clicks in different places
- believe heard sounds were masked (vice versa)
- phonemic restoration
- sometimes coughs
- Words recognised faster than non-words
- impossible nonwords excluded faster than possible non-words
- accuracy vs. speed - inverse correlation
general findings
- word frequency effects
- more frequent words recognised faster
- word supremacy effect
- actual words faster than non-words
- context effect
- words recognised in context (primed)
- distortion effects
- location of distortion matters
- beginning more damaging
- in written, ends and beginnings important
- clause boundaries have special status
- phrase boundaries less so
- we process speech very quickly
- including corrections, and at same time as speech
Models
a theory of how the mind does certain things
sometimes designed to be used in computers to test
top down or bottom up
logical assumption
- speech understanding bottom up (sounds processed)
- Broca
- speech production top down
- Wernicke
- Wernicke
(Wernicke seems to be activated first)
Modular
language as a module
separate modules within language module
non-modular
fully connectionist
serial modular
module works
send output to next module
module stops working
interactive modular
module can work with previous module
McClelland & Rumelhart
Phonology
Speech sounds = phones
depends on: position in a word, accent, etc.
some variation leads to different meaning
representation of speech sounds (in minds) = phonemes
abstract mental representations
Phonemes
the phoneme /t/ can be
top
stop
butter
in English substitution doesn't alter meaning
babies
- at 1 mo can distinguish phonemes
- prefer own name, mothers voice, etc.
- at 4mo can distinguish phonemes from other languages
- parents can't
- 6-8 mo no longer distinguish vowels from other languages
- 8-12 mo no longer distinguish consonants from other lanugages
problems recognising words
- speed
- english 20-30 phonemes per second
- non-speech sounds at 1.5 per second
- 160-200 wpm (more inf rench)
- speaker variation
- variations in pronounciation (regional, formality, emotion, mode e.g. conversation or reading
- no two utterances are identical
- segmentation problem
- continuous stream of sounds
- no predictable pauses between words
- catgorical perception
categorical perception
- perception of different thigns as the same category
- different pronounciations of /k/ as all being k
- why lack of invariance not a problem
- not limited to speech
- not limited to humans
impoverished signal
- vowels perceived better than Cs
- stronger/louder
- stop consonants in vowels
highly redundant speech
beeb can be diff from bab
- vowels duration & coarticulation important
- lip shapes for koo kee kah
acoustic cues
place of articulation
experiment
listen for errors
errors noticed
- more at beginnings
- in predictable sequences more than unpredictable
- more in certain types of sound switches
Word Recognition
Need models that account for these
Need models that account for:
- word frequency effects
- (im)possible word effect #
- priming effect
- semantic priming effect
- beginnings of words more important#
- more than one word at a time is considered
- blends
- shadowing
- slips of the ear (hear something incorrectly)
McGurk effect
visual input helps decide sound