lecture 5 - speech production Flashcards
tongue twisters are tricky because:
rapid sequence of overlapping neural patterns
strategies for fluent speech (to reduce processing demands while planning what to say):
1) preformulation = producing phrases used before (accounts for 70% of speech)
2) underspecification = full meaning not explicit e.g. ‘something, ‘things’, ‘them’
3) repetition
spreading activation theory (Dell): 4 levels to speech production
1) semantic level = planning level
2) syntactic level = grammatical structure of words (planning)
3) morphological level = morphemes (basic units of meaning)
4) phonological level = phonemes (basic units of sound)
SSMP
mild cog impairment - how to monitor disease progression
monitor reduction in semantic content
how can we account for mixed findings?
flexibility: speakers are flexible and depends on their immediate goals and situational demands
spoonerisms
when initial letters if 2 words are switched - phoneme exchange error
Freudian slips
speech errors products of unconscious/repressed thoughts
limitation: where Freud stressed intrusion of thoughts from outside language system, research suggests most errors reflect inner workings of language system
semantic substitution errors
when correct word is replaced by word of similar meaning - suggests we plan grammatical structure (syntactic level) before finding precise words to fit that structure
morpheme exchange errors
inflections or suffixes remaining in place but attached to wrong words (e.g. I randomed some samply’) - word stems seem to be planned before inflections are added (so processed separately - inflections/suffixes are altered to fit with new word stems)
number-agreement errors
singular verbs are mistakenly used with plural subjects & vice versa (e.g. the government ‘have’ made a mess, instead of ‘has)
cause of errors
misapplications for combining one type of unit with another - misapplication occurs within a given level (so similar units exchanged)
speech errors demonstrate that an utterance is planned
Support for this idea comes from anticipation errors: a particular unit (e.g., a phoneme) is activated as the utterance is being planned, but is produced too early
e.g. “leading list” - the phoneme “l” is produced earlier than intended (and competes with and ultimately replaces the “r”)
speech dysfluencies examples
breaks and irregularities within the flow e.g. false starts, restarted/repeated phrases, repeated syllables, fillers ‘uh’, ‘well’, repaired utterances (correcting their slips)
error detection: Perceptual loop theory (Levelt)
speakers detect own errors by listening to themselves and discovering what they said differed from what they intended (same when detecting others’ speech with their comprehension system)
error detection is slow
error detection: Conflict-based theoretical account (Nozari)
error detection relies on info generated by the speech-production system - conflict monitoring during competition among possible words at time of response
error detection is faster: monitor inner speech
evidence is more supportive of detecting errors by conflict monitoring:
1) success of brain-damaged patients in detecting own speech errors depends more on their speech-production ability than their comprehension ability
limitations of research in this area:
1) hard to carry out systematic research when occurrence of speech erros and corrections are unpredictable
2) processes involved occur rapidly
theories of speech production:
Levelt’s theoretical approach & WEAVER (1989)
computational model: WEAVER++ (Word-form Encoding by Activation & VERification)
focuses on processes involved in producing individual spoken words
WEAVER assumptions: 3 main levels
1) feed-forward activation-spreading network - from meaning to sound (highest-lowest)
2) 3 levels:
highest level = nodes representing lexical concepts
second level = nodes representing lemmas (abstract words) from the mental lexicon (word representations specified syntactically & semantically but phonologicaly)
lowest level = nodes representing morphemes & their phonemic (sound-based) segments
speech errors are avoided by monitoring speech
lexical selection depends on (WEAVER):
competitve process based on number of lexical units activated - lemma is selected when it is activated more than others
morphological encoding & phonological encoding (WEAVER)
1) basic word form of selected lemma is activated
2) word’s syllables are computed
therefore lexicalisation happens = we turn thoughts underlying words into sounds
processing is discrete (separate)
Parallel activation will lead to ________ errors
anticipation errors (i.e. word exchange errors and spoonerisms)
we can see distinction between lemma and word in the tip-of-the-tongue state
occurs when semantic processing is successful but phonological processing is not
(lemma has been activated but the word form (lexeme) cannot be accessed)
have a concept in mind but can’t find the word
Weaknesses of WEAVER
cannot account for the finding that info about the sounds of words also influence which words speakers produce or that semantic word substitutions involve words that are more phonologically similar than is expected by chance
Advanatge of model
1) can predict speed of word production
2) focuses on error-free speech - makes sense
Ferreira & Griffin (2003)
incomplete sentences shown followed by unrelated pic
had to name the pic (not complete the sentence)
primed sentence completions that were semantically related to pic were substituted for picture names more often than unrelated completions (e.g. nun for priest)
primed sentence completions that were homopones of semantic competitors also substituted for pic names (e.g. none for priest)
- shows that semantic AND phonological info can jointly influence lexical selection (supports interactive theories)
spreading activation model (Dell)
nodes (correspond to words or concepts) vary in activation
when a node/word is activated, activation spreads to to other related nodes
can also account for sounds
spreading activation model: 4 levels
speech planning happens at same time at all 4 levels
1) semantic level = meaning/abstract representation
2) syntactic level = grammatical structure of words
3) morphological level = morphemes
4) phonological level = phonemes (pronunication info)
SSMP
categorical rules at each level
a lexicon in the form of a connectionist network - nodes for words, morphemes and phonemes
insertion rules (spreading activation model):
1) most activated node belonging to most appropriate category is chosen
2) after selection, its activation level reduces to zero - prevents it from being selected repeatedly
weaknesses of spreading activation model
1) focuses on indvidual words
2) focused on errors
3) can’t say much about time take to produce words
advantages of spreading activation model
1) predicts nature and number of speech errors