Cross-linguistic Effects in the
Perception of Assimilated Speech
Lexical access in speech perception
• Rapid identification: 250-500 ms
• Large lexicon: 50,000+ words
• Crowded space
– e.g., “cat” has over 25 neighbours differing on the
identity of a single phoneme (e.g., rat, cut, cap)
• So, little room for error
– Even subphonemic manipulations can affect the
activation of a word (Andruski, Blumstein & Burton,
1993; Davis, Marslen-Wilson & Gaskell, 2002)
Phonological variation
• Problem of crowded space compounded by natural
variation in surface form of speech
• Focus in particular on assimilation of place of
articulation:
– Changes place of articulation of word-final coronal
consonants to become more like following context
• lean bacon  leem bacon
• dress shop  dresh shop
– Crosses word boundaries
– Can be a continuum of assimilation
• full assimilation creates strongest ambiguity
Contextual Viability
• Perceptual system makes use of contextual
viability in determining likely underlying
sequence
– e.g., Gaskell & Marslen-Wilson (1996)
• “Leam bacon” perceived as “Lean bacon” [shared POA]
• But “Leam gammon” causes mismatch in perception
[mismatching POA]
– Compensation effect observed in a wide range of
circumstances (e.g., Gow, 2001; Mitterer & Blomert, 2003)
Accounts of compensation
• Models that deal with context effect differ in terms of
their reliance on statistical learning
– Gow (2002, 2003)
•
•
•
•
separate out cues to assimilated consonant using feature parsing
works best for incomplete assimilation
similar perceptual account given by Mitterer & Blomert (2003)
applies to assimilations irrespective of language background
– Gaskell (2003)
• learn circumstances of assimilation from statistics of language and
compensate accordingly
• applies to both complete and incomplete assimilation
• predicts cross-linguistic differences (statistics are tailored to native
language)
Data on cross-linguistic effects
• Some studies found little or no effect of native language on
perception of assimilation
– Gow & Im (2004) – Hungarian/Korean/English
– Mitterer et al. (2006) – Hungarian/Dutch
• short sequences, simple perceptual tasks
• Others found clear effects of language background
– Darcy et al., (2007, in press) – French/English
• longer sentential context, word detection tasks
• Key factor may be the degree to which the full range of
utterance processes can be engaged
– But with broader sentence context, it becomes harder to control all
aspects of stimuli for two sets of speakers
• Goal of current research: maintain tight control of stimuli while
engaging sentential level processing
General Methodology
• Speakers of two languages learn an artificial lexicon
(words refer to abstract objects)
– equates lexical knowledge
• During testing, embed same two-word sequences
varying on degree of assimilation into sentence
contexts using the speaker’s native language
– equates sentential context and phonetic properties
• Examine lexical preferences using the visual world eyetracking paradigm
– look of evidence of cross-linguistic differences in highly
comparable circumstances
Choice of assimilation phenomenon
• Need assimilation that:
– shows significant differences in extent or conditions across
two languages
– involves consonants that are phonetically similar in the two
languages
• Selected sibilant assimilation in English and French (cf.
Holst & Nolan, 1996)
– standard view is that English shows strong regressive
alveolar to postalveolar assimilation (e.g., dress shop →
dresh shop)
– whereas such assimilation is absent in French
Phonetic Study
Niehbuhr, Clayards, Meunier, Lancia (in revision)
4 speakers of both languages
Measured spectral centre of gravity (CoG) and duration
ENGLISH
FRENCH
Regressive
“See how the glass shines”
STRONG
“C’etait une classe chargée”
WEAK
Progressive
“She likes the British south”
ABSENT
“Tu te cache sous le lit”
WEAK
Type
s  sh
s  sh
Perceptual study - predictions
Language specific contextual compensation should develop in cases
where complete assimilation causes ambiguity
Predict cross-linguistic differences in regressive (following
context) but not progressive (preceding context) assimilation
ENGLISH
FRENCH
Regressive
“See how the glass shines”
STRONG
“C’etait une classe chargée”
WEAK
Progressive
“She likes the British south”
ABSENT
“Tu te cache sous le lit”
WEAK
Type
s  sh
s  sh
Experimental Design
Equate Stimuli
•
Artificial lexicon
•
•
1 French and 1 English native speaker
Listeners hear both speakers in both languages
Compare complete and partial assimilation
•
cross-spliced intermediate CoG as well as endpoints
/ʃ/
/s/
1 2 3 4 5 6 7
Compare following and preceding context
Experimental Design
objects buttons
cavees
caveesh
pidas
tamash
nalip
remop
Rendezthe
le nalip
Render
nalip
shinnow
vous
shinnows’il
please
plait
samal
shamal
sival
shinnow
pagoon
pentuf
Word 1 Word 2
Experimental Design
Task from Pirog Revill et. al 2008
Render the nalip
shinnow please
Experimental Design
Testing
objects first
Render the cavees pagoon please
Following context
buttons first
Render the nalip samal please
Preceding context
Training
Inclusion criteria
• Participants with more than 25% errors on
endpoints for control condition excluded
– Both groups of listeners did better with the French
speaker
speaker
English French total
English 13
20
25
French 13
22
26
Total
43
27
buttons
listener
listener
objects
speaker
English French total
English 18
20
25
French 15
22
26
Total
42
33
– Main analyses run on just data for French speaker
(any bias goes against key prediction)
Mouse click responses (following context)
/ʃ/
“Render the cavee...
English listeners
1
2
3
4
5
6
7
/s/
control
shinnow”
sival”
French listeners
Stats (logistic mixed effects model):
Effect of test context (s vs sh)*
French speaker
only
Interaction
with listener group***
Mouse click responses (preceding context)
/ʃ/
control
“Render the tamash
pidas
English listeners
1
2
3
4
5
6
7
/s/
...amal”
French listeners
Stats (logistic mixed effects model):
Effect of test context (s vs sh)***
French speaker
only
Interaction
with listener group (ns)
Time-course of effects (regressive)
English listeners
Occulo-motor
delay (200 ms)
This spectrogram illustrates the control
condition (1 sibilant, 120 ms long)
step
step x following
context
following context
Time bin
300400
400500
500600
600700
700800
800900
*
*
*
*
*
*
following context
*
*
*
step x following context
*
step
200300
*
*
Time-course of effects (progressive)
French listeners
Occulo-motor
delay (200 ms)
This spectrogram illustrates one of the test
conditions (2 sibilants, 200ms long)
preceding context
step
Time bin
200300
step
preceding context
*
300400
400500
500600
600700
700800
800900
*
*
*
*
*
*
*
*
Conclusions
• French and English listeners differ in their use of following
context in regressive assimilation
– English listeners show contextual viability effects across a broad
continuum, including complete assimilations
– French listeners show little or no viability effect
• In the progressive case, both sets of listeners use sibilant cues
contrastively, with no cross-linguistic difference
• These data suggest that listeners adapt to the ambiguities
typical of their language
– Complete assimilation in production leads to compensatory perceptual
effects
– In other cases, generic feature parsing/compensation for coarticulation
applies
Thanks to...
• Other members of the Marie Curie Sound to
Sense group, especially:
– Uli Frauenfelder
– Sarah Hawkins
– Christine Meunier
– Noel Nguyen
• Eyetrackers:
– Gerry Altmann and Dirk Kerzel
Phonetic results
English
Niehbuhr, Clayards, Meunier, Lancia (in revision)
French
Descargar

Slide 1