An Adaptive, Dynamical Model of Linguistic Rhythm

1
An Adaptive, Dynamical Model of Linguistic Rhythm

• Sean McLennan
• GLM 040312

2
Underlying Intuitions

• Somewhere between the signal and low level speech
  recognition, linguistic time is imposed upon real
  time.
• Linguistic time is more relevant to speech
  recognition than real time.
• Not all segments are created equal - certain
  points / intervals in the speech stream are more
  important for recognition than others.

3
What Rhythm Is and Is Not

• Rhythm - historically based primarily on the
  perception that different languages are
  temporally organized differently
• Three recognized rhythmic types stress-timed
  (English), syllable-timed (French), and
  mora-timed (Japanese)
• Rhythm implies underlying isochrony which turns
  out to be absent (ex. Dauer, 1983)

4
Recent Views of Rhythm

• Ramus and colleagues
• examined three factors V ?V ?C
• V proportion of vocalic intervals in the
  signal
• ?V variation of length of vocalic intervals
• ?C variation of length of consonantal
  intervals

5
Recent Views of Rhythm
6
Recent Views of Rhythm
7
Recent Views of Rhythm
8
Rhythm and Segmentation

• Cutler and Colleagues
• study the question of how rhythm type impacts on
  the segmentation of words from the speech stream
• implication being that a naïve listener (i.e. an
  infant) uses rhythm as a bootstrap for early
  stages of acquisition

9
Rhythm and Segmentation

• Syllable Effect
• French speakers spot ba- in balance faster than
  in balcon
• French speakers spot bal- in balcon faster than
  in balance
• rigorously reproduced, even on non-French words
• stubbornly absent in English

10
Rhythm and Segmentation

• Stress Effect
• Native English speakers find mint faster in
  mintesh than in mintayve
• Native English speakers find mint slower in
  mintayf than in mintef and thin in thintayf
  or thintef.
• In missegmentations - tend to insert before a
  stressed syllable (in vests) or delete before a
  weak syllable (bird in)

11
Rhythm and Segmentation

• Mora Effect
• Native Japanese speakers find ta- in tanishi
  faster than in tanshi
• Native Japanese speakers find tan- faster in
  tanshi than in tanishi.
• Native Japanese speakers can find uni in
  gyanuni and gyaouni but fail to find it in
  gyabuni.
• Native English speakers have no problem with the
  Japanese task
• Native French speakers show the same cross-over
  effect with the Japanese task as in French and
  English

12
The Proposed Model

• hopefully a bridge between Cutler et al and Ramus
  et al - why should V ?V ?C impact on
  segmentation?
• can a naïve adaptive model responsive to V ?V
  and ?C produce behavior consistent with
  segmentation based on rhythm-type?

13
The Proposed Model

• V ?V and ?C need two points to be consistently
  tracked vocalic onsets and offsets

14
The Proposed Model

• Use these spikes to drive two adaptive
  oscillators (habituating neurons?)
• Unlikely to entrain but will make predictions

15
The Proposed Model

• The accuracy of prediction will be a measure of
  ?C and ?V
• Difference in the period will be a measure of V

16
The Proposed Model

• ?C ?V and V in turn determine the size of an
  attentional window
• the attentional window is a metaphor for stimulus
  decay
• The smaller ?C and ?V and closer V is to 50,
  the more periodic the rhythm, the narrower the
  window can be
• The larger ?C and ?V and more divergent V is
  from 50, the less periodic the rhythm, the wider
  the window must be

17
The Proposed Model

• Attentional window size (hopefully) would
  correlate with rhythm type and would predict
  different types of segmentation / recognition

18
The Proposed Model

• Predictions, questions, and other benefits
• consistent with the correlation between rhythmic
  type and consonant cluster complexity
• consistent with ambisyllabicity
• perhaps attractor states predict categorical
  differences
• suggests manner in which to manipulate tasks to
  force effects
• single language-independent mechanism