Theoretical Perspectives to Speech Perception

1
Theoretical Perspectives to Speech Perception

• Mark C. Flynn
• The University of Canterbury

2
Question to be answered

• How are we able to perceive the acoustic speech
  wave and transform it into a linguistically coded
  message in our brain?

3
Basic Problem is

• Human communication is not just a transfer of
  information like two fax machines connected with
  a wire. It is a series of alternating displays of
  behaviour by sensitive, scheming, second
  guessing, social animals.
• - Steven Pinker

4
Spectrogram
5
The speech input consists of

• Frequency range 50-5600Hz
• Critical band filters
• Dynamic range 50dB
• Temporal resolution of 10ms
• Smallest detectable change in F0 2Hz
• Smallest change in F1 40Hz
• Smallest change in F2 100Hz
• Smallest change in F3 150Hz

6
Issues in models of speech perception

• Bottom-up vs top-down processing
• Acoustic phonetic invariance
• Segmentation of the signal into phonetic units
• Time normalization
• Talker normalization
• Lexical representations for optimal search
• Phonological recoding of words in sentences
• Dealing with errors in the initial representation
• Interpretation of prosodic cues

7
Bottom-up processing

• Peripheral processing
• Acoustic property detectors
• Phonetic feature detectors
• Segmental analysis
• Lexical search
• Syntactic and semantic analysis

8
Top-down processing

• Higher level processing
• Lexical cues
• Contextual cues
• World knowledge
• Cognitive skills

9
Acoustic phonetic non-invariance

• Intra-speaker variability
• Co-articulation (phonetic context)
• Which acoustic cues are used?
• How are the acoustic cues combined?

10
Segmentation into phonetic units

• Segments overlap (co-articulation)
• Segments are not always well defined acoustically
• Acoustic segments do not always match phonetic
  segments
• Errors are difficult to recover from

11
Overlapping speech segments
12
Effects of consonants on vowel transitions
13
Time normalization

• Duration of segments can vary up to two or three
  times
• Speaking rate
• Syllable stress
• Syntactic boundaries
• Co-articulation
• Duration is a phonetic cue that interacts with
  the other factors above

14
Talker normalization

• Formant frequencies depend of anatomical size
• Dialects vary
• Moderate amounts of noise, and distortion are
  common, but do not affect the accuracy of speech
  perception in normally hearing listeners
• Indexical communication

15
Lexical normalization

• Perceptual units?
• features
• phonemes
• syllables
• words
• beyond?

16
Phonological recoding

• The phonetic representation of a word depends on
  the sentence/word context.
• For example, /p/ is normally aspirated but in
  some contexts it is unreleased as in /apt/ or
  unaspirated as in /spIt/ (Kess, 1992).

17
Interpretation of prosodic cues

• Prosodic cues (F0, segmental durations, intensity
  contour) provide syntactic, semantic and
  pragmatic information
• The prosodic cues modify the acoustic phonetic
  cues
• How is the prosodic information preserved and
  incorporated

18
Models of Speech Perception

• Motor Theory (Liberman et al., 1967 Liberman
  Mattingly, 1985).
• TRACE (McClelland Elmann, 1986 Elmann, 1989).
• Neighborhood Activation Model (Luce, 1986 Luce,
  Pisoni, Goldinger, 1989).
• Logogen Theory (Morton, 1969 1979).

19
(No Transcript)
20
(No Transcript)
21
The Speech Chain
22
Why perception is not all phonetic!

• Woman Im leaving you
• Main Who is he?
• Blundy gets chair for wife
• Isle of view

23
Predicted vs Actual CNC words.
24
Effect of Context (at 5dBSNR).
Participants (in order of PTA)
25
Distortion and Audibility.