A hierarchical and ITfriendly

1
A hierarchical and IT-friendly approach to speech
segmentation Sven Mattys Experimental
Psychology University of Bristol
2
Illustration of Speech continuity
3
The segmentation challenge
Physical continuity (speech signal) versus Perc
eptual discreteness (words)
They had been comprehensively defeated
1 2 3 4 5
4
I. EXPLICIT SEGMENTATION

• Segmentation precedes recognition.
• Listeners use explicit sub-lexical
• segmentation cues
• (a) coarticulation
• (b) phonotactics
• (c) allophonic variations
• (d) prosody

5
II. IMPLICIT SEGMENTATION

• Segmentation is a by-product of word
• recognition.
• Trade-off Multiple lexical activation.
• Lexical competition naturally settles on best
• parsing solution.
• (e.g., obey-convince vs. o-bacon-vince).

6
Current theories are rather bi-polar
Implicit Segmentation
Explicit Segmentation
but it doesnt have to be that way
7
Interpretive conditions
Segmentation strategies
Pragmatic, semantic, and syntactic context
Optimal
Poor contextual information
Lexical knowledge
Phonotactics Allophones Coarticulation
Poor lexical information
Poor segmental information
Prosody
8
Segmentation cues/strategies are hierarchically
organised
Pragmatic, semantic, and syntactic context
Lexical knowledge
Phonotactics Allophones Coarticulation
Prosody
9
Example of cue conflict Stress vs. coarticulation
comproNO-ta compa-deTE comproNO-ta compa-deTE
Intact
Coartic. gt Stress
Degraded
Stress gt Coartic.
Probed segmentation point
10
CONCLUSION

• The architecture of the model is
  language-general but its components are
  language-specific.

• Developmentally, the model may build up from the
  bottom levels up to more lexically- and
  contextually-driven strategies over time.