Working memory and prosodic phrasing

1
Working memory and prosodic phrasing

• Merle Horne, Johan Frid and Mikael Roll
• Dept. of Linguistics Phonetics, Lund University

2
  The role of function words in spontaneous 
speech processing   SPEECH TECHNOLOGY PROGRAM
PROJECT Funding 2003-2004 MERLE HORNE  JOHAN
FRID GÖSTA BRUCE BIRGITTA LASTOW MIKAEL
ROLL ADINA SVENSSON www.ling.lu.se/projects/ProSe
g2.html
.
3
ATTh
THAT PAUSE just to work (with) that ONE
PAUSE likes to do
A T T
PAUSE just å jobba det MAN
PAUSE trivs med å göra
4
AND EH PAUSE even a little erobics IN
INHALE the wintertime WHEN EH PAUSE
soccer is over
OCH EH PAUSE även lite gympa PÅ INHALE
vinterhalvåret NÄR EH PAUSE fotbollen är slut
5
Spontaneous speech data (Swedish)
Observation speech fragments (production
units) following hesitations are often ca. 2-2.5
sec long Question Are there timing restrictions
on speech production units?
6
Support from memory research

• Baddeley (1997)
• Part of working memory where speech coding takes
  place (phonological loop) has a time limit of
  around 2 seconds (inner speech)
• Cues from neurolinguistic research
• Ackermann Hertrich (2003)
• Temporal organization of internal speech is
  controlled by the cerebellum (internal clock)

7
(No Transcript)
8

• Evidence for timing restriction on production
  units
• PROSODIC
• Very often inhalation after 2-2.5 sec speech
• often a pause after 2-2.5 sec speech
• often a F0-declination pattern spanning over
  2-2.5 sec. and often boundary tone (H/L) after
  2-2.5 sec speech (cf. Intonation(al) phrase,
  tone unit )
• often final lengthening/laryngealization after
  2-2.5 sec speech
• SYNTACTIC
• often a constituent-boundary after ca.
• 2-2.5 sec speech

9
If a timing restriction on speech coding exists,
it can be used in developing algorithms for
parsing spontaneous speech

• Basic assumptions
• A 2-2.5 sec speech production unit can contain
  internal pauses
• A 2-2.5 sec speech production unit does not
  contain internal inhalations, i.e. inhalations
  occur only at the edges of production units
• A 2-2.5 sec speech production unit optimally
  corresponds to a clause or a constituent

10
so that then we moved down the shop
there
så att då flyttade vi ner butiken dit
ltt2gt
ltt2/gt
Ca. 2.5 sec.
11
people are so PAUSE close
INHALE
man är så pass PAUSE nära INHALE
ltt2gt
ltt2/t2gt
Ca. 2.5 sec.
12
PAUSE now we do something EH
good here
PAUSE nu gör vi något EH bra
här
lt/t2gtltt2gt
lt/t2gt
Ca. 2.5 sec.
13
that it would just be a
PAUSE
therapy-assistent course INHALE
att det skulle bara bli en PAUSE
terapibiträdeskurs INHALE
Ca. 2 sec.
Ca. 2 sec.
14
then they had to help very much with it
Since then we had to get up PAUSE redo INHALE
då fick de ju hjälpa till väldigt mycket med det
för att då skulle vi ju ha upp PAUS göra om INHALE
H
ltt2/gtltt2gt
Ca. 2 sec.
Ca. 2 sec.
15
Conclusions

• Evidence for isochronal 2-2.5 sec speech
  production units useful in segmentation of speech

• Breathing seems to play an important role in
  delimitation of these units Inhalations only
  occur at edges and can thus function as anchors
  for the grouping of speech into 2-2.5 sec speech
  segments

• Local prosodic information (Pauses, boundary
  tones (H/L) timing restriction, can be used
  to make a further segmentation of spontaneous
  speech into 2-2.5 sec production units

16
Follow-up studies..

• Designing an algorithm for segmenting speech
  using timing restriction other prosodic
  parameters
• Testing algorithm on more speakers
• Neurophysiological experiments to look for
  external support for timing unit boundaries