Formant Synthesis of Speaker Age ( a question regarding prosodic timing)

1
Formant Synthesisof Speaker Age( a question
regarding prosodic timing)
Swedish National Graduate School of Language
Technology

• Susanne Schötz
• Centre for Languages and Literature
• Lund University
• susanne.schotz_at_ling.lu.se

2
A Prototype System for Analysis of
Speaker Age by Formant Synthesis
3
background (1)speaker age

• acoustic cues to speaker age in almost every
  phonetic dimension
• relative importance of these cues not fully
  explored
• one reason the lack of an adequate analysis tool
  (where a large number of potential age parameters
  can be varied systematically and studied in
  detail)

4
background (2)formant synthesis

• robust flexible, but not as natural-sounding as
  concatenation synthesis
• GLOVE (OVE III with improved glottal source, dev.
  at Royal Institute of Technology, Stockholm) used
  in experiments of voice variation since 1989
  (Carlson et al., 1991 Karlsson, 1992)
• data-driven formant synthesis (GLOVE)(Sjölander,
  2001 Sigvardson, 2002 Öhlin, 2004)

5
purpose aim

• develop a prototype system for analysis of
  speaker age by data-driven formant synthesis
• use as a tool to analyse, model and synthesize
  speaker age

6
material

• 1 word själen '???l?n (the soul)
• 4 speakers (same dialect family)
• speaker 1 a granddaughter (aged 6)
• speaker 2 a daughter (aged 36)
• speaker 3 a mother (aged 66)
• speaker 4 a grandmother (aged 91)

7
materialacustic pre-analysis

• F0-contours for '???l?n F1-F2 plot for the
  steady-state part of ??

8
method (1)

• data-driven analysis by synthesis(GLOVE, Praat,
  Perl)
• automatic extraction of 23 GLOVE parameters
  (once every 10 ms)
• formant synthesis (GLOVE)
• audio-visual comparison of synthesis to original
• parameter adjustment rules

9
resultsspeaker 1 (6 years)

• original synthesis
• fricative, creak
• - formant error, amplitude

10
resultsspeaker 2 (36 years)

• original synthesis
• formants, F0
• - dull

11
resultsspeaker 3 (66 years)

• original synthesis
• formants, F0
• - amplitude, dull

12
resultsspeaker 4 (91 years)

• original synthesis
• formants, F0 (incl. creak)
• - amplitude, fricative, dull

13
method (2)

• weighted linear interpolation between two source
  speakers to synthesize a target age (Praat,
  Java)
• ex. target age 51
• source speakers 2 (36) and 3 (66)
• age weights for each source speaker 0.5
• duration interpolation for segment 1(source
  speaker 2 dur 100 ms, source speaker 3 dur
  200 ms)
• target dur 100 x 0.5 200 x 0.5 150 ms

14
results (interpolation)

• at a first glance similarities
• linear interpolation not optimal aging is not
  linear!
• only first attempt
• how evaluate?
• - perception tests (31 students)
• age
• (naturalness)

15
evaluation results natural speakers
CA PA
6 7
36 30
66 35/36
91 74
16
evaluation results data-driven synthesis
CA PA
6 12
36 41
66 44
91 69
17
evaluation results linear age interpolation
CA PA
10 13
20 63
30 54
40 42
50 42
60 54
70 58
80 70
18
summary discussion (1)

• both similarities and differences between
  original and synthesis
• a good start, but needs more work
• formants
• amplitudes
• voice source parameters
• try other interpolation algorithms (spline?)
• if developed further, may be used to to model and
  synthesize speaker age

19
Question about prosodic timing

• Bruce (1983) strong Swedish dialect -gt later
  tonal peaks
• older speakers often sound more dialectal
  (Stölten Engstrand, 2003)
• could prosodic timing be age-related?
• my hypothesis older speakers -gt
• later tonal peaks.
• investigated 1 word Nordanvinden (4 speakers)

20
F0 contours for nordanvinden
66
6
91
36
21
F0 contours for nordanvinden
66
6
H
H
91
36
H
H
22
Questions

• What have I missed?
• synchronisation to segments?
• context effects?
• ?
• How do I proceed?
• suggestions?
• Do you (as prosodic experts) believe that there
  may be a relation between age and prosodic timing?

23

• thank you!
• questions, answers and comments welcome!