1. Acoustics´08 Paris, 29 June – July 2008
Identifying speaker-dependent acoustic parameters in Spanish vowels
Victoria Marrero
Universidad Nacional de Educación a Distancia
Elena Battaner
Universidad Rey Juan Carlos
Juana Gil
Consejo Superior de Investigaciones Científicas
Joaquim Llisterri, María Machuca,
Montserrat Marquina, Carme de la Mota, Antonio Ríos
Universitat Autònoma de Barcelona
Acoustic and Perceptual Study of Inter and Intra-speaker Variation in Spanish (VILE-II Project)

2. Identifying speaker-dependent acoustic parameters in Spanish vowels
VILE Projects (Inter-and-Intra-Speaker-Variation-in-Spanish)
1st. stage:
BFF VILE I: Estudio acústico de la variación inter e intralocutor en español ( )
Acoustic Study of Inter and Intra-Speaker Variation in Spanish
2nd. stage:
HUM /FILO VILE II: Estudio perceptivo de la variación inter e intralocutor en español ( )
Perceptual Study of Inter and Intra-Speaker Variation in Spanish

3. Identifying speaker-dependent acoustic parameters in Spanish vowels
Which acoustic parameters of the vowels are
More dependent on individual characteristics of the speaker
Less dependent on linguistic variables
Variations of the standard deviation (SD)
Different groupings of the variables

4. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

5. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

6. Identifying speaker-dependent acoustic parameters in Spanish vowels
Acoustic parameters of the vowels
F1, F2, F3, F4 F0
Individual or linguistic?
Literature
High formants (F3 - F4)  Individual information
Low formants (F1 – F2)  Vowel quality
F0  Vowel quality (intrinsic F0)
 Suprasegmental variations (intonation, tone, stress)
 Speaker identification

7. Identifying speaker-dependent acoustic parameters in Spanish vowels
The sample
Inter-session comparison: same text -- 3 sessions
The grouping
By session (1st, 2nd, 3rd.)
By speaker (from 1 to 30)
By phoneme (/i, e, a , o/, /`i, `e, `a, `o/).
The analysis
Variations on Standard Deviation (SD)  statistical dispersion  grouping
Inferences
Small SD = little variation
Phoneme clustering – small SD  Dependency on vowel quality
Speaker clustering - small SD  Dependency on individual parameters

8. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

9. Identifying speaker-dependent acoustic parameters in Spanish vowels
When grouping data by speaker or session, individual parameters will show less SD than vowel quality parameters
When grouping data by phoneme, individual parameters will show more SD than vowel quality parameters
F0 will show some characteristics of both individual parameters and vowel quality parameters

10. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

11. Identifying speaker-dependent acoustic parameters in Spanish vowels
Sample
AHUMADA database
30 male speakers
Reading a phonetically balanced text
3 different recording sessions
Automatic extraction (Praat) + manual supervision
F0, F1, F2, F3 and F4
/i, e, a, o/
Context: [p, t, k, s] _ [p, t, k, s]
Lexically stressed / unstressed

12. Identifying speaker-dependent acoustic parameters in Spanish vowels
Sample
By phoneme
/i/
/e/
/a/
/o/
Total
Syllable
Stressed
Unstr.
Unstr
Contexts number 1 6 3 5 21
Segments 90 89 84
536
270 83
450
265
1867
By acoustic parameter F0 F1 F2 F3 F4
Total
Items
1848
1849
1847
9240

13. Identifying speaker-dependent acoustic parameters in Spanish vowels
Grouping
By sound = Baseline groups = 30 speakers * 3 sessions * 21 contexts Minimal variation (same speaker, same sound, same context and same stress)
By session = One voice in one moment 90 groups = 3 sessions * 30 speakers No vowel / syllable / context differences
By speaker = One voice in whatever moment 30 groups = 30 speakers. No vowel / syllable / context / session differences
By phoneme = Enhancement of linguistic factors (voice quality and stress) 8 groups = 4 vowels * 2 syllables (stressed / unstressed) No speaker / session differences
All together (total) = Maximum variation 5 groups = F1, F2, F3, F4, F0

14. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

15. Identifying speaker-dependent acoustic parameters in Spanish vowels
Standard Deviation (Hz) F0 F1 F2 F3 F4
a. By sound
7.22
26.00
97.75
151.71
157.88
b.By phoneme
12.44
82.13
411.79
228.44
225.18
c. By session
13.46
86.07
407.45
239.96
235.61
d. By speaker
19.58
39.37
183.39
263.62
320.15
e. Total
21.06
84.03
410.14
288.19
332.73
lower
higher

16. Identifying speaker-dependent acoustic parameters in Spanish vowels
Statistical relevance of differences between clusters
Sound - Phoneme
Sound - Session
Sound - Speaker
Phoneme - Session
Phoneme - Speaker
Session - Speaker F0
0.0000
0.0008
0.3775 F1
0.0431
0.8116 F2
0.0262
0.7102 F3
0.0333
0.0014
0.3484
0.5013
0.5948 F4
0.0015
0.0036
0.0078
0.5721
No difference Session / Speaker
No difference Stressed / Unstressed vowels for F1, F2, F3
Low sensibility of F3

17. Identifying speaker-dependent acoustic parameters in Spanish vowels
Proportion of differences in SD F0 F1 F2 F3 F4
a. By sound
1.00
b. By phoneme
2.71
1.51
1.88
1.74
2.03
c. By session
1.71
3.15
4.43
1.50
1.42
d. By speaker
1.86
3.31
4.38
1.58
1.49
e. Total
3.00
3.23
4.41
1.89
2.11
Proportion of differences  By sound (= baseline) / other clusters in the same column (parameter)

18. Identifying speaker-dependent acoustic parameters in Spanish vowels
F2 - F1 – F0 are more variable than F3 – F4
F1 – F2  dramatic reduction => by phoneme F0 F1 F2 F3 F4
a. By sound
1.00
b. By phoneme
2.71
1.51
1.88
1.74
2.03
c. By session
1.71
3.15
4.43
1.50
1.42
d. By speaker
1.86
3.31
4.38
1.58
1.49
e. Total
3.00
3.23
4.41
1.89
2.11
F3: same tendency but less pronunced
10% lower by session / speaker than by phoneme.
F0  highest SD by phoneme
 30% lower by session/speaker
F4: same tendency than F0:
high variation by phoneme
27% lower by session/speaker

19. Identifying speaker-dependent acoustic parameters in Spanish vowels
a. By sound
1.00
b. By phoneme
2.71
1.51
1.88
1.74
2.03
c. By session
1.71
3.15
4.43
1.50
1.42
d. By speaker
1.86
3.31
4.38
1.58
1.49
e. Total
3.00
3.23
4.41
1.89
2.11
Conclusion:
Variations in SD in F0, F4 and F3 can be explained by the factor ‘speaker/session’
F1 and F2 variations can be explained better by the factor ‘phoneme’

20. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

21. Identifying speaker-dependent acoustic parameters in Spanish vowels
a) The most frequently-used parameters in forensic speaker identification / verification:
F0 (Jiang 1996)
Mean
Standard Deviation
Long-term mean / SD (Rose 1991)
Likelihood ratio
Base-value factor (Traunmüller & Eriksson 1995)
Alternative baseline (Lindh 2007, 2006)
Formants or formant trajectories (Hollien 1990, Kuwabara & Tagaki 1991, Kreiman & Papcun 1991)
b) SD in speaker verification
Score normalization techniques - text-independent speaker verification systems (Mariéthoz & Bengio 2005, Saeta & Hernando 2003)

22. Identifying speaker-dependent acoustic parameters in Spanish vowels
This study:
SD of F0 – Formant Mean
grouped
By speaker = individual variable
By phoneme = linguistic variable
Acoustic parameters which characterize a voice
F2 & F1
Higher SD
» cluster by phoneme  variation decreases drastically
F3 & F4
Lower SD
» cluster by speaker  lowest variation

23. Identifying speaker-dependent acoustic parameters in Spanish vowels
Fundamental frequency (F0)
The most sensitive parameter to the individual grouping
F3, a special case
Unexpected dependency on vowel quality
Relation to the size of the area behind the lower teeth (Sundberg 1973, 1974)
Highest values  /i/
Lowest values  /a/

24. Identifying speaker-dependent acoustic parameters in Spanish vowels
Introduction
Hypotheses
Method
Results
Discussion
Conclusions

25. Identifying speaker-dependent acoustic parameters in Spanish vowels
Do our hypothesis have been confirmed?
Mean SD by speaker/session:
F4 < F3 < F0 < F1 < F2
High formants -F4- are more dependent on the speaker's voice than on vowel quality
F2 and F1 are close related to vowel quality differences
When grouping data by speaker or session, individual parameters will show less SD than vowel quality parameters

26. Identifying speaker-dependent acoustic parameters in Spanish vowels
Do our hypothesis have been confirmed?
Mean SD by phoneme:
F1 < F3 < F2 < F4 < F0
When grouping data by phoneme, individual parameters will show more SD than vowel quality parameters
F1  strongest relationship with vowel quality
F3  high dependency on vowel quality, even more than F2

27. Identifying speaker-dependent acoustic parameters in Spanish vowels
Do our hypothesis have been confirmed?
F0 will show some characteristics of both individual parameters and vowel quality parameters
Like vowel quality parameters: high level of general SD (total)
Like individual parameters: highest level of SD when grouping by phoneme,
even when stressed /unstressed vowels are computed separately F0 F1 F2 F3 F4
e. Total
3.00
3.23
4.41
1.89
2.11

28. Identifying speaker-dependent acoustic parameters in Spanish vowels
Acknowledgments
The authors would like to extend their appreciation to Carme Carbó, Natalia Madrigal and Montserrat Riera for their previous work on the research project VILE (Variación acústica inter e intralocutor en español).
This research has been achieved with the financial support of the Ministerio de Educación y Ciencia. Project Nº HUM FILO.
Thank you!