Expanding the set of acoustic features

1
Expanding the set of acoustic features of the
post-vocalic voicing contrast in English B.
Rodgers, T. Purnell J. Salmons, University of
Wisconsin, Madison
1.0 Introduction
3.1 Methodology
2.2 Results

• Same data set as Experiment 1
• Acoustic Measures
• Frequency Characteristics
• Change in F1, F2, F3 frequency prior to formant
  end
• Change in f0 frequency prior to formant end
• Temporal Energy Characteristics
• H1 end time to 15 dB drop in first harmonic
• H2 end time to 15 db drop in second harmonic
• Formant end 15 dB drop in F2 from local maximum
• H1 energy average energy over 30 msec prior to
  formant end for 150-350 Hz (Fig. 7)

• ANOVA, means for acoustics (Table 1, plt0.05, df
  1, 104)
• Energy factors significant interactions (Fig. 4)
  for F2 end (F30.3) and RMS knee (30.3)
• Frequency factors only significant main effect
  for f0 change for VOICING (F5.3)

• Traditional View of phonetic laryngeal contrasts
  (/t//d/, VOICING) F0 drop, F1 drop, pulsing in
  the gap, CV Ratio, etc. (Kingston et al. 2008,
  Lisker 1986, Raphael 1972, Port Dalby 1982,
  Stevens 1998)
• Often have samples that dont conform to norms
  (Fig. 1)
• No single measure found to universally correlate
  with voicing
• Alternative View termination of the preceding
  vowel, nature of RMS drop (Parker 1974)
• Goal
• Better characterize VOICING with measures
  available
• Evaluate measures (RMS) associated with vowel
  transition

Gesture F2 db RMS F1 f0 end end knee ? ? (ms
) (ms) (ms) (Hz) (Hz) /t/ Isolation 266 231 201 19
7 47 (31) (27) (29) (174) (95) Sentence 269 219
188 151 19 (30) (28) (19) (180) (71) /d/ Isolati
on 328 313 317 226 -6 (36) (37) (39) (89) (10) S
entence 262 236 231 192 14 (31) (31) (35) (70) (
61)
F1
F1
F2
F2
F3
F3
h1
h1
b)
a)
a)
b)
Figure 7. Formant energy and first harmonic
energy (a) bat (b) bad. Shading shows energy in
first harmonic at formant end (vertical aqua
line).
Table 1. Means and standard deviations.
3.2 Results

• ANOVA, means for acoustics (Table 3, plt0.05, df
  1, 104)
• Energy factors significant main effect for
  VOICING h1-h2 (F37.6), energy h1-h2 (F16.5)
  and mean energy h1(F361.1) for Condition h1-h2
  (F40.1)
• Frequency factors significant main effect for
  VOICING F2 change (F4.1) and f0 change (F5.3)

d)
c)
Figure 1. Formant plots (a) normal /bat/ (b)
normal /bad/ (c) abnormal /bat/ (d) abnormal
/bad/. Blue plane is RMS knee, green plane is
lingual closure.
F2 f0 h1-h2 mean ? ? (ms) energy h1
(Hz) (Hz) (dB) /t/ Isolation 98 47 181 -67 (42
7) (95) (102) (3) Sentence 60 19 -73 -67 (383) (
71) (139) (3) /d/ Isolation -16 -6 337 -52 (132)
(10) (183) (3) Sentence -59 14 174 -54 (91) (61
) (231) (5)
Figure 4. Comparison of Gesture end with F2 end.
2.0 Experiment 1

• Covariation with lingual closure (r2, Table 2,
  plt0.05)
• Energy factors significant for both factors in
  all conditions except /t/ in sentences (Fig. 4)
• Frequency factors no significant covariation

• Conventional wisdom Frequency values
• Formant transitions will show trend for /t/ vs.
  /d/
• f0 transition will show well-defined trend for
  /t/ vs. /d/
• Test energy vs frequency information

r2 F2 dB end RMS knee F1 ? f0
? /t/ Isolation 0.76 0.72 n.s.
n.s. Sentence 0.40 n.s. n.s.
n.s. /d/ Isolation 0.94 0.92 n.s.
n.s. Sentence 0.93 0.59 n.s. n.s.
2.1 Methodology

• N27 female speakers from Upper Midwest (MN, WI)
• Context bat and bad, isolated and sentence (Use
  this new bat/Make this one bad)
• X-ray microbeam tongue-tip data
• Acoustic Measures
• Frequency Characteristics
• Change in F1 frequency prior to formant end
• Change in f0 frequency prior to formant end
• Temporal Energy Characteristics
• Formant end 15 dB drop in F2 from local maximum
  (Fig. 2)
• RMS knee RMS derivative exceeds negative
  threshold of -0.33 dB/ms (Fig. 3)

Table 3. Means and standard deviations for select
measures.
Table 2. Covariation with lingual closure.
(plt0.05)
2.3 Discussion

• Final /t/ characteristics
• RMS knee prior to occlusion (Fig. 3a)
• phonation terminates prior to occlusion
• Final /d/ characteristics
• RMS knee at time of occlusion if present (Fig.
  3b)
• phonation continues past occlusion
• Formant transitions not significant for voicing
• f0 transition not significant for voicing

Figure 8. Comparison of F2 end with mean H1
energy. Box identifies region of /t/ distinct
from /d/, regardless of context.

• Interpretation
• Energy for final /d/ determined by
  supra-laryngeal gestures (lingual closure and
  enhancing gestures)
• Energy for final /t/ determined by laryngeal
  gesture (Fig. 5)
• Traditional measures of frequency for f0 and F1
  weak correlation with voicing at best

F3
15 dB
F2
3.3 Discussion
F1

• /t/ vs /d/ difference
• Vowel end determined by lingual closure for /d/
• Vowel end determined by laryngeal termination
  gesture for /t/
• Relation between H1 energy prior to vowel end
  and VOICING is significant (Fig. 8)
• Implications
• Confirm importance of vowel duration
• Energy levels as opposed to frequency values
  (Parker 1974)
• Energy measures greater correlation with VOICING
  than traditional measures(f0 delta, F1 delta)
• Energy measures during transition informative
  for consonant voicing in addition to vowels
  (Jenkins et al. 1983)

/d/
phonation
Figure 2. Formant plot. End of vowel defined as
end of second formant point where second formant
15 dB down from local maximum.
enhancing gesture
hold
approach
release
/t/
phonation
termination gesture
Figure 5. Timing diagram. Shows supra-laryngeal
gesture for syllable final /d/, laryngeal
termination gesture for final /t/.
4.0 Conclusion

• Work to date on speech perception in post-vocalic
  stops has focused on frequency transitions,
  percent closure voicing, vowel duration.
• Energy characteristics can provide more robust
  acoustic measures for voicing than traditional
  measures.
• /t/ is characterized by a phonation termination
  gesture (glottal stop) which has implications for
  phonological theories.
• Going forward investigate importance of energy
  in perception.

3.0 Experiment 2

• Test role of phonation and relation between
  energy and frequency
• Exploit difference in lower harmonic behavior
  between final /t/ and /d/ (Fig. 6)
• Examine energy behavior in lower harmonics
  relative to energy in formants

References Jenkins, J. et al., 1983, Perception
Psychophysics, 34, 441-450. Kingston, J. et al.,
2008, J Phonetics, 36, 28-54. Lisker, L., 1986.
Language Speech, 29, 3-11. Parker, F., 1974, J
Phonetics 2, 211-221. Peterson, G. Lehiste, I.,
1960, JASA, 32, 693-703. Port, R. Dalby, J.,
1982, Perception Psychophysics, 32,
141-152. Raphael, L., 1972, JASA, 51,
1296-1303. Stevens, K. 1998. Acoustic Phonetics,
Cambridge, MA MIT Press. Acknowledgements We
would like to thank Eric Raimy for discussion.
All mistakes are our own. Contact
Info brodgers_at_wisc.edu tcpurnell_at_wisc.edu
jsalmons_at_wisc.edu
h3
h3
b)
h2
h2
h1
h1
a)
b)
Figure 3. RMS and rate of change of RMS (a) bat
(b) bad. (a) Syllable final /t/ shows RMS knee
prior to occlusion (vertical green) while (b)
final /d/ shows RMS knee close to occlusion.
Figure 6. Harmonic plots (a) bat (b) bad. Final
/t/ shows simultaneous end of lower 3 harmonics,
final /d/ shows continuation of h1 relative to
higher harmonics.