topics in applied phonetics - simon fraser universitymjmunro/courses/l411h10.pdf · topics in...
TRANSCRIPT
11/4/16
1
LING 411Topics in Applied Phonetics
MIDTERM EXAM: Nov. 8
Topics:Non-expert Speaker ID; Expert Speaker ID, Speaker Profiling, Intra-speaker Variability, Vocal disguise
Format: Essay (long answer) w/choice of topics
Notes permitted:You may bring one page of computer-printed notes, which
must be handed in with your exam paper.
NB: All responses to questions must be based on what you’ve learned this term. Speculative answers will not receive
points.
11/4/16
2
R. Rodman (2000). Computer Recognition of Speakers Who Disguise Their Voice
1989 -1994
Germany: Disguise used in 52% of cases where offender spoke and
expected to be recorded
blackmail cases ≈ 69%
R. Rodman (2000). Computer Recognition of Speakers Who Disguise Their Voice
Disguise classification
deliberate vs non-deliberate [intent vs effect of transmission medium]
electronic vs non-electronic [electronic device intentionally employed, rather than some
“natural” disguise]
11/4/16
3
Source: R. Rodman (2000). Computer Recognition of Speakers Who Disguise Their Voice
Examples of bite blocks
11/4/16
4
Talkers = 40 adult male
Sentence-length readings
Speaking Conditions (demonstrated to talkers)
UND = undisguisedOA = “old age,” 70-80 years oldH = hoarse voiceN = nasalizedSR = slow speaking rateFD = free disguise
11/4/16
5
Listeners
Naïve: 24 undergradsSophisticated: 3 grad students + 3 profs in Speech/Hearing Sciences
Stimuli
pairedtraining given in a same voice/ different voice task
Testing
different voices, 360 pairs, judged in 3 sessions, different days
same/different task + confidence ratings
UND = undisguisedOA = “old age,” 70-80 years oldH = hoarse voiceN = nasalizedSR = slow speaking rateFD = free disguise
significantly different from all others
three significant differences
11/4/16
6
Weak to moderate tendency for confidence ratings to predict accuracy
NB: a full understanding of the data requires a look at false positives & false negatives as well
FROM: http://wwwhomes.uni-bielefeld.de/ttrippel/hths/hths_acoustic_praat.html
Source-Filter Theory:Vowels
source = rich sound fromvibrating vocal folds
INDEPENDENT OF
filter function = resonances due to SLVT
configuration
11/4/16
7
Theoretical effect of pitch-based disguise on formants?
None, because source and filter are independent.(also, disguises that affect formants should not alter pitch)
So disguises might be analyzed in terms of whether they affect the source or the filter.
Disguise Mechanism Affects Acoustic Effect
Raise/lower pitch vary vocal fold tension SOURCE change in F0
Modify perceived body size lower/raise larynx FILTER shift in formant
frequencies
Modify perceived body size
inhale helium(or sulphur hexafluouride)!
FILTER increase in formant frequencies
Modify voice qualitylaryngeal settings: whisper, hoarse/breathy voice, creaky
voiceSOURCE changes in F0, spectral
properties, HNR
Non-specific bite block/object in oral cavity FILTER changes in spectral and temporal properties
Non-specific electronic voice changer BOTHchanges in F0, formants, other spectral properties,
reverberation, noise
11/4/16
8
Overcoming speech disguise in FSID
• look for phonetic aspects of speech that are minimally affected by the type of disguise used
• identify specific acoustic phonetic consequences of particular types of disguise and try to reverse them
• use automatic speaker recognition to match disguised and undisguised voices
11/4/16
9
Andruski et al.
influenced by articulation?
marker of segment identity?
marker of speaker identity?
F1, F2 greatly + –
F3, F4 minimally – +
F0 minimally – +
F0: a common mechanism of vocal disguise, so F0 measurements may be unhelpful in FSID.
Andruski et al.
10 speakers (5F, 5M)
produced a reading passage 3 times each using:
1) normal voice2) lowered pitch3) vocal disguise of their own choosing
11/4/16
10
11/4/16
11
includes a change in articulation
11/4/16
12
Andruski et al.
Data: Measurements of F0, F1, F2, F3, F4 for all speech samples.
Statistical approach: Discriminant analysis
Which measures can be used to classify items into distinct groups (i.e., match same voices and separate different voices)?
Andruski et al.
Findings:
In the ‘freely chosen’ disguise condition:
all speakers raised F0all speakers showed high F0 variability
11/4/16
13
Andruski et al.
Findings:
For all speakers:lowest F1/F2 variability: low F0 conditionhighest F1/F2 variability: disguise condition
The high variability may be due to imitating an accent and using less stable vowel qualities.
Andruski et al.
Findings:
In low F0 condition, all formants (F1-F4) decreased in frequency!!
BUT...
11/4/16
14
Andruski et al.Findings:
In low F0 condition, all formants (F1-F4) decreased in frequency!!
The S-F Theory says that F0 and formants are independent of each other.
Does this mean the S-F Theory is wrong?
Andruski et al.
Findings:
In low F0 condition, all formants (F1-F4) decreased in frequency!!
Does this mean the S-F Theory is wrong?
No! When the speakers lowered F0, they also tended to lower their larynges, thus increasing effective vocal tract length. (Part of strategy to “sound bigger”)
11/4/16
15
Andruski et al.
S-F Theory models normal speech production by assuming independence of source and filter.
NOT the same as saying
that speakers always control source and filter independently of each other (especially during vocal disguise).
Andruski et al.
F3 and F4 are only modestly effective for matching disguised/undisguised voices (36%), but they are about twice as effective as F1 and F2.
11/4/16
16
Creaky productions from two female German speakers
11/4/16
17
Moosmüller (2001)
Talkers = 5 F; 4 M
Produced sentences ina) modal voiceb) creaky voice
Target vowels were analyzed from selected words
Difference in F2 and F3 between modal and creak was determined
Converted to Bark units to facilitate comparisons
11/4/16
18
Moosmüller (2001)
Similar observations across vowels:
Females show consistent lowering of F2Males and other formants – no clear pattern
Conclusion ?
Moosmüller (2001)
Similar observations across vowels:
Females show consistent lowering of F2Males and other formants – no clear pattern
Conclusion:
11/4/16
19
L411 Term Paper References
• must be published in scholarly venues (journals, books, conference proceedings). If in doubt, just ask.
• must clearly relate to forensic phonetics and not mainly to other areas of forensic linguistics
• must not include papers or chapters that we have covered in-depth in class
Is there a telephone effect on vowel formants in both perception and production?
11/4/16
20
/i/, /æ/, and /u/ productions
Statistical shift in F1 values for the close vowels
11/4/16
21
% shift in F1, F2, F3 values for the close vowels
Filled markers are ‘direct’
11/4/16
22
Perceptual Evaluation
Method
No perceived difference!Filled markers are ‘direct’
Phoneticians judgments