Content uploaded by Raphael Werner
Author content
All content in this area was uploaded by Raphael Werner on Jul 14, 2022
Content may be subject to copyright.
Methods
Introduction
Speaker discrimination and classification
in breath noises by human listeners
IAFPA –Prague
▪audible breathing frequent around speech [1, 2] or
during effortful actions [3]
▪as vital function, perhaps less affected by disguising voice
▪breath rarely used for forensic purposes (e.g. [4, 5])
▪speaker identification by neural networks looks
promising [6, 7]
[1] Rochet-Capellan, A., & Fuchs, S. (2013). The interplay of linguistic structure and breathing in German spontaneous speech. Interspeech, 2014–2018.[2] Kuhlmann, L. L., & Iwarsson, J. (2021). Effects of speaking rate on
breathing and voice behavior. Journal of Voice. [3] Trouvain, J., & Truong, K. P. (2015). Prosodic characteristics of read speech before and after treadmill running. Interspeech, 3700–3704.[4] Kienast, M., & Glitza, F. (2003).
Respiratory sounds as an idiosyncratic feature in speaker recognition. ICPhS, 1607–1610.[5] Braun, A. (2017). Nonverbal vocalisations –a forensic phonetic perspective. Laughter and other non-verbal vocalisations
workshop, pp.19–23. 2020. [6] Lu, L. et al. (2020). I sense you by breath: speaker recognition via breath biometrics. IEEE Transactions on Dependable and Secure Computing, 17(2), 306–319.[7] Zhao, W., Gao, Y., & Singh, R.
(2017). Speaker identification from the sound of the human breath. [8] van Son, R. et al. (2008). The IFADV corpus: a free dialog video corpus. LREC. 501–508.[9] Prolific. (2014). URL https://www.prolific.co. Accessed:
17/05/2022.[10]Finger, H. et al. (2017). LabVanced: a unified JavaScript framework for online studies. In International Conference on Computational Social Science, 2016–2018.[11]Jessen, M. (2007). Speaker classification
in forensic phonetics and acoustics. In: Müller, C. (eds) Speaker classification I. Lecture Notes in Computer Science, vol 4343. Springer, Berlin, Heidelberg.
References
▪research questions:
A. how well can listeners discriminate between same vs
different breathers?
B. how well can listeners guess a breather‘s age (young
vs old) and sex (male vs female)?
Raphael Werner, Jürgen Trouvain, Bernd Möbius
{rwerner|trouvain|moebius}@lst.uni-saarland.de
July 10-13, 2022
▪breath noises annotated in conversations [8]
▪5 oral(+nasal) inhalations each from 6 young (20-29 yrs;
3f, 3m) and 6 old speakers (age: 59-65 yrs; 3f, 3m)
▪33 participants (22f, 10m, 1 other; age: 20-71 yrs,
median: 31 yrs) via Prolific [9] and Labvanced [10]
▪overall correctness rate: mean = 64.3% (sd: 11.8%)
▪confidence rating: mean = 3.5 (sd: 0.76)
▪sex differences seem more perceivable than age
differences
▪different age + same sex even far below chance
▪young, female speakers stand out
Results
▪speaker discrimination possible, but not with high
accuracy
▪classification: sex > age (in line with findings for regular
speech from [11])
▪only binary distinctions for two categories here
A. discrimination task: 2 breath noises, separated by 500
ms silence →same or different speaker? how confident
(1-5)?
B. classification task: 1 breath noise →speaker young/old?
male/female? how confident each (1-5)?
Discussion and Conclusion
discrimination task classification task
sex
age
same_male
same_female
different
total
same_old
79.5% (73) 76.8% (69) 60.0% (35)
74.6% (177)
same_young
73.1% (67) 53.1% (64) 65.0% (40)
63.7% (171)
different
31.8% (22) 35.3% (34) 63.8% (58)
49.1% (114)
total
70.4% (162)
59.3% (167)
63.2% (133)
64.3% (462)
▪overall correctness rate:
▪age: mean = 50.2% (sd: 9.1%); confid. = 3.0 (sd: 0.75)
▪sex: mean = 66.7% (sd: 13.5%); confid. = 3.2 (sd: 0.77)
▪confounding factors: biological vs chronological age?
height/weight?
▪implications for using breath noises in synthetic speech
▪breath noises relevant in real-world forensic applications
(e.g. rape, black box)
Table: Correctness rate by speaker sex and age in percent. Numbers in
brackets indicate number of stimuli per cell.