Post on 19-Mar-2020
Analysis ofmutual durationandnoise effects in speaker recognition:benefits of condition-matchedcohort selection in scorenormalization
Andreas Nautsch? Rahim Saeidi†
Christian Rathgeb? Christoph Busch?
? da/sec – Biometrics and Internet Security Research Group, Hochschule Darmstadt, Germany† Department of Signal Processing and Acoustics, Aalto University, Finland
andreas.nautsch@{cased|h-da}.de
Motivation
• Real-life conditions in commercial andforensic applications
• Vast broadness of conditions e.g.:
– Sample completeness (duration)– Ambient noise (AC / CROWD)
• True sample condition remains unknowndue to e.g., SNR estimations
• Unified Audio Characteristics seempromising for condition-matching score-normalization
Research Questions
Q1: How extensive are mutual effects to the per-formance of a State-of-the-Art system?
Q2: Does condition-informed cohort selectionbenefit from statistic approaches ratherthan from condition-matched cohorts?
Q3: Do conditions affect basic i-vector proper-ties, i.e. mean i-vectors?
Related Work
Unified Audio Characteristics
• Single multivariate Gaussian models inoriginal (raw) i-vector spaceΛi ∼ N (µi,Σ), i = 1, . . . , 55
– Condition-dependent µi– Shared full Σ by pooling
• Condition quality vector (q-vector) ~q asposterior probabilities for i-vector ~w:
~q(i) =P (~w |Λi)∑55i=1 P (~w |Λi)
Conventional AS Score Normalization
• Adaptive & Symmetric
– Refs vs. probe-alike cohort set
– Prbs vs. reference-alike cohort set
– Computation of: µref, σref, µprb, σprb
• AS-normalized scores SAS by averagedsymmetric zero-norm of score S:
SAS =1
2
(S − µref
σref+S − µprb
σprb
)Cohort Selection Criterion
• Symmetric Kullback-Leibler divergence:
1
2
55∑i=1
~qa(i) log~qa(i)
~qb(i)+ ~qb(i) log
~qb(i)
~qa(i)
• Selecting top-c cohort q-vectors
Experimental Set-up & Results
Condition 1 2 3 4 5 6 7 8 9 10 11 . . . 15 16 . . . 30 31 . . . 55
Duration 5 s 10 s 20 s 40 s full 5 s 10 s 20 s . . . full 5 s . . . full
Noise clean Air Conditioner (AC) CROWDSNR 0 dB 5 dB 10 dB 15 dB 20 dB 0 . . . 20 dB 0 dB . . . 20 dB 0 dB . . . 20 dB
I4U Data of NIST SRE’12
• Condition-dependent sampleversions from long-duration &clean samples
• VAD labels of clean samples
• LDA: 400 to 200
• PLDA: 200 factors
• No calibration
Baseline Results: impact of mutual conditions
5 s 10 s 20 s 40 s full
0.01
0.10.2
0.512
5
Duration
Cllr
5 s 10 s 20 s 40 s full
0.01
0.10.2
0.512
5
Duration
Cllr
0 dB
5 dB
10 dB
15 dB
20 dB
Clean
Cllr
Cminllr
(a) AC noise (b) CROWD noise
Comparison of AS-norm approaches
0 5 10 15 20
0.1
0.3
0.5
0.7
SNR (in dB)
Cm
inllr
5 s 10 s 20 s 40 s full
0.1
0.3
0.5
0.7
Duration
Cm
inllr
(c) 10s/CROWD (d) CROWD-0 dB
no norm uAS (matched cohorts)cAS uAS (automatic cohort selection)
Automatic cohort pre-selection
200 400 600 800
10
20
30
40
50
Cohort speaker ID
Con
dition
(e) Unique selections on 10s/CROWD-0 dB
Cohort pre-selection by:
• Similar conditions
• SNR more relevant than duration
Analysis: i-vector pool mean shift
• Multi-variate Student t-test:
t2 =nx nynx + ny
(~̄x− ~̄y)′W−1(~̄x− ~̄y),with:
W =
∑nx
i=1 (~xi − ~̄x)(~xi − ~̄x)′ +∑ny
i=1 (~yi − ~̄y)(~yi − ~̄y)′
nx + ny − 2
• P-values by cdf F of χ2 (D = 200):
t2 ∼ χ2D, p = 1− Fχ2
D(t2)
10 20 30 40 50
10
20
30
40
50
Condition
Con
dition
0
5 e3
10 e3
15 e3
20 e3
(f) t2 scores of Hotelling’s T 2-test
⇒ High significance: p
{= 1 same condition: same mean i-vector< 10−13 cross condition: different mean i-vector
Conclusion
• Mutual effects with performance impactsby log-duration and log-SNR
• Quality-based cohort pre-selection yieldssignificant gains
• Performance gap between degraded andnon-degraded samples still as open chal-lenge
• Means of cross-condition i-vectors differ
Acknowledgements
We would like to thank the I4U consortium fordatabase sharing.
This work has been partially funded by the Cen-ter for Advanced Security Research Darmstadt(CASED), the Hesse government (project no.467/15-09, BioMobile) and the Academy ofFinland (project no. 256961 and 284671).