Voice Biometrics - how to recognize a speaker.

Voic

ePIN

.com

ul.K

raku

sa 1

1,

30-5

35

tel: 7

26 5

03 4

03

mail:in

fo@

voic

epin

.com

Voice Biometrics

No sensors required

Various solution scenarios

Cheap

Comfortable

Natural

Known basic applications:

Identity verification

• access controll

• biometric PIN

• password reset (30%-40% cases)

Identity identification

• fraud detection

• service personlization

• other

HybridPrompted

Textindependent

Textdependent

How voice authentication works ?

Voice signal

Voice signalspectrogram

Voice signalsmoothed spectrogram

Voice features MFCCMel-Freq Cepstral Coeff.

50 100 150 200 250 300 350 400

2

4

6

8

10

Performance• FRR - False Rejection Rate, % • FAR - False Acceptance Rate, % • EER - Equal Error Rate, % when the decision treshold fixed as to assure FAR = FRR.

• Accuracy - %, percentage of (any) correctdecisions depends on the evaluationscenario and decision threshold settings

• Example of the performance specificationFAR < 0.1%, and FRR< 5%

• Statistical significance: How to assess the risk of rarely

occuring phenomena ?

Security vs usability issue

% errors

Decision threshold

Attacks (FAR)Rejections(FRR)

EER

password, 123, 0000, love

NA7;zSrluz, Mj[LAX}i]O, 9622535008, 594772359571

Evaluation scenario

Ellen SierraVoiceprint

False acceptances % (succesful attacks)

False rejections % (unsuccesful genuine verification attempts)

Decision score (-100, 100)

Learning Methods

Statistical - GMM

SVM – Supervectors

Factor Analysis (i-Vectors)

Deep learning (DNN)

GMM UBM MAP Framework

GMM MAP Adaptation

Supervector Factor Analysis

Supervector Factor Analysis

Deep Learning

Deep Learning results – DET plot

[email protected]+48 781 503 403