1 best practices for voice authentication charles r. jankowski jr., ph. d. speechtek west 2007...
TRANSCRIPT
1
Best Practices for Voice Authentication
Charles R. Jankowski Jr., Ph. D.SpeechTek West 2007February 21, 2007
2
Design for Voice Authentication
Dialog design is the single most critical aspect of voice authentication performance
3
Make Verifier SECUREMake Verifier SECURE
Make Verifier EASY to useMake Verifier EASY to use
Verifier Dialog and Application Design
4
Verifier Performance Measures
Automation Rate– Percentage of repeat true speakers automated by system– 100 – False Reject Rate
• Security Rate– 100 – False Accept Rate
5
Baseline Performance
Automation Rate
94% 99%93%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Very High High MediumSecurity Level
9
Summary of design takeaways
1. Maintain consistency between enrollment and verification dialogs
2. Collect enough data during enrollment and verification
3. Use variable length verification
4. Use knowledge verification
5. Use speaker adaptation
10
1. Maintain consistency between enrollment and verification
SSN/Account#
Authentication
1-9
1-9
1-9
KV selection
SSN/Account#
1-9
1-9
KV
• Only using 1-9 for voiceprint creation and checking
• Could also use account number, phone number, etc…
11
Choice of Verification Utterance
• Why use identity claim?– One shorter dialog step– 80% of users claim and verify in one step
• Why use common?– Many users do not like to speak their identity claim in public
• Identity theft• Allow option to touch-tone identity claim
– Multiple options for Identification• Enrollment very complex and long
– Common utterance performance more predictable• Out-of-the-box performance better understood• ID claim may vary in length
– Evaluating performance with identity claim more complex• Impostors MUST speak same text as enrollee• Identity claim -> Need live impostors
12
2. Collect enough data during enrollment and verification
SSN/Account#
Authentication
1-9
1-9
1-9
KV selection
SSN/Account#
1-9
1-9
KV
• 3 Enrollment utterances instead of 2: 1-3% lower automation
• 1 Verification utterance instead of 2: 1-3% lower automation
13
3. Use variable length verification
SSN/Account#
Authentication
1-9
1-9
1-9
KV selection
SSN/Account#
1-9 80%
1-9 10%
• Same verification performance as with 2 required utterances
14
4. Use knowledge verification
SSN/Account#
Authentication
1-9
1-9
1-9
KV selection
SSN/Account#
1-9
1-9
80%
10%
KV 10%
• Examples: PIN, Secret Date
• Verifier not turned on
• Increases Automation
• Increases Effectiveness of Adaptation
15
5. Use speaker adaptation
SSN/Account#
Authentication
1-9
1-9
1-9
KV selection
SSN/Account#
1-9
1-9
KV
16
Speaker Verification Performance
• Speaker Adaptation– Online: Score very well, adapt and get even better
– Manual: Unsure, backoff to knowledge, adapt if knowledge passes
• Cross-channel
Digit Voiceprints
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
40.00%
0 2 4 6 8 10 12
Number of adaptation calls
Fa
lse
re
jec
t ra
te (
FR
)
Online all data
Online same channel only
Manual all data
Manual same channel only