Revolution in Accuracy and Speedof Voice Biometrics
Seminář AFCEAPraha, 17. 1. 2019
About Phonexia
Turning voice to knowledge!
19+ years inspeech processing
12+ years of commercial
activities
4
Worldwideactivity
Private companyfully owned
by 5 founders
50+ people
5
60+
And many more…
Our Clients and Partners
Strategic research partner
Brno University of Technology, Faculty of Information Technology, Speech@FIT research group
Success in U.S. NIST evaluation, teamed together with Brno University of Technology
A complete set of state-of-the-art speech technologies in a single software platform
9
Phonexia Speech
Platform
Speech Transcription
Speaker Identification
Language Identification
Diarization
Age Estimation
Speech Quality Estimation
Voice Activity Detection
Keyword Spotting
Denoiser
Gender Identification
PHONEXIASPEECH PLATFORMFOR GOVERNMENT
man
English
20-30yearsold
extremism
jewish
nine elevenkilledwar in afghanistan
Keyword Spotting
Speech Transcription
islam
Speaker Identification
Speaker Score
95,7%Mohammed
Emwazi
GovernmentUse Cases
Different Users, Different Needs
StrategicMI/CI
SIGINTCOMINT
Massive interception
National security
TacticalMI/CI
TacticalSIGINT
Localinterception
Immediate local threats
Archive SearchLEA/Police
Organizedcrime
Anti-narcotics
Investigation
ForensicForensic Labs
Courtcases
Quantifiedevidences
Detailed reports
INPUT AUDIO STREAM
14
VAD VoiceActivity Detection
SQE SpeechQuality Estimator
GID GenderIdentification
LID LanguageIdentification
SID SpeakerIdentification
SPEAKERSOF MY
INTEREST
Example: Strategic Speaker Spotting
1 000 000 calls per day
100 calls per day
CaseStudies
16
Speaker Identification for OSINTSpeaker Identification - Use Case Demo
17
Explore relations among people using several phones
Additional data for investigation
Link Analysis
18
Speaker Clustering
Biometrics
How?
Revolution in Voice Biometrics in 2018
Voice
iVectorPHONEXIA
DEEP EMBEDDINGS™23
PAST NOW
SPEED (FTRT) 5
EER ON NIST16 DATA SET 13%
2.07RAM CONSUMPTION (GB)
25
7%
0.29
5 times fasterthan iVector
double accuracythan iVector
7 times lowerRAM consumption
DeepEmbeddingsTMiVector
24
Hardware SavingsPAST NOW
25
From 5 to 1 server.
Measurements
26
NOW
PAST
NET SPEECH (s)
EQUA
L ER
ROR
RATE
(%)
Team EfficiencyPAST NOW
27
• Less manual job• More time for other
analytics tasks.
Higher Percentage of Resolved CasesPAST NOW
28
Revealing most of the network.
Key Takeaways
Complete portfolio of Speech Technologiesin a single software platform.
Rapid increase of accuracy and speedof the new generation of Speaker Identification technology.
Unlock new possibilitiesfor security and defense use cases.
Thank youfor your attention
Tomáš BiaHead of Product
T +420 605 279 125 E [email protected]
phonexia.com