View
14
Download
0
Category
Preview:
Citation preview
Voice Biometry standard proposal
Honza Černocký
Brno University of Technology,
BUT Speech@FIT,
Czech Republic
Sep 8th 2015, Interspeech VBS meeting
Program
Honza Cernocky – intro, “why?”
Ondrej Glembek – Technical description
Petr Schwarz – Phonexia remarks
Discussion
Honza Cernocky – next steps
End 16.00, no buffet, drinks, entertainment
BUT Speech@FIT Honza Cernocky 05/2015 2/56
Situation
• In the last 10 years, scientific advances in speaker recognition (JFA, iVectors, PLDA) allowed for producing precise and robust SRE systems
• Quickly adopted by vendors, producing solutions that are successful on the market.
• R&D never stopping
• Everyone continuously improving performance of their system, robustness, calibration, etc
• New versions of engines released
A vibrant community working in cooperative/competitive mode both for R&D labs and vendors.
BUT Speech@FIT Honza Cernocky 05/2015 3/56
It works
BUT Speech@FIT Honza Cernocky 05/2015 4/56
SIDScore, hard decision …
Pe
piV
ec
Pe
piV
ec
Pe
piV
ec
Pe
piV
ec
Score, hard decision …
Co
ca
iVe
c
Co
ca
iVe
c
Co
ca
iVe
c
Co
ca
iVe
c
SID
It does not work
BUT Speech@FIT Honza Cernocky 05/2015 5/56
SIDP
ep
iVe
c
Co
ca
iVe
c
Co
ca
iVe
c
Co
ca
iVe
c
SID
MISMATCH
Making it work
BUT Speech@FIT Honza Cernocky 05/2015 6/56
Co
ca
iVe
c
Co
ca
iVe
c
Co
ca
iVe
c
SIDScore, hard decision …
Co
ca
iVe
c
Making it really work – standardized iVectors
BUT Speech@FIT Honza Cernocky 05/2015 7/56
SID
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
SIDScore, hard decision …
Making it really work – standardized iVectors
BUT Speech@FIT Honza Cernocky 05/2015 8/56
SID
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
SID
Score, hard decision …
Making it really work – standardized iVectors
BUT Speech@FIT Honza Cernocky 05/2015 9/56
SID
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
VB
SiV
ec
SID
Score, hard decision …
The main thing
BUT Speech@FIT Honza Cernocky 05/2015 10/56
I-VECTOREXTRACTION(VENDOR 1)
COMPARISON
AUDIO 1
I-VECTOREXTRACTION(VENDOR 2)
AUDIO 2
SCORE
SPEAKER IDENTITYNO CONTENT
SPEAKER IDENTITYAND CONTENT
i-vector
i-vector
What is needed
• Fix the core iVector extraction algorithms
• Fix the necessary parameters
• Do the necessary minimum, let people freedom to use their (own, best) VAD and scoring.
• Do it well for the core condition – telephone, not trying to address everything.
BUT Speech@FIT Honza Cernocky 05/2015 11/56
We WANT
• Users
• Having interoperable systems
• Being able to exchange speaker information without compromising content
• within companies/agencies, across companies/agencies and across borders
• Vendors
• Increasing the whole market (think about introduction of USB!)
• R&D labs
• sharing iVectors between labs without lengthy discussions on configuration (not excluded though!)
• Giving a working recipe to juniors to play with.
• Obtaining massive data from the users
BUT Speech@FIT Honza Cernocky 05/2015 12/56
We DON’T WANT
• stop R&D (both academic and commercial) of speaker recognition technology by saying that this will be the only iVector extraction scheme forever.
• all of us are trying to push the field further, sometimes as collaborators, sometimes as competitors.
• We want to define a snap-shot of the best practice up to day on which we could agree.
• Earn money on licenses or patents – the proposed standard is license and patent-free
• Have something too complex and too relying on a proprietary and/or 3rd party technology.
• Present this as an ultimate forensic solution.
BUT Speech@FIT Honza Cernocky 05/2015 13/56
What is there
• http://voicebiometry.org/ - technical description, Python code with all necessary parameters (feature extraction, UBM, T-matrix)
• Google group http://groups.google.com/d/forum/voice-biometry-standard - please subscribe
BUT Speech@FIT Honza Cernocky 05/2015 14/56
Program
Honza Cernocky – intro, “why?”
Ondrej Glembek – Technical description
Petr Schwarz – Phonexia remarks
Discussion
Honza Cernocky – next steps
End 16.00, no buffet, drinks, entertainment
BUT Speech@FIT Honza Cernocky 05/2015 15/56
Program
Honza Cernocky – intro, “why?”
Ondrej Glembek – Technical description
Petr Schwarz – Phonexia remarks
Discussion
Honza Cernocky – next steps
End 16.00, no buffet, drinks, entertainment
BUT Speech@FIT Honza Cernocky 05/2015 16/56
Next steps
• If interested, sign-up to the google-group:
• http://groups.google.com/d/forum/voice-biometry-standard (no more personal emails).
• take the code and test it on your data
• Report anything that you'd like to improve.
• Please bug-fixes, not complete changes …
• To the g-group or personally to Ondraglembek@fit.vutbr.cz
• Tell us if we can add your lab/company as supporter on the web-page.
• Please attach a logo in reasonable resolution and a web-link.
• You might need to consult your management.
• Vendors: implement it to your systems
BUT Speech@FIT Honza Cernocky 05/2015 17/56
Ext steps II.
• The real normalization (ISO/IEC, NIST, W3C …)
• Yes, but only if it has wide industrial and academic support.
• Will need help …
BUT Speech@FIT Honza Cernocky 05/2015 18/56
BUT Speech@FIT Honza Cernocky 05/2015 19/56
Thank you for your attention !
http://voicebiometry.org/
Recommended