VBS Documentation and Implementation
The full standard initiative is located atwww.voicebiometry.orgQuick descriptionStandard manual with detailed description and a quick user guide to…The reference demo packageContains full speaker-recognition (demo) pipeline
i-vectors
Information-richLow-dimensionalFixed-lengthVector of real numbersBased on statistical modelEasy to compareEasy to storeNot recoverable to speech
Dehak, N.,et al., Support Vector Machines versus Fast Scoring in the Low-Dimensional Total Variability Space for Speaker Verification In Proc Interspeech 2009, Brighton, UK, September 2009
I-VECTOREXTRACTOR(SITE 1)
COMPARISON
AUDIO 1
I-VECTOREXTRACTOR(SITE 2)
AUDIO 2
SCORE
SPEAKER IDENTITYNO SPEECH CONTENT
SPEAKER IDENTITYAND SPEECH CONTENT
i-vector
i-vector
VOICEACTIVITYDETECTION
I-VECTOREXTRACTION
POST-PROCESSING
COMPARISON
FEATUREEXTRACTION
COLLECTIONOFSTATISTICS
TMATRIX
GMMUBM
AUDIO
ALGORITHMS
PARAMETERS
STANDARDIZED
NOTSTANDARDIZED
Standardization objectives
Acoustic feature extractioni-vector extraction algorithmi-vector extraction parameters(GMM parameters,i-vector extractor parameters)The data exchangeformats(tuned for telephone speech)
Feature Extraction
Pre-emphasis25ms windowing with 10ms shiftHamming window24 Mel filter-banks in the range of 125 – 3800 Hz19-dimensional MFCC coefficients + C0Delta + Double-deltaShort-timeCepstralMean and Variance NormalizationOver 3 second window
Universal Background Model
2048 Gaussian mixture componentsDiagonalcovariancesTrained on 1156 hours of the NIST SRE 2004-2008 data (gender independent)Trained using gradual Gaussian splitting with 10 EM steps in each splitThe UBM is used to extract the sufficient statistics for thei-vector extractor and to normalize (whiten) these statistics
i-vector extractor
Trained on the same data as UBM + Switchboard 2 (phases 2 and 3) + Fisher English (phases 1 and 2)600 dimensional10 iterations of EM and MD steps
Reference implementation
Reference implementation
Python codeReadability and ease of understandingExtensibilityStandard Python packagesNumpy+Scipy
Systemperfofmance
NISTSRE 2010,cond5, female
i-vector compatibility
i-vectors produced by one system are incompatible with those generated by a different systemWe run experiments for training ani-vector transformation to migratei-vectors of one systems to anotherTake it as an invitation for tomorrow’s talk:“Migratingi-vectors Between Speaker Recognition Systems UsingRegression Neural Networks”
THANK YOU
0
Embed
Upload