Component Score Weighting for GMM based TextIndependent Speaker Verification Liang Lu

1
Component Score Weighting for GMM based
Text-Independent Speaker Verification
Liang Lu
luliang07_at_gmail.com

• SNLP Unit, France Telecom RD Beijing
• 2008-01-21

2
Outline

• Introduction
• Conventional LLR and Motivation for detailed
  score processing
• Component Score Weighting
• Experimental Results
• Conclusion

3
Introduction

• State of the art GMM-UBM framework
• GMM based model construction
• Log-likelihood Ratio (LLR) based decision making
• Score Normalisation (Tnorm, Hnorm, etc) for
  robustesses

4
Introduction

• Major challenges
• Limited data for speaker model training
• Mismatch between training and testing data

5
Motivation for Component Score Weighting

• Motivation
• The insufficiency of training data and mismatch
  between training and testing condition make the
  mixtures in GMM different in discriminative
  capability
• The LLR just sum the score of each mixture
  without considering its reliability
• Does it helpful if LLR considers the
  discriminative capability of each mixture?

Question If it does, how to explore the
discriminative capabilities of Gaussian
Component Mixtures
6
Component Score Weighting

• Our Method
• First, scatter the LLR to each Gaussian mixture
• Where, the k-th mixture is dominant for frame ,
  namely,

Let we call is the dominant score and is
the residual score
7
Component Score Weighting

• Extend the original LLR
• After doing this, the original LLR will be
  spitted into two score serials, dominant score
  serial and residual score
  serial
• Original
• If we consider the discriminative capacity of
  each Gaussian mixture
• Extended

in original LLR
8
Component Score Weighting

• Now the question is
• How can we know the discriminative capability
  of each Gaussian mixture and what the
  should be?

Our assumption We believe that the high
dominant scores will have better discriminative
capability and should be highlighted.
9
Component Score Weighting
Why the high dominant scores?

• If the test utterance is from the target
  speaker, then more components in GMM should get
  high value compared with UBM.
• If the utterance is form imposter, then
  high-valued components in GMM are hardly more
  UBM.
• If the test utterance is from the target
  speaker, the low-valued components in GMM is due
  to the mixtures are not well trained or mismatch
  exists between training and testing data.

10
Component Score Weighting
We simply used an exponential function as the
weighting function The residual scores have
little importance and we ignore them finally. The
final LLR score is as follows
Restrained
Emphasized
11
Experimental Results
Table Results for GMM baseline and GMM with
Component Score Weighting with TNorm
Experiments are performed in the 1conv4w-1conv4w
task of the 2006 NIST SRE corpora
12
Conclusion

• Split the LLR score and consider the
  discriminative capacity of Gaussian mixtures is
  helpful to cope with the insufficiency of
  training data and mismatch between training and
  testing condition.
• The score weighting function should be
  coincident with the component score distribution
  and discriminative capacity.
• The exponential weighting function used in this
  investigation is not universal and also may not
  optimal. More work is needed to explore an
  optimal weighting function.

13
Thanks