Title: Benchmarking Feature Selection Techniques on the Speaker Verification Task
1Benchmarking Feature Selection Techniques on the
Speaker Verification Task
- T. Ganchev, P. Zervas, N. Fakotakis, G.
Kokkinakis - Wire Communications Laboratory,
- University of Patras, Greece
2Presentation Outline
- Problem definition
- Approaches for feature evaluation
- Feature ranking techniques
- Feature selection
- Experimentations
- Conclusion
3Problem Definition
- Speaker verification is a typical detection task
- Does the speaker possess the identity s/he
claims? - Our problem
- The speech features contribute differently for
speaker differentiation (i.e. have different
discrimination power) - Our target
- To identify and keep all potentially relevant
features
4Approaches for Feature Evaluation
- Filters act on pre-processing step
- Wrappers utilize the classifier to score
subsets of features according to their
discriminative power - Embedded methods perform feature selection
during the training process
5Feature ranking techniques 1/4
- Information Gain (IG) attribute evaluation,
- Gain Ratio (GR) attribute evaluation,
- Symmetrical Uncertainty (SU) attribute
evaluation, - Correlation-based Feature Selection (CFS),
- Support Vector Machine Recursive Feature
Elimination (SVM-RFE).
6Feature ranking techniques 2/4
The Entropy of Y is defined as
The Entropy of Y after observing X is
7Feature ranking techniques 3/4
- Information Gain (IG) ? mutual information
- Symmetrical Uncertainty (SU)
8Feature ranking techniques 4/4
- Correlation-based Feature Selection
- SVM Recursive Feature Elimination
train the SVM classifier by optimizing the
weights w.r. the cost function
compute the ranking criterion for all features
remove the feature with smallest ranking
criterion
9Feature selection
- IG, GR, and SU does not perform feature selection
but only feature ranking, - we experimented with the forward selection and
bi-directional search. - In the case of CFS, we used genetic search of the
feature space. - Finally, we performed full search to find out the
appropriate number of attributes.
10Experimental Setup
- The WCL-1 speaker verification system
- PNN-based expert for each target user
- Speech parameterization
- logarithm of fundamental frequency,
- logarithm of energy,
- 31 Mel-Frequency Cepstral Coefficients (where the
0-th coefficient is excluded) - The 2001 NIST SRE database
- 74 male speakers, 2 minutes of training data
- various transmission channels TDMA, CDMA,
Cell, GSM, Land - various length of trials 00-15, 16-25,
26-35, 36-45, and 46-60 sec - different environmental conditions inside,
outside, vehicle
11Experimental results Feature ranking
- Feature ranking
- IG - Information Gain,
- GR - Gain Ratio,
- SU - Symmetrical Uncertainty,
- CFS - Correlation-based Feature Selection,
- SVM-RFE - Support Vector Machine Recursive
Feature Elimination, - Ref - Genuine feature set.
12Experimental results All experiments
13Experimental results SV performance
- Comparison among the most successful feature sets
for the evaluated feature ranking methods.
14Conclusion
- We evaluated the relevance of a set of speech
features for the task of speaker verification. - Feature subsets obtained from various ranking
techniques were compared - Useful knowledge was gained regarding features
that do not contribute significantly to the task.
- Several subsets were observed to provide a
competitive performance, which opens ways to
lower the computational and memory demands
15Comments and Questions
Thank you for your attention ! Any Comments or
Questions will be welcomed...