Download presentation
Presentation is loading. Please wait.
1
Decision Making Based on Cohort Scores for
Speaker Verification Lantian Li CSLT / RIIT, Tsinghua University Co-work with Renyu Wang, Caixia Wang and Thomas Fang Zheng IEEE APSIPA ASC, Dec , 2016 APSIPA presentation 1/13/2019 APSIPA ASC 2016
2
Outline Introduction Cohort-based decision making framework
A single-score decision making Solution by a multi-score decision making Cohort-based decision making framework Cohort selection Feature design Discriminative model training Experiments Conclusions 1/13/2019 APSIPA ASC 2016
3
Introduction Speaker recognition Decision making
The single-score decision is simple and efficient Quite sensitive to variations (score variation) Text contents, channel, speaking styles. Difficulty in choosing an appropriate threshold Error-pron decision In a typical GMM-UBM system, the score is often computed as the log likelihood ratio that the test utterance being generated from the GMM of the claimed speaker and the UBM. Leads to the err-pron decsion 1/13/2019 APSIPA ASC 2016
4
Introduction Score normalization techniques
Bayes’ theorem (Z-norm, T-norm, etc.) Cohort normalization Cohort replaces the UBM The alternative hypothesis more accurately It is also simply averaged to normalize the target score. Still a single-score approach 1/13/2019 APSIPA ASC 2016
5
Motivations Our idea A new cohort approach
Cohort normalization is not just a mean average. Cohort scores: distributions, ranks, spanning areas, etc. A new cohort approach Decision on the whole cohort sets Employ a powerful discriminative model A true and reliable multi-score decision making 1/13/2019 APSIPA ASC 2016
6
Cohort-based decision making framework
Cohort selection How to select a cohort for each claimed speaker. Feature design How to fully use these cohort scores. Discriminative model training How to build a more powerful decision model. Add three parts in the typical GMM-UBM system. 1/13/2019 APSIPA ASC 2016
7
Cohort-based decision making framework
Cohort selection Vector quantization (VQ) K-L distance Minimize the within-class cost stopping criterion 1/13/2019 APSIPA ASC 2016
8
Cohort-based decision making framework
Feature design Distribution of scores Score normalization Rank position Sorted score differences … The distribution of scores between target test and imposter test is different. 1/13/2019 APSIPA ASC 2016
9
Experiments Database (‘CSLT-DSDB’: all recordings is the text-prompted digit strings.) Training set: 200 females and 200 males for UBM training. Development set: 145 speakers including 280 enrollment and 2,874 test utterances. Cohort selection and feature design. Evaluation set: 92 speakers including 1,220 target trials and 111,020 non-target trials. Experimental setups 13-dim MFCCs 256 Gaussian components ∆ ∆∆ 1/13/2019 APSIPA ASC 2016
10
Experiments Two cohort selection criterion Global clustering
Sibling speakers Nearest neighbor 1/13/2019 APSIPA ASC 2016
11
Experiments Feature design Rank position Sorted score differences
1/13/2019 APSIPA ASC 2016
12
Experiments Discriminative model training SVM-based scoring
NN-based scoring 1.621 (Baseline) 1.621 (Baseline) ‘norm’ is the standard T-norm methods, ‘r-pos’ is the ‘Rank position’, ‘s-diff’ is the ‘Sorted score differences’. 1/13/2019 APSIPA ASC 2016
13
Conclusions Decision making method Future work
Distribution of Cohort scores Design score-level features (‘sorted score differences’) Powerful discriminative models Stable and better performance than GMM-UBM baseline Future work Feature designing and cohort selection 1/13/2019 APSIPA ASC 2016
14
Thank you lilt.cslt.org IEEE APSIPA ASC Dec. 13-16, 2016, Jeju, Korea
1/13/2019 APSIPA ASC 2016
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.