Difference between revisions of "Lantian Li 14-10-27"

From cslt Wiki
Jump to: navigation, search
(以“Weekly Summary 1. SVM training data: each true speaker test with the top N imposter making up the training data. N = 2,5,10. 2. Four types of scoring domain featur...”为内容创建页面)
 
 
Line 3: Line 3:
 
1. SVM training data: each true speaker test with the top N imposter making up the training data. N = 2,5,10.
 
1. SVM training data: each true speaker test with the top N imposter making up the training data. N = 2,5,10.
  
2. Four types of scoring domain feature: 1).system score; 2). tnorm score; 3).system score + cohort scores; 4).system score + cohort score + detal scores;
+
2. Five types of scoring domain feature: 1).system score; 2). tnorm score; 3).system score + cohort scores; 4).system score + cohort score + detal scores;
  
 
5).system score + cohort score + detal scores + tnorm score;
 
5).system score + cohort score + detal scores + tnorm score;

Latest revision as of 12:27, 27 October 2014

Weekly Summary

1. SVM training data: each true speaker test with the top N imposter making up the training data. N = 2,5,10.

2. Five types of scoring domain feature: 1).system score; 2). tnorm score; 3).system score + cohort scores; 4).system score + cohort score + detal scores;

5).system score + cohort score + detal scores + tnorm score;

3. For a given test set, the results show that for linear SVM, the EER of 3)/4) is similar and a little better than 2). and 5) gets the best performance.

However, due to the test set is relatively small, the subsequent validation should be required.

Another phenomenon is that there exists overfitting problem using 'rbf' and 'poly' kernel. The training effect is very good while the test result is so bad.

Next Week

1. Additional experiments shoule be done to prove the effectiveness of the method.

2. Try to analyse the overfitting problem and solve it.