Difference between revisions of "09-30 Lantian Li"

From cslt Wiki
Jump to: navigation, search
(以“1. To go on studying a scoring method on GMM-UBM aiming to design a cohort reference speaker models. 1). Implement the K-means algorithem to cluster the training se...”为内容创建页面)
 
Line 1: Line 1:
 
1. To go on studying a scoring method on GMM-UBM aiming to design a cohort reference speaker models.
 
1. To go on studying a scoring method on GMM-UBM aiming to design a cohort reference speaker models.
  
1). Implement the K-means algorithem to cluster the training set in order for organizing the cohort set.
+
1). On the basis of testing on a small scale, a total-batch experiment was made.
  
2). Make the verfication score results dividied into four parts. --"Real True Speaker"/"Sensitive True
+
The result for "true speaker" shows that there exists a score gap between the real-speaker and cohort
  
Speaker"/"Sensitive Imp Speaker"/"Abosulte Imp Speaker".
+
speaker. Besides, the variance from the real-true-speaker is larger than the sen-true-speaker based on the
  
3). Re-score for the four parts on the cohort set.
+
cohort set. For imposter-speaker having the similar result, However, it does not meet the original hypothesis
  
4). Score ranking for each part and draw score-rank distrubution diagrams.
+
that if the speaker is not the true speaker, it may have the similar scoring between this hypothesis
 +
 
 +
speaker model and corhort models. But experimental result shows that there still has a gap.
  
 
Next Week
 
Next Week

Revision as of 02:32, 29 September 2014

1. To go on studying a scoring method on GMM-UBM aiming to design a cohort reference speaker models.

1). On the basis of testing on a small scale, a total-batch experiment was made.

The result for "true speaker" shows that there exists a score gap between the real-speaker and cohort

speaker. Besides, the variance from the real-true-speaker is larger than the sen-true-speaker based on the

cohort set. For imposter-speaker having the similar result, However, it does not meet the original hypothesis

that if the speaker is not the true speaker, it may have the similar scoring between this hypothesis

speaker model and corhort models. But experimental result shows that there still has a gap.

Next Week

1. Go on the task1 to explore the inherent law of re-scoring results and use the cohort set to

reduce the error rate on the "Sensitive True Speaker"/"Sensitive Imp Speaker".