Difference between revisions of "Lantian Li 14-12-01"

From cslt Wiki
Jump to: navigation, search
(以“Weekly Summary 1. Compare the performance between SVM and MLR, and the result is that MLR is worse than SVM. I think there are two reasons. 1/ the training dataset...”为内容创建页面)
 
Line 1: Line 1:
 
Weekly Summary
 
Weekly Summary
  
1. Compare the performance between SVM and MLR, and the result is that MLR is worse than SVM.
+
1. Model Cluster: three ways to measure the distance between two models.
  
I think there are two reasons. 1/ the training dataset is small.
+
2. Explore the different score method between UBM minus GMM or GMM minus UBM, and the performance EER
  
2/ This issue based on GMM-UBM is not applied to complex non-linear model.
+
shows that GMM minus UBM is a bit better.  
  
2. Compute the training accuarcy. For true speaker, the training accuray is about 4%, and for imp speaker, it is about 1%.
+
3. With the help of Z.-Y Zhang, using DNN-Decoder to decode the phoneme of each digital utterance.
 
+
The EER is 2%. So there exists a difference between the true traning accuracy and imp training accuracy.
+
 
+
Now I still don't know whether to need to adjust the training dataset.
+
 
+
3. Help Jun Wang test the performance of PLDA-based classifier, results is baseline < SVM < DNN.
+
 
+
So I learn DNN method from him.
+
  
 
Next Week
 
Next Week
  
1. Continue to look for distinguishing characteristics
+
1. Using the phoneme results and lexicon to position each digit and segment each utterance.
 
+
1) Improve K-means algorithm.
+
  
2) Implement the UBM segmentation score method.
+
2. Make UBM adaptation to get 9 digit-dependent UBMs.
  
3) Add original GMM score to feature vector.
+
3. Experiments this digit-dependent system.

Revision as of 11:33, 1 December 2014

Weekly Summary

1. Model Cluster: three ways to measure the distance between two models.

2. Explore the different score method between UBM minus GMM or GMM minus UBM, and the performance EER

shows that GMM minus UBM is a bit better.

3. With the help of Z.-Y Zhang, using DNN-Decoder to decode the phoneme of each digital utterance.

Next Week

1. Using the phoneme results and lexicon to position each digit and segment each utterance.

2. Make UBM adaptation to get 9 digit-dependent UBMs.

3. Experiments this digit-dependent system.