Lantian Li 14-12-01 - 版本历史

2014年12月1日 (一) 11:33 Lilt

2014-12-01T11:33:38Z

2014年12月1日 (一) 11:33 Lilt

2014-12-01T11:33:20Z

Lilt：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

2014-12-01T11:21:59Z

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

新页面

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 is small.

2/ This issue based on GMM-UBM is not applied to complex non-linear model.

2. Compute the training accuarcy. For true speaker, the training accuray is about 4%, and for imp speaker, it is about 1%.

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

1. Continue to look for distinguishing characteristics

1) Improve K-means algorithm.

2) Implement the UBM segmentation score method.

3) Add original GMM score to feature vector.

@@ 第3行： / 第3行： @@
 . Model Cluster: three ways to measure the distance between two models.
-. Explore the different score method between UBM minus GMM or GMM minus UBM, and the performance EER
+. Explore the different score methods between UBM minus GMM or GMM minus UBM, and the performance EER
 shows that GMM minus UBM is a bit better.

@@ 第1行： / 第1行： @@
 Weekly Summary
-. Compare the performance between SVM and MLR, and the result is that MLR is worse than SVM.
+. Model Cluster: three ways to measure the distance between two models.
-I think there are two reasons. 1/ the training dataset is small.
+. Explore the different score method between UBM minus GMM or GMM minus UBM, and the performance EER
-/ This issue based on GMM-UBM is not applied to complex non-linear model.
+shows that GMM minus UBM is a bit better.
-. Compute the training accuarcy. For true speaker, the training accuray is about 4%, and for imp speaker, it is about 1%.
+. 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.
-−
-. Help Jun Wang test the performance of PLDA-based classifier, results is baseline < SVM < DNN.
-−
-So I learn DNN method from him.
 Next Week
-. Continue to look for distinguishing characteristics
+. Using the phoneme results and lexicon to position each digit and segment each utterance.
-−
-) Improve K-means algorithm.
-) Implement the UBM segmentation score method.
+. Make UBM adaptation to get 9 digit-dependent UBMs.
-) Add original GMM score to feature vector.
+. Experiments this digit-dependent system.