Difference between revisions of "Lantian Li 15-08-11"

From cslt Wiki
Jump to: navigation, search
(以“Weekly Summary 1. Prepare for three deep speaker embedding tasks: 1). large-scale deep speaker vector framework -- hold. 2). knowledge transfer for i-vector -- in...”为内容创建页面)
 
 
Line 15: Line 15:
 
We think it is due to the training-data is insufficient.  
 
We think it is due to the training-data is insufficient.  
  
3). derive binary i-vectors using Hamming distance learning
+
3). derive binary i-vectors using Hamming distance learning (identification efficiency Exp.)
  
 
obtain the code from Xing Chao, and start to modify the test-code and design the experiments.
 
obtain the code from Xing Chao, and start to modify the test-code and design the experiments.

Latest revision as of 08:19, 11 August 2015

Weekly Summary

1. Prepare for three deep speaker embedding tasks:

1). large-scale deep speaker vector framework -- hold.

2). knowledge transfer for i-vector -- in process.

The ivector knowledge transfor(IKT) both on DNN and RNN has been done.

The results show that the performance of DNN-IKT is better than former d-vector system,

but still worse than i-vector baseline. And the RNN-IKT does not work well.

We think it is due to the training-data is insufficient.

3). derive binary i-vectors using Hamming distance learning (identification efficiency Exp.)

obtain the code from Xing Chao, and start to modify the test-code and design the experiments.

Next Week

1. Go on the task 1 and repeat the DNN-IKT of 2) on Fisher and SRE database.

2. Go on the task 1 and attempt to complete 3).