Difference between revisions of "Lantian Li 15-06-02"

From cslt Wiki
Jump to: navigation, search
(以“Weekly Summary 1. Make preliminary experiments on Speech watermarking. 2. Prepare for three deep speaker embedding tasks: 1). large-scale deep speaker vector fr...”为内容创建页面)
 
 
Line 3: Line 3:
 
1. Make preliminary experiments on Speech watermarking.
 
1. Make preliminary experiments on Speech watermarking.
  
2. Prepare for three deep speaker embedding tasks:
+
Results show that with the increase of the number of interpolation points, system performance will gradually decline.
  
  1). large-scale deep speaker vector framework: complete the first (i-vector clusters) and second (phone posterior extraction) steps.
+
2. Make experiments of the discriminative scoring training.  
  2). RNN based deep speaker vector: use kaldi-RNN tools to embed utterance-level deep speaker vector.
+
  3). derive binary i-vectors using Hamming distance learning. complete the fisrt step (data preparation).
+
Next Week
+
  
1. Go on the task 1 and plan to complete it next week.
+
Results show that no matter the traditional system scoring method or SVM/DNN classifier, the performance of raw-score is almost the same.
 +
 
 +
It proves that these selected features are effective and available.
 +
 
 +
3. Prepare for three deep speaker embedding tasks:
 +
 
 +
  1). large-scale deep speaker vector framework: In process.
 +
 
 +
  2). RNN based deep speaker vector: use kaldi-RNN tools to embed utterance-level deep speaker vector.
 +
 
 +
  Results show that the r-vector is not better than d-vector.
 +
 
 +
  3). derive binary i-vectors using Hamming distance learning. In process.
 +
 
 +
Next Week
  
2. Go on the task 2 and learn the RNN framework.
+
1. Go on the task 3.

Latest revision as of 08:40, 3 June 2015

Weekly Summary

1. Make preliminary experiments on Speech watermarking.

Results show that with the increase of the number of interpolation points, system performance will gradually decline.

2. Make experiments of the discriminative scoring training.

Results show that no matter the traditional system scoring method or SVM/DNN classifier, the performance of raw-score is almost the same.

It proves that these selected features are effective and available.

3. Prepare for three deep speaker embedding tasks:

 1). large-scale deep speaker vector framework: In process.
 2). RNN based deep speaker vector: use kaldi-RNN tools to embed utterance-level deep speaker vector.
 Results show that the r-vector is not better than d-vector.
 3). derive binary i-vectors using Hamming distance learning. In process.
 

Next Week

1. Go on the task 3.