“Deep Generative Factorization For Speech Signal(ICASSP21)”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
第44行: 第44行:
 
   NF  |  0.013  |  0.410  |  0.397  ||  0.612  |  0.489  |  -0.123  
 
   NF  |  0.013  |  0.410  |  0.397  ||  0.612  |  0.489  |  -0.123  
 
   DNF  |  0.013  |  0.619  |  0.606  ||  0.612  |  0.335  |  -0.277   
 
   DNF  |  0.013  |  0.619  |  0.606  ||  0.612  |  0.335  |  -0.277   
   f-DNF |  0.013  |  0.636  |  0.623  ||  0.612  |  0.536  |  -0.076   
+
   f-DNF |  0.013  |  <b>0.636</b>   |  <b>0.623</b>   ||  0.612  |  <b>0.536</b>   |  <b>-0.076</b>  
 
-----------------------------------------------------------------
 
-----------------------------------------------------------------
  
第50行: 第50行:
 
                         <b>Speaker Manipulation</b>
 
                         <b>Speaker Manipulation</b>
 
   Model |  <i>p(s<sub>2</sub>|x)</i> | <i>p(s<sub>2</sub>|x')</i> |  <i>&delta;(s<sub>2</sub>)</i>  ||  <i>p(q|x)</i> |  <i>p(q|x')</i> |  <i>&delta;(q)</i>
 
   Model |  <i>p(s<sub>2</sub>|x)</i> | <i>p(s<sub>2</sub>|x')</i> |  <i>&delta;(s<sub>2</sub>)</i>  ||  <i>p(q|x)</i> |  <i>p(q|x')</i> |  <i>&delta;(q)</i>
   VAE  |  0.013 |  0.312   |  0.299   ||  0.612 |  0.454   |  -0.158
+
   VAE  |  0.010 |  0.303   |  0.293   ||  0.520 <b>0.509</b>   |  <b>-0.011</b>
   NF  |  0.013 |  0.410   |  0.397   ||  0.612 |  0.489   |  -0.123
+
   NF  |  0.010 |  0.435   |  0.425   ||  0.520 |  0.484   |  -0.036
   DNF  |  0.013 |  0.619   |  0.606   ||  0.612 |  0.335   |  -0.277  
+
   DNF  |  0.010 |  0.700   |  0.690   ||  0.520 |  0.349   |  -0.171  
   f-DNF |  0.013 |  0.636   |  0.623   ||  0.612 |  0.536   |  -0.076  
+
   f-DNF |  0.010 <b>0.710</b>   |  <b>0.700</b>   ||  0.520 |  0.503   |  -0.017  
 
-----------------------------------------------------------------
 
-----------------------------------------------------------------
  

2020年10月23日 (五) 07:49的版本

Introduction

  • This paper presented a speech information factorization method based on a novel deep generative model that we called factorial discriminative normalization flow.

Qualitative and quantitative experimental results show that compared to all other models, the proposed factorial DNF can retain the class structure corresponding to multiple information factors, and changing one factor will cause little distortion on other factors. This demonstrates that factorial DNF can well factorize speech signal into different information factors.

Members

  • Haoran Sun, Lantian Li, Yunqi Cai, Yang Zhang, Thomas Fang Zheng, Dong Wang

Publications

  • Haoran Sun, Lantian Li, Yunqi Cai, Yang Zhang, Thomas Fang Zheng, Dong Wang, "Deep Generative Factorization For Speech Signal", 2020. pdf

Source Code

  • xxx

Factorial DNF

  • xxx

Experiments

Data

  • xx

Encoding

  • xx

Fdnf tsne.png

Factor manipulation


                        Phone Manipulation
 Model |  p(q2|x) | p(q2|x') |  δ(q2)  ||  p(s|x) |  p(s|x') |   δ(s)
  VAE  |   0.013  |  0.312   |  0.299   ||  0.612  |  0.454   |  -0.158 
  NF   |   0.013  |  0.410   |  0.397   ||  0.612  |  0.489   |  -0.123 
  DNF  |   0.013  |  0.619   |  0.606   ||  0.612  |  0.335   |  -0.277  
 f-DNF |   0.013  |  0.636   |  0.623   ||  0.612  |  0.536   |  -0.076  


                        Speaker Manipulation
 Model |  p(s2|x) | p(s2|x') |  δ(s2)  ||  p(q|x) |  p(q|x') |   δ(q)
  VAE  |   0.010  |  0.303   |  0.293   ||  0.520  |  0.509   |  -0.011 
  NF   |   0.010  |  0.435   |  0.425   ||  0.520  |  0.484   |  -0.036 
  DNF  |   0.010  |  0.700   |  0.690   ||  0.520  |  0.349   |  -0.171  
 f-DNF |   0.010  |  0.710   |  0.700   ||  0.520  |  0.503   |  -0.017  

Future Work

  • Test factorial DNF on larger datasets.
  • Establish general theories for deep generative factorization.