Sinovoice-2014-03-25 - 版本历史

2014年3月25日 (二) 06:45 Wangd

2014-03-25T06:45:54Z

Wangd：/* Train Analysis */

2014-03-25T06:28:12Z

‎Train Analysis

Wangd：/* Environment setting */

2014-03-25T06:13:06Z

‎Environment setting

Wangd：/* 1000h Training */

2014-03-25T06:12:43Z

‎1000h Training

2014年3月25日 (二) 06:10 Wangd

2014-03-25T06:10:56Z

Wangd：以内容“=Environment setting= * Raid215 is a bit slow. Move some den-lattice and alignment to Raid212. =Corpora= * Labeling Beijing Mobile? * Now totally 1229h (470 + 346 +...”创建新页面

2014-03-25T06:09:26Z

以内容“=Environment setting= * Raid215 is a bit slow. Move some den-lattice and alignment to Raid212. =Corpora= * Labeling Beijing Mobile? * Now totally 1229h (470 + 346 +...”创建新页面

新页面

=Environment setting=

* Raid215 is a bit slow. Move some den-lattice and alignment to Raid212.

=Corpora=

* Labeling Beijing Mobile?
* Now totally 1229h (470 + 346 + 105BJ mobile + 200 PICC + 108h HBTc) telephone speech is ready.
* 16k 6000h data: 978h online data from DataTang + 656h online mobile data + 4300h recording data.
* LM corpus preparation done.

=Acoustic modeling=

==Telephone model training==

===1000h Training===

* Jietong phone, 200 hour seed, 10k states training:

*Baseline: 8k states, 470+300 MPE4, 20.29
*MPE1: 21.91
*MPE2: 21.71
*MPE3: 21.68
*MPE4: 21.86

* CSLT phone, 8k states traiing
*MPE1: 20.60
*MPE2: 20.37

===PICC dedicated training===

* Need to collect financial text data and retrain the LM
* Need to comb word list and training text

==6000 hour 16k training==

===Training progress===

* 6000h/CSLT phone set alignment/denlattice completed
* 6000h/jt phone set alignment/denlattice completed
* MPE is kicked off

===Train Analysis===
* The Qihang model used a subset of the 6k data
:* 2500+950H+tang500h*+20131220, approximately 1700+2400 hours

* GMM training using this subset achieved 22.47%. Xiaoming's result is 16.1%.
:* Seems the database is still not very consistent
:* Xiaoming will try to reproduce the Qihang training using this subset

* Tested the 1700h model and 6000h model on the T test sets

<pre>
model/testcase | ditu | due1| entity1 | rec1 | shiji | zaixian1 | zaixian2 | kuaisu
------------------------------------------------------------------------------------------------
1700h_mpe | 12.18 | 12.93 | 5.29 | 3.69 | 21.73 | 25.38 | 19.45 | 12.50
------------------------------------------------------------------------------------------------
6000h_xEnt | 11.13 | 10.12 | 4.64 | 2.80 | 17.67 | 27.45 | 23.23 | 10.98
</pre>

:* 6000h model is general better than the 1700h for careful reading or domain specific recording
:* 6000h with MPE/jt phone set is still on training, but better performance is expected
:* This indicates that we should prepare domain-specific AM (not only 8k/16k). The online test prefers online training data
:* Suggest test the 6000 model on jidong data

=Language modeling=

* Training data ready
* First focus on PICC test set, try to improve the PPL

=DNN Decoder=

==Online decoder adaptation==

* Incremental training finished (stream mode)
* 8k sentence test

<pre>
non-stream baseline MPE5: %WER 9.91 [ 4734 / 47753, 235 ins, 509 del, 3990 sub ]
stream: MPE1：%WER 9.66 [ 4612 / 47753, 252 ins, 490 del, 3870 sub ]
stream: MPE2: %WER 9.48 [ 4529 / 47753, 251 ins, 477 del, 3801 sub ]
stream: MPE3: %WER 9.43 [ 4502 / 47753, 230 ins, 484 del, 3788 sub ]
stream: MPE4: %WER 9.39 [ 4482 / 47753, 236 ins, 475 del, 3771 sub ]
</pre>

←上一版本		2014年3月25日 (二) 06:13的版本
第1行：		第1行：
	=Environment setting=		=Environment setting=
−
−	* Raid215 is a bit slow. Move some den-lattice and alignment to Raid212.

	=Corpora=		=Corpora=

@@ 第15行： / 第15行： @@
 ===1000h Training===
 *Baseline: 8k states, 470+300 MPE4, 20.29
 :*MPE1: 21.91
 :*MPE2: 21.71

@@ 第3行： / 第3行： @@
 =Corpora=
-* Labeling Beijing Mobile?
+* Labeling Beijing Mobile.
 * Now totally 1229h (470 + 346 + 105BJ mobile + 200 PICC + 108h HBTc) telephone speech is ready.
 * 16k 6000h data: 978h online data from DataTang + 656h online mobile data + 4300h recording data.

@@ 第47行： / 第47行： @@
 :* Seems the database is still not very consistent
 :* Xiaoming will try to reproduce the Qihang training using this subset
-:* Suggest test the 6000 model on jidong data
+:* Test the 6000 model on jidong data, obtained 2% absolute improvement.
-−
-−
 =Language modeling=

@@ 第19行： / 第19行： @@
 *Baseline: 8k states, 470+300 MPE4, 20.29
-*MPE1: 21.91
+:*MPE1: 21.91
-*MPE2: 21.71
+:*MPE2: 21.71
-*MPE3: 21.68
+:*MPE3: 21.68
-*MPE4: 21.86
+:*MPE4: 21.86
-* CSLT phone, 8k states traiing
+* CSLT phone, 8k states training
-*MPE1: 20.60
+:*MPE1: 20.60
-*MPE2: 20.37
+:*MPE2: 20.37
 ===PICC dedicated training===
@@ 第50行： / 第50行： @@
 :* Seems the database is still not very consistent
 :* Xiaoming will try to reproduce the Qihang training using this subset
 :* Suggest test the 6000 model on jidong data
@@ 第71行： / 第57行： @@
 * Training data ready
-* First focus on PICC test set, try to improve the PPL
+* Xiaoxi from CSLT may be involved after some patent writing
 =DNN Decoder=
@@ 第85行： / 第71行： @@
   stream:              MPE2: %WER 9.48 [ 4529 / 47753, 251 ins, 477 del, 3801 sub ]
   stream:              MPE3: %WER 9.43 [ 4502 / 47753, 230 ins, 484 del, 3788 sub ]
-  stream:              MPE4:  %WER 9.39 [ 4482 / 47753, 236 ins, 475 del, 3771 sub ]
+  stream:              MPE4: %WER 9.39 [ 4482 / 47753, 236 ins, 475 del, 3771 sub ]
 </pre>