2013-08-10 - 版本历史

Cslt：/* Embedded progress */

2013-08-20T05:17:03Z

‎Embedded progress

Cslt：/* Embedded progress */

2013-08-20T05:08:40Z

‎Embedded progress

Cslt：/* Embedded progress */

2013-08-20T04:56:42Z

‎Embedded progress

Cslt：/* DNN Confidence estimation */

2013-08-20T03:40:55Z

‎DNN Confidence estimation

Cslt：/* Stream decoding */

2013-08-20T03:39:34Z

‎Stream decoding

Cslt：/* Discriminative DNN */

2013-08-16T03:16:43Z

‎Discriminative DNN

Cslt：/* Tencent exps */

2013-08-16T02:52:34Z

‎Tencent exps

2013年8月16日 (五) 02:51 Cslt

2013-08-16T02:51:47Z

Cslt：以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Discriminative DNN === * Running 1200-3620 NN, graph generation is done. Should be don...”创建新页面

2013-08-16T02:51:17Z

以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Discriminative DNN === * Running 1200-3620 NN, graph generation is done. Should be don...”创建新页面

新页面

== Data sharing ==

* LM count files still undelivered!

== DNN progress ==

=== Discriminative DNN ===

* Running 1200-3620 NN, graph generation is done. Should be done in 3 days.

=== Sparse DNN ===

* Iterative sparse sticky training runs. More sparsity is expected.

=== Tencent exps ===

==DNN Confidence estimation==

* Distribution graph is obtained. The performance seems bad.
* A possible reason is that the decoding is LM-based, and the confidence is only acoustic related. So (1) the errors in linguistic layer are not really errors in the acoustic layer (2) the search will automatically choose the almost-correct phones/states.
* The conclusion is that the DNN confidence is most suitable for grammar-based applications, or at least LM information is not very important.

* To be done:
# CI phone confidence, on going
# No-tone confidence, on going

===GFCC DNN ==

*GFCC computing is highly slow. 100 hour speech costs 16 hour cpu time. RT is around 0.2. It is intolerable.
*GFCC-based DNN training for 100 hour speech data is done. Need to test the noise-robust performance in 2 days.

==Stream decoding==
* the code is done. Simple testing is completed.
* Problem 1: CMN initialization is not prefect. Need to train a better initial CMN model.
* Problem 2: balance for posterior-based silence detection.

==Subgraph integration==

* G.fst integration is done. Initial test passed. Looks like the zero-probability is better for the NUM class.
* HCLG integration is done. A bug fixed, passed initial test.
* Online integration cost is 1 minute. Need to optimize.
* Need thorough testing with the Tencent test suite.
* Need to tune the subgraph feeding probability.

== Embedded progress ==

* GFCC-based engine test
* Attain a performance curve: RT,memory size,package size Vs vocabulary size.
* A new demo released for 4600 song names.

@@ 第50行： / 第50行： @@
 * GFCC-based engine test
 * Attain a performance curve: RT,memory size,package size Vs vocabulary size.
-* A new demo released for [[4600 song names]].
+* A new demo released for [[4600 song names]]. [http://cslt.riit.tsinghua.edu.cn/csltdemo/public/release/easr/easr.v1.0.song.apk download here]

←上一版本		2013年8月16日 (五) 02:52的版本
第14行：		第14行：

	=== Tencent exps ===		=== Tencent exps ===
		+	* online support
		+	* garbage model training
		+	* VAD optimization

	==DNN Confidence estimation==		==DNN Confidence estimation==

←上一版本		2013年8月16日 (五) 02:51的版本
第26行：		第26行：


−	===GFCC DNN ==	+	==GFCC DNN ==

	*GFCC computing is highly slow. 100 hour speech costs 16 hour cpu time. RT is around 0.2. It is intolerable.		*GFCC computing is highly slow. 100 hour speech costs 16 hour cpu time. RT is around 0.2. It is intolerable.

@@ 第22行： / 第22行： @@
 * Distribution graph is obtained. The performance seems bad.
 * A possible reason is that the decoding is LM-based, and the confidence is only acoustic related. So (1) the errors in linguistic layer are not really errors in the acoustic layer (2) the search will automatically choose the almost-correct phones/states.
-* The conclusion is that the DNN confidence is most suitable for grammar-based applications, or at least LM information is not very important.
+* The conclusion is that the DNN confidence is most suitable for grammar-based applications, or at least LM information is not very strong.
 * To be done:
 # CI phone confidence, on going
 # No-tone confidence, on going
 ==GFCC DNN ==

@@ 第36行： / 第36行： @@
 ==Stream decoding==
 * the code is done. Simple testing is completed.
-* Problem 1: CMN initialization is not prefect. Need to train a better initial CMN model.
+* Problem 1: CMN initialization is not perfect. Need to train a better initial CMN model.
 * Problem 2: balance for posterior-based silence detection.

@@ 第7行： / 第7行： @@
 === Discriminative DNN ===
-* Running 1200-3620 NN, graph generation is done. Should be done in 3 days.
+* Running 1200-3620 NN, graph generation is done. DT training should be done in 3 days.
 === Sparse DNN ===