2013-04-26 - 版本历史

2013年4月26日 (五) 07:14 Cslt

2013-04-26T07:14:14Z

2013年4月26日 (五) 07:13 Cslt

2013-04-26T07:13:13Z

2013年4月26日 (五) 07:12 Cslt

2013-04-26T07:12:45Z

2013年4月26日 (五) 07:08 Cslt

2013-04-26T07:08:21Z

2013年4月26日 (五) 06:50 Cslt

2013-04-26T06:50:26Z

2013年4月26日 (五) 06:49 Cslt

2013-04-26T06:49:45Z

2013年4月26日 (五) 06:49 Cslt

2013-04-26T06:49:25Z

2013年4月26日 (五) 06:48 Cslt

2013-04-26T06:48:37Z

2013年4月26日 (五) 06:43 Cslt

2013-04-26T06:43:50Z

2013年4月26日 (五) 06:42 Cslt

2013-04-26T06:42:59Z

@@ 第81行： / 第81行： @@
 *Status:
 :# QA LM training, done.
-:# PocketSphinx migration done, using PocketSphinx default Chinese model. After migrating to the smart phone, the test shows that the decoding is very slow. RT=7.0.
+:# PocketSphinx migration done, using PocketSphinx default Chinese model. After migrating to the smart phone, the test shows that the decoding (with an LM involving 90k words) is very slow (RT=7.0).
 *To be done
-:# Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.
+:# substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.
 :# Need to train a new AM.

@@ 第84行： / 第84行： @@
-*To bedone
+*To be done
 :# Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.
 :# Need to train a new AM.

@@ 第6行： / 第6行： @@
 ===400 hour DNN training===
 {| class="wikitable"
-!Test Set!! Tencent Baseline!! bMMI!! fMMI !! BN(with fMMI) !! Hybrid
+!Test Set!! Tencent Baseline!! bMMI!! fMMI !! BN(with fMMI) !! Hybrid (DNN/HMM)
 |-
 |1900||8.4  || 7.65 || 7.35||6.57 || 7.27

@@ 第52行： / 第52行： @@
 *To be done:
-:#CSLT reproduce phone-clustered based NN
+:#CSLT: reproduce phone-clustered NN (Eryu's results)
-:#CSLT investigate performance of different epochs.
+:#CSLT: investigate performance of different epoches.
 :#Tencent: feature comparison.
-:#FBank with PLP. With or without LDA.
+:#Tencent: FBank with PLP. With or without LDA.

@@ 第39行： / 第39行： @@
 :LM: 88k LM
 :Test case: general
-:#param=700k
+:param=700k
 {|class="wikitable"

@@ 第46行： / 第46行： @@
 |PLP(-5,+5) [Eryu]           || 47 || 38.4   || 26.5 || 23.8 ||300*1200*1200*1200*1200]*1700
 |-
-|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -|| 34 ||300*[1007*1007*1007*1007]*3xxx
+|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -|| 34 || - ||300*[1007*1007*1007*1007]*3xxx
 |-
 |}

@@ 第44行： / 第44行： @@
 !Feature !! GMM !!GMM-bMMI !! DNN !! DNN-MMI !! DNN structure
 |-
-|PLP(-5,+5) [Eryu]           || 47 || 38.4   || 26.5 || 23.8 ||300*[1200*1200*1200*1200]*1700
+|PLP(-5,+5) [Eryu]           || 47 || 38.4   || 26.5 || 23.8 ||300*1200*1200*1200*1200]*1700
 |-
-|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -      || 34 ||300*[1007*1007*1007*1007]*3xxx
+|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -|| 34 ||300*[1007*1007*1007*1007]*3xxx
 |-
 |}

@@ 第26行： / 第26行： @@
 |}
-*Tencent baseline is with 700h online data+ 700h 863 data, HLDA+MPE, 88k lexicon
+*Note
-*Our results are with 400 hour AM, 88k LM. ML+bMMI.
+:Tencent baseline is with 700h online data+ 700h 863 data, HLDA+MPE, 88k lexicon
-*The CSLT structure: 300*[1200*1200*1200*40*1200]*4850.
+:Our results are with 400 hour AM, 88k LM. ML+bMMI.
-*The CSLT feature: MFCC+delta MFCC
+:The CSLT structure: 300*[1200*1200*1200*40*1200]*4850.
-*To be done: compare with the traditional structure 300*[1200*1200*1200*1200*1200]*4850.
+:The CSLT feature: MFCC+delta MFCC
 ===Tencent test result===
-:  AM: 70h training data
+:AM: 70h training data
-:  LM: 88k LM
+:LM: 88k LM
-:  Test case: general
+:Test case: general
 {|class="wikitable"
-!Feature !! GMM !!GMM-bMMI !! DNN !! DNN-MMI
+!Feature !! GMM !!GMM-bMMI !! DNN !! DNN-MMI !! DNN structure
 |-
-|PLP(-5,+5) [Eryu]           || 47 || 38.4   || 26.5 || 23.8
+|PLP(-5,+5) [Eryu]           || 47 || 38.4   || 26.5 || 23.8 ||300*[1200*1200*1200*1200]*1700
 |-
-|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -      || 34
+|PLP+LDA+MLLT(-5,+5)[Jingbo] || 47 || -      || 34 ||300*[1007*1007*1007*1007]*3xxx
 |-
 |}
-* Note
-:#Tencent NN structure:
-::300*[1200*1200*1200*1200]*1700, #param=700k
-::300*[1007*1007*1007*1007]*3xxx  #param=700k
 *To be done:
 :#CSLT reproduce phone-clustered based NN
@@ 第59行： / 第59行： @@
 ===GPU & CPU merge===
-: Investigate the possibility to merge GPU and CPU code.
+# Investigate the possibility to merge GPU and CPU code.
-: CUDA code merged to CPU.
+:Decision: CUDA code merged to CPU.
 ===L-1 sparse initial training===
-:  L-1=1e-5, starting from 6th iteration, converged with another 3 iterations. The performance is generally worse than the case where l1=0, except one test suite.
+*Initial trial
-:  L-1=1e-6, the same results obtained, means le-6 is too small to be effective.
+:#  L-1=1e-5, starting from 6th iteration, converged with another 3 iterations. The performance is generally worse than the case where l1=0, except one test suite.
-:  L-1=1e-4, start from the first iteration. crashed. Need more investigation.
+:#  L-1=1e-6, the same results obtained, means le-6 is too small to be effective.
 ==Kaldi/HTK merge==
-:* HTK2Kaldi: hold.
+* HTK2Kaldi: hold.
-:* Kaldi2HTK: done with implementation. A bug fixed. gConst was computed in a wrong way. The current HDecode result is 14.9%; The tencent model is 11%; Kaldi decoder 7%.
+* Kaldi2HTK: done with implementation. A bug fixed. gConst was computed in a wrong way. The current HDecode result is 14.9%; The tencent model is 11%; Kaldi decoder 7%.
-:* Possibly the SP model issue, due to the complicated structure of silence in Kaldi.
+* Possibly the SP model issue, due to the complicated structure of silence in Kaldi.
 ==Embedded progress==
-:* PocketSphinx migration done, using PocketSphinx default Chinese model. After migrating to the smart phone, the test shows that the decoding is very slow. RT=7.0.
+*Status:
-:* QA LM training, done.
+:# QA LM training, done.
-:* Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.
+*To bedone
-:* Need to train a new AM.
+:# Next substitute the LM with JSGF grammar involving 1000 words. Finish the initial test.

@@ 第30行： / 第30行： @@
 *The CSLT structure: 300*[1200*1200*1200*40*1200]*4850.
 *The CSLT feature: MFCC+delta MFCC
-* compare with the traditional structure 300*[1200*1200*1200*1200*1200]*4850.
+*To be done: compare with the traditional structure 300*[1200*1200*1200*1200*1200]*4850.
 ===Tencent test result===
@@ 第50行： / 第50行： @@
 :300*[1200*1200*1200*1200]*1700, #param=700k
 :300*[1007*1007*1007*1007]*3xxx  #param=700k
+:*To be done:
-:*CSLT reproduce phone-clustered based NN
+::CSLT reproduce phone-clustered based NN
-:*CSLT investigate performance of different epochs.
+::CSLT investigate performance of different epochs.
-:*Tencent: feature comparison.
+::Tencent: feature comparison.
-:*FBank with PLP. With or without LDA.
+::FBank with PLP. With or without LDA.