2013-05-17 - 版本历史

Cslt：/* Tencent exps */

2013-05-17T06:24:30Z

‎Tencent exps

Cslt：/* Embedded progress */

2013-05-17T06:23:31Z

‎Embedded progress

Cslt：/* Embedded progress */

2013-05-17T06:11:15Z

‎Embedded progress

Cslt：/* Tencent exps */

2013-05-17T06:00:13Z

‎Tencent exps

Cslt：/* Kaldi/HTK merge */

2013-05-17T05:42:54Z

‎Kaldi/HTK merge

2013年5月17日 (五) 05:40 Liuc

2013-05-17T05:40:14Z

2013年5月17日 (五) 05:33 Liuc

2013-05-17T05:33:21Z

2013年5月17日 (五) 05:15 Cslt

2013-05-17T05:15:06Z

2013年5月17日 (五) 04:52 Cslt

2013-05-17T04:52:42Z

Cslt：以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Experiments === * setups for input layer : s1: mfcc(13), splice +-5[143] : s2: mfcc(13),...”创建新页面

2013-05-17T04:48:07Z

以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Experiments === * setups for input layer : s1: mfcc(13), splice +-5[143] : s2: mfcc(13),...”创建新页面

新页面

== Data sharing ==

* LM count files still undelivered!

== DNN progress ==
=== Experiments ===
* setups for input layer
: s1: mfcc(13), splice +-5[143]
: s2: mfcc(13), splice +-5(143), LDA[143]
: s3: mfcc(13), delta(39), splice +-5(429), LDA[143]
: s4: mfcc(13), delta(39), splice +-5(429), LDA[300]
* setups for alignment
: tri1: triphone training, feature input: mfcc(13), delta[39]. #pdfs 1651, #gaussians 10028
: tri2: LDA/MLLT training, feature input: mfcc(13), delta(39), splice +-4(351), LDA[40]. #pdfs 3536, #gaussians 39995
* other notes
: about 100 hours training data
: 88k LM, biglm decoding (1e-5 / 1e-9)
: gpu-based nnet training, in-1200-1200-1200-1200-out
* results
{| class="wikitable"
! Test Set !! fMMI !! s1/tri1 !! s2/tri1 !! s3/tri1 !! s4/tri1 !! s2/tri2 !! s4/tri2 !! cpu-based (like s4/tri1)
|-
|map || 28.58 || 25.38 || 24.47 || 26.16 || 26.20 || 22.85 || 24.27 || 26.45
|-
|2044 || 24.79 || 23.58 || 22.82 || 23.84 || 24.13 || 21.45 || 22.76 || 24.66
|-
|notetp3|| 21.64 || 16.08 || 14.89 || 15.92 || 15.97 || 14.89 || 14.79 || 16.14
|-
|1900 || 8.19 || 8.55 || 8.43 || 8.66 || 8.90 || 7.30 || 7.91 || 8.23
|-
|general|| 39.63 || 36.18 || 34.79 || 35.88 || 35.90 || 33.06 || 33.79 || 38.02
|-
|online1|| 35.19 || 34.68 || 33.90 || 33.45 || 33.38 || 32.93 || 32.43 || 33.00
|-
|online2|| 28.30 || 27.27 || 26.61 || 26.26 || 26.36 || 25.94 || 25.69 || 26.63
|-
|speedup|| 28.45 || 24.97 || 24.40 || 24.55 || 25.42 || 23.04 || 23.67 || 27.17
|-
|}
* conclusion
#the GPU approach is comparable with the CPU approach (see s4/tri1 & CPU results). The former works slightly better in most cases.
#the fine training leads to significant better performance than the rough training (see s2/tri1 vs s2/tri2 & s4/tri1 vs s4/tri2)
#the delta features do not help, actually harm the perfromance (see s2/tri1 vs s3/tri1 & s4/tri1)
#the identical-dimension LDA helps the performance (see s1/tri1 vs s2/tri1)

'''the best system is s2/tri2: without delta, apply a linear LDA.'''

*to be done
#experiment with Tencent feature
#migrate the bottlenck structure to the 100k task with s2/tri2.

=== Tencent exps ===

#手动将NN网络的W权重，较小的置零，在保留30%左右的较大权重的条件下，系统性能未见明显衰减。
#按照HTK模型的结构，以及HTK align的结构，修改Kaldi的GPU接口，验证并无问题，已开始较大规模数据训练（1000小时），网络结构前后5帧扩展，4个隐层，每层2048节点，输出15000个状态。使用mpe模型alignment，特征为plp特征，未做任何映射。
#解码器仍在CLG结构下，修改声学模型计算接口，接入DNN模型，验证无问题，已开始效率优化。

待做实验：
#验证不同学习率调节策略，指数下降衰减方式，newbob方式。
#验证不同特征在大数据上的作用。
#最后层不过softmax，降维得到BN特征实验，类似IBM BN做法

=== GPU & CPU merge ===

#just started

=== L-1 sparse initial training ===
* experiments on L1 and L2 penalty
#LM: 4G gigabyte LM, 1e-5 small LM
#AM: 100hour tri4b_nn

{| class="wikitable"
! Test Set !! 0 !! 1.00E-06 !! 1.00E-05 !! 2.50E-05 !! 5.00E-05 !! 7.50E-05 !! 1.00E-04
|-
|map || 61.06 || 61.18 || 61.31 || 61.72 || 62.89 || 62.84 || 62.48
|-
|2044 || 49.53 || 49.58 || 49.84 || 49.94 || 50.71 || 51.08 || 51.08
|-
|notetp3|| 43.44 || 43.44 || 43.55 || 44.58 || 45.22 || 44.95 || 45.76
|-
|1900 || 38.50 || 38.54 || 39.00 || 39.21 || 40.33 || 40.53 || 40.60
|-
|general|| 61.24 || 61.22 || 61.69 || 61.84 || 62.71 || 62.83 || 62.93
|-
|online1|| 58.02 || 58.05 || 58.31 || 58.23 || 58.83 || 59.06 || 59.46
|-
|online2|| 53.62 || 53.70 || 54.03 || 53.93 || 54.65 || 54.94 || 55.51
|-
|speedup|| 57.51 || 57.49 || 57.93 || 58.31 || 59.75 || 60.08 || 59.52
|-
|}
* Conclusions
#L1 and L2 penalty do not work in the current nnet-GPU code.
#will work to check the L1 code and change the penalty scheme.

== Kaldi/HTK merge ==

* HTK2Kaldi: hold.
* Kaldi2HTK: stuck. various sp models tried but don't help
* Need thorough debug this week.

== Embedded progress ==
*Status:
:# first embedded demo done, 1000 words take 3.2M memory.
:# accuracy test not yet finished
:# training acoustic model for sphinx. The toolkit runs well, need prepare data and start parallel training.

*To be done
:# finish AM training
:# run offline test

@@ 第55行： / 第55行： @@
 *Status:
 :# first embedded demo done, 1000 words take 3.2M memory.
-:# accuracy test finished. The test data involves 3 speakers recorded in a car, 1000 address names.
+:# accuracy test finished. The test data involves 3 speakers recorded in a car with Chongqing dialect, 1000 address names.
 :# training acoustic model for sphinx. The an4 training process is done, while the test seems problematic.
@@ 第70行： / 第70行： @@
 *To be done
-:# finish 400hour AM training
+:# finish the large scale AM training

@@ 第55行： / 第55行： @@
 *Status:
 :# first embedded demo done, 1000 words take 3.2M memory.
-:# accuracy test finished
+:# accuracy test finished. The test data involves 3 speakers recorded in a car, 1000 address names.
 :# training acoustic model for sphinx. The an4 training process is done, while the test seems problematic.

@@ 第50行： / 第50行： @@
 * HTK2Kaldi: hold.
 * Kaldi2HTK: pdf error problem.
 * workaround; use the BN feature to train HTK models, so without kaldi training.
 == Embedded progress ==

@@ 第58行： / 第58行： @@
 :# accuracy test finished
 :# training acoustic model for sphinx. The an4 training process is done, while the test seems problematic.
 *To be done
 :# finish 400hour AM training

@@ 第31行： / 第31行： @@
 === Tencent exps ===
-#手动将NN网络的W权重，较小的置零，在保留30%左右的较大权重的条件下，系统性能未见明显衰减(how much in number?)。
+#手动将NN网络的W权重，较小的置零，在保留30%左右的较大权重的条件下，系统性能未见明显衰减(how much in number? it would be interesting to re-train the net after pruning the weights)。
 #按照HTK模型的结构，以及HTK align的结构，修改Kaldi的GPU接口，验证并无问题，已开始较大规模数据训练（1000小时），网络结构前后5帧扩展，4个隐层，每层2048节点，输出15000个状态。使用mpe模型alignment，特征为plp特征，未做任何映射。
 #解码器仍在CLG结构下，修改声学模型计算接口，接入DNN模型，验证无问题，已开始效率优化。
@@ 第38行： / 第38行： @@
 #验证不同学习率调节策略，指数下降衰减方式，newbob方式。
 #验证不同特征在大数据上的作用。
-#最后层不过softmax，降维得到BN特征实验，类似IBM BN做法 (this is the linear dim-reduction, it might be worse than the BN...)
+#最后层不过softmax，降维得到BN特征实验，类似IBM BN做法 (if this is a linear dim-reduction, it might be worse than the BN...)
 === GPU & CPU merge ===

@@ 第8行： / 第8行： @@
 {| class="wikitable"
-! Test Set !! fMMI !! s1/tri1 !! s2/tri1 !! s3/tri1 !! s4/tri1 !! s2/tri2 !! s4/tri2 !! cpu-based (like s4/tri1)
+! Test Set !! fMMI !! s1/tri1 !! s2/tri1 !! s3/tri1 !! s4/tri1 !! s2/tri2 !! s4/tri2 !! cpu-based (like s4/tri1) !! plp-s2/tri2
 |-
-|map    || 28.58 || 25.38 || 24.47 || 26.16 || 26.20 || 22.85 || 24.27 || 26.45
+|map    || 28.58 || 25.38 || 24.47 || 26.16 || 26.20 || 22.85 || 24.27 || 26.45 || 23.86
 |-
-|2044   || 24.79 || 23.58 || 22.82 || 23.84 || 24.13 || 21.45 || 22.76 || 24.66
+|2044   || 24.79 || 23.58 || 22.82 || 23.84 || 24.13 || 21.45 || 22.76 || 24.66 || 22.68
 |-
-|notetp3|| 21.64 || 16.08 || 14.89 || 15.92 || 15.97 || 14.89 || 14.79 || 16.14
+|notetp3|| 21.64 || 16.08 || 14.89 || 15.92 || 15.97 || 14.89 || 14.79 || 16.14 || 16.46
 |-
-|1900   ||  8.19 ||  8.55 ||  8.43 ||  8.66 ||  8.90 ||  7.30 ||  7.91 ||  8.23
+|1900   ||  8.19 ||  8.55 ||  8.43 ||  8.66 ||  8.90 ||  7.30 ||  7.91 ||  8.23 ||  7.68
 |-
-|general|| 39.63 || 36.18 || 34.79 || 35.88 || 35.90 || 33.06 || 33.79 || 38.02
+|general|| 39.63 || 36.18 || 34.79 || 35.88 || 35.90 || 33.06 || 33.79 || 38.02 || 34.12
 |-
-|online1|| 35.19 || 34.68 || 33.90 || 33.45 || 33.38 || 32.93 || 32.43 || 33.00
+|online1|| 35.19 || 34.68 || 33.90 || 33.45 || 33.38 || 32.93 || 32.43 || 33.00 || 33.60
 |-
-|online2|| 28.30 || 27.27 || 26.61 || 26.26 || 26.36 || 25.94 || 25.69 || 26.63
+|online2|| 28.30 || 27.27 || 26.61 || 26.26 || 26.36 || 25.94 || 25.69 || 26.63 || 26.20
 |-
-|speedup|| 28.45 || 24.97 || 24.40 || 24.55 || 25.42 || 23.04 || 23.67 || 27.17
+|speedup|| 28.45 || 24.97 || 24.40 || 24.55 || 25.42 || 23.04 || 23.67 || 27.17 || 23.62
 |-
 |}

@@ 第6行： / 第6行： @@
 === Experiments ===
 * setups for mfcc/plp
 {| class="wikitable"
 ! Test Set !! fMMI !! s1/tri1 !! s2/tri1 !! s3/tri1 !! s4/tri1 !! s2/tri2 !! s4/tri2 !! cpu-based (like s4/tri1)
@@ 第26行： / 第27行： @@
 |-
 |}

@@ 第5行： / 第5行： @@
 == DNN progress ==
 === Experiments ===
-* setups for input layer
+* setups for mfcc/plp
-: s1: mfcc(13), splice +-5[143]
-: s2: mfcc(13), splice +-5(143), LDA[143]
-: s3: mfcc(13), delta(39), splice +-5(429), LDA[143]
-: s4: mfcc(13), delta(39), splice +-5(429), LDA[300]
-* setups for alignment
-: tri1: triphone training, feature input: mfcc(13), delta[39]. #pdfs 1651, #gaussians 10028
-: tri2: LDA/MLLT training, feature input: mfcc(13), delta(39), splice +-4(351), LDA[40]. #pdfs 3536, #gaussians 39995
-* other notes
-: about 100 hours training data
-: 88k LM, biglm decoding (1e-5 / 1e-9)
-: gpu-based nnet training, in-1200-1200-1200-1200-out
-* results
 {| class="wikitable"
 ! Test Set !! fMMI !! s1/tri1 !! s2/tri1 !! s3/tri1 !! s4/tri1 !! s2/tri2 !! s4/tri2 !! cpu-based (like s4/tri1)
@@ 第38行： / 第26行： @@
 |-
 |}
 * conclusion
 #the GPU approach is comparable with the CPU approach (see s4/tri1 & CPU results). The former works slightly better in most cases.
@@ 第67行： / 第56行： @@
 #just started
@@ 第103行： / 第61行： @@
 * HTK2Kaldi: hold.
-* Kaldi2HTK: stuck. various sp models tried but don't help
+* Kaldi2HTK: pdf error problem.
-* Need thorough debug this week.
+* workaround; use the BN feature to train HTK models, so without kaldi training.
@@ 第110行： / 第68行： @@
 *Status:
 :# first embedded demo done, 1000 words take 3.2M memory.
-:# accuracy test not yet finished
+:# accuracy test finished
-:# training acoustic model for sphinx. The toolkit runs well, need prepare data and start parallel training.
+:# training acoustic model for sphinx. The an4 training process is done, while the test seems problematic.
 *To be done
-:# finish AM training
+:# finish 400hour AM training
-:# run offline test