“2013-04-19”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
第8行: 第8行:
 
*Tencent baseline:
 
*Tencent baseline:
 
<pre>
 
<pre>
700 hour online data + 700 863 data , HLDA+MPE; 88k lexicon:
+
:*700 hour online data + 700 863 data , HLDA+MPE; 88k lexicon:
 
+
:record1900: 8.4
record1900: 8.4
+
:2044:      22.4
2044:      22.4
+
:online 1:  35.6
online 1:  35.6
+
:online 2:  29.6
online 2:  29.6
+
:map:        24.5
map:        24.5
+
:notepad:    16
notepad:    16
+
:general:    36
general:    36
+
:speedup:    26.8
speedup:    26.8
+
 
</pre>
 
</pre>
 
*bMMI
 
*bMMI

2013年4月25日 (四) 04:10的版本

Data sharing

  • AM/lexicon/LM are shared.
  • LM count files are still in transfering.

DNN progress

400 hour BN model

  • Tencent baseline:
:*700 hour online data + 700 863 data , HLDA+MPE; 88k lexicon:
:record1900: 8.4
:2044:       22.4
:online 1:   35.6
:online 2:   29.6
:map:        24.5
:notepad:    16
:general:    36
:speedup:    26.8
  • bMMI
exp/tri4b_mmi_b0.1/decode_tlm_biglm:
map: %WER 27.54 [ 4029 / 14628, 63 ins, 533 del, 3433 sub ]
2044: %WER 24.44 [ 5681 / 23241, 313 ins, 844 del, 4524 sub ]
notetp3: %WER 19.81 [ 367 / 1853, 8 ins, 48 del, 311 sub ]
record1900: %WER 7.65 [ 909 / 11888, 17 ins, 377 del, 515 sub ]
general: %WER 38.52 [ 14490 / 37619, 182 ins, 1314 del, 12994 sub ]
online1: %WER 34.66 [ 9855 / 28433, 398 ins, 1895 del, 7562 sub ]
online2: %WER 27.23 [ 16092 / 59101, 623 ins, 2954 del, 12515 sub ]
speedup: %WER 27.88 [ 1465 / 5255, 32 ins, 332 del, 1101 sub ]
  • fMMI
exp/tri4b_fmmi_indirect/decode_tlm_it7_biglm:
map: %WER 27.69 [ 4050 / 14628, 61 ins, 538 del, 3451 sub ]
2044: %WER 24.03 [ 5584 / 23241, 316 ins, 817 del, 4451 sub ]
notetp3: %WER 21.75 [ 403 / 1853, 7 ins, 53 del, 343 sub ]
record1900: %WER 7.35 [ 874 / 11888, 31 ins, 347 del, 496 sub ]
general: %WER 38.90 [ 14635 / 37619, 206 ins, 1331 del, 13098 sub ]
online1: %WER 34.33 [ 9762 / 28433, 424 ins, 1888 del, 7450 sub ]
online2: %WER 26.80 [ 15837 / 59101, 648 ins, 2902 del, 12287 sub ]
speedup: %WER 26.81 [ 1409 / 5255, 35 ins, 284 del, 1090 sub ]
  • DNN-bn
exp/tri4d_fmmi_indirect/decode_tlm_it4_biglm:
map: %WER 23.79 [ 3480 / 14628, 58 ins, 465 del, 2957 sub ]
2044: %WER 21.77 [ 5060 / 23241, 297 ins, 711 del, 4052 sub ]
notetp3: %WER 15.81 [ 293 / 1853, 8 ins, 35 del, 250 sub ]
record1900: %WER 6.57 [ 781 / 11888, 18 ins, 325 del, 438 sub ]
general: %WER 33.61 [ 12645 / 37619, 191 ins, 968 del, 11486 sub ]
online1: %WER 31.44 [ 8940 / 28433, 311 ins, 1619 del, 7010 sub ]
online2: %WER 24.10 [ 14245 / 59101, 523 ins, 2417 del, 11305 sub ]
speedup: %WER 22.82 [ 1199 / 5255, 39 ins, 241 del, 919 sub ]

Tencent test result

AM: 70h training data(2 day, 15 machines, 10 threads)
LM: 88k LM
Test case: general
gmmi-bmmi: 38.7%
dnn-1: 28% 11 frame window, phone-based tree
dnn-2: 34% 9 frame window, state-based tree


GPU & CPU merge

Invesigate the possibility to merge GPU and CPU code. Try to find out an easier way. (1 week)

L-1 sparse initial training

Start to investigating.

Kaldi/HTK merge

  • HTK2Kaldi: the tool with Kaldi does not work.
  • Kaldi2HTK: done with implementation. Testing?

Embedded progress

  • Some large performance (speed) degradation with the embedded platform(1/60).
  • Planning for sparse DNN.
  • QA LM training, still failed. Mengyuan need more work on this.