Yannick Estève / ONTRAC-Kaldi

Blame view

egs/wsj/s5/local/run_sgmm2.sh 7.13 KB
edit raw normal view history
8dcb6dfcb   Yannick Estève   first commit 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
 
  #!/bin/bash
  
  # This script is invoked from ../run.sh
  # It contains some SGMM-related scripts that I am breaking out of the main run.sh for clarity.
  
  . ./cmd.sh
  
  # Note: you might want to try to give the option --spk-dep-weights=false to train_sgmm2.sh;
  # this takes out the "symmetric SGMM" part which is not always helpful.
  
  # SGMM system on si84 data [sgmm5a].  Note: the system we aligned from used the si284 data for
  # training, but this shouldn't have much effect.
  
  (
    steps/align_fmllr.sh --nj 30 --cmd "$train_cmd" \
      data/train_si84 data/lang exp/tri4b exp/tri4b_ali_si84 || exit 1;
  
    steps/train_ubm.sh --cmd "$train_cmd" \
      400 data/train_si84 data/lang exp/tri4b_ali_si84 exp/ubm5a || exit 1;
  
    steps/train_sgmm2.sh --cmd "$train_cmd" \
      7000 9000 data/train_si84 data/lang exp/tri4b_ali_si84 \
      exp/ubm5a/final.ubm exp/sgmm2_5a || exit 1;
  
    (
      utils/mkgraph.sh data/lang_test_tgpr exp/sgmm2_5a exp/sgmm2_5a/graph_tgpr
      steps/decode_sgmm2.sh --nj 10 --cmd "$decode_cmd" --transform-dir exp/tri4b/decode_tgpr_dev93 \
        exp/sgmm2_5a/graph_tgpr data/test_dev93 exp/sgmm2_5a/decode_tgpr_dev93
    ) &
  
    steps/align_sgmm2.sh --nj 30 --cmd "$train_cmd" --transform-dir exp/tri4b_ali_si84 \
      --use-graphs true --use-gselect true data/train_si84 data/lang exp/sgmm2_5a exp/sgmm2_5a_ali_si84 || exit 1;
    steps/make_denlats_sgmm2.sh --nj 30 --sub-split 30 --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si84 \
      data/train_si84 data/lang exp/sgmm2_5a_ali_si84 exp/sgmm2_5a_denlats_si84
  
    steps/train_mmi_sgmm2.sh --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si84 --boost 0.1 \
      data/train_si84 data/lang exp/sgmm2_5a_ali_si84 exp/sgmm2_5a_denlats_si84 exp/sgmm2_5a_mmi_b0.1
  
    wait
  
    for iter in 1 2 3 4; do
      steps/decode_sgmm2_rescore.sh --cmd "$decode_cmd" --iter $iter \
        --transform-dir exp/tri4b/decode_tgpr_dev93 data/lang_test_tgpr data/test_dev93 exp/sgmm2_5a/decode_tgpr_dev93 \
        exp/sgmm2_5a_mmi_b0.1/decode_tgpr_dev93_it$iter &
    done
  
    steps/train_mmi_sgmm2.sh --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si84 --boost 0.1 \
     --update-opts "--cov-min-value=0.9" data/train_si84 data/lang exp/sgmm2_5a_ali_si84 exp/sgmm2_5a_denlats_si84 exp/sgmm2_5a_mmi_b0.1_m0.9
  
    for iter in 1 2 3 4; do
      steps/decode_sgmm2_rescore.sh --cmd "$decode_cmd" --iter $iter \
        --transform-dir exp/tri4b/decode_tgpr_dev93 data/lang_test_tgpr data/test_dev93 exp/sgmm2_5a/decode_tgpr_dev93 \
        exp/sgmm2_5a_mmi_b0.1_m0.9/decode_tgpr_dev93_it$iter &
    done
  
  ) &
  
  
  (
  # The next commands are the same thing on all the si284 data.
  
  # SGMM system on the si284 data [sgmm5b]
    steps/train_ubm.sh --cmd "$train_cmd" \
      600 data/train_si284 data/lang exp/tri4b_ali_si284 exp/ubm5b || exit 1;
  
    steps/train_sgmm2.sh --cmd "$train_cmd" \
     11000 25000 data/train_si284 data/lang exp/tri4b_ali_si284 \
      exp/ubm5b/final.ubm exp/sgmm2_5b || exit 1;
  
    (
      utils/mkgraph.sh data/lang_test_tgpr exp/sgmm2_5b exp/sgmm2_5b/graph_tgpr
      steps/decode_sgmm2.sh --nj 10 --cmd "$decode_cmd" --transform-dir exp/tri4b/decode_tgpr_dev93 \
        exp/sgmm2_5b/graph_tgpr data/test_dev93 exp/sgmm2_5b/decode_tgpr_dev93
      steps/decode_sgmm2.sh --nj 8 --cmd "$decode_cmd" --transform-dir exp/tri4b/decode_tgpr_eval92 \
        exp/sgmm2_5b/graph_tgpr data/test_eval92 exp/sgmm2_5b/decode_tgpr_eval92
  
      utils/mkgraph.sh data/lang_test_bd_tgpr exp/sgmm2_5b exp/sgmm2_5b/graph_bd_tgpr || exit 1;
      steps/decode_sgmm2.sh --nj 10 --cmd "$decode_cmd" --transform-dir exp/tri4b/decode_bd_tgpr_dev93 \
        exp/sgmm2_5b/graph_bd_tgpr data/test_dev93 exp/sgmm2_5b/decode_bd_tgpr_dev93
      steps/decode_sgmm2.sh --nj 8 --cmd "$decode_cmd" --transform-dir exp/tri4b/decode_bd_tgpr_eval92 \
        exp/sgmm2_5b/graph_bd_tgpr data/test_eval92 exp/sgmm2_5b/decode_bd_tgpr_eval92
    ) &
  
  
   # This shows how you would build and test a quinphone SGMM2 system, but
    (
     steps/train_sgmm2.sh --cmd "$train_cmd" \
        --context-opts "--context-width=5 --central-position=2" \
      11000 25000 data/train_si284 data/lang exp/tri4b_ali_si284 \
       exp/ubm5b/final.ubm exp/sgmm2_5c || exit 1;
     # Decode from lattices in exp/sgmm2_5b
      steps/decode_sgmm2_fromlats.sh --cmd "$decode_cmd"  --transform-dir exp/tri4b/decode_tgpr_dev93 \
         data/test_dev93 data/lang_test_tgpr exp/sgmm2_5b/decode_tgpr_dev93 exp/sgmm2_5c/decode_tgpr_dev93
      steps/decode_sgmm2_fromlats.sh --cmd "$decode_cmd"  --transform-dir exp/tri4b/decode_tgpr_eval92 \
         data/test_eval92 data/lang_test_tgpr exp/sgmm2_5b/decode_tgpr_eval92 exp/sgmm2_5c/decode_tgpr_eval92
    ) &
  
  
    steps/align_sgmm2.sh --nj 30 --cmd "$train_cmd" --transform-dir exp/tri4b_ali_si284 \
      --use-graphs true --use-gselect true data/train_si284 data/lang exp/sgmm2_5b exp/sgmm2_5b_ali_si284
  
    steps/make_denlats_sgmm2.sh --nj 30 --sub-split 30 --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si284 \
      data/train_si284 data/lang exp/sgmm2_5b_ali_si284 exp/sgmm2_5b_denlats_si284
  
    steps/train_mmi_sgmm2.sh --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si284 --boost 0.1 \
      data/train_si284 data/lang exp/sgmm2_5b_ali_si284 exp/sgmm2_5b_denlats_si284 exp/sgmm2_5b_mmi_b0.1
  
    for iter in 1 2 3 4; do
      for test in eval92; do # dev93
        steps/decode_sgmm2_rescore.sh --cmd "$decode_cmd" --iter $iter \
          --transform-dir exp/tri4b/decode_bd_tgpr_${test} data/lang_test_bd_fg data/test_${test} exp/sgmm2_5b/decode_bd_tgpr_${test} \
          exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_${test}_it$iter &
       done
    done
  
    steps/train_mmi_sgmm2.sh --cmd "$decode_cmd" --transform-dir exp/tri4b_ali_si284 --boost 0.1 \
      --drop-frames true data/train_si284 data/lang exp/sgmm2_5b_ali_si284 exp/sgmm2_5b_denlats_si284 exp/sgmm2_5b_mmi_b0.1_z
  
    for iter in 1 2 3 4; do
      for test in eval92 dev93; do
        steps/decode_sgmm2_rescore.sh --cmd "$decode_cmd" --iter $iter \
          --transform-dir exp/tri4b/decode_bd_tgpr_${test} data/lang_test_bd_fg data/test_${test} exp/sgmm2_5b/decode_bd_tgpr_${test} \
          exp/sgmm2_5b_mmi_b0.1_z/decode_bd_tgpr_${test}_it$iter &
       done
    done
  
  ) &
  
  wait
  
  # Examples of combining some of the best decodings: SGMM+MMI with
  # MMI+fMMI on a conventional system.
  
  local/score_combine.sh data/test_eval92 \
     data/lang_test_bd_tgpr \
     exp/tri4b_fmmi_a/decode_tgpr_eval92_it8 \
     exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3 \
     exp/combine_tri4b_fmmi_a_sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it8_3
  
  
  # %WER 4.43 [ 250 / 5643, 41 ins, 12 del, 197 sub ] exp/tri4b_fmmi_a/decode_tgpr_eval92_it8/wer_11
  # %WER 3.85 [ 217 / 5643, 35 ins, 11 del, 171 sub ] exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3/wer_10
  # combined to:
  # %WER 3.76 [ 212 / 5643, 32 ins, 12 del, 168 sub ] exp/combine_tri4b_fmmi_a_sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it8_3/wer_12
  
  # Checking MBR decode of baseline:
  rm -r exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3.mbr 2>/dev/null
  cp -r exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3{,.mbr}
  local/score_mbr.sh data/test_eval92 data/lang_test_bd_tgpr exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3.mbr
  # MBR decoding did not seem to help (baseline was 3.85).  I think this is normal at such low WERs.
  %WER 3.86 [ 218 / 5643, 35 ins, 11 del, 172 sub ] exp/sgmm2_5b_mmi_b0.1/decode_bd_tgpr_eval92_it3.mbr/wer_10