Yannick Estève / ONTRAC-Kaldi

Blame view

egs/hub4_english/s5/local/run_segmentation_wsj.sh 12.8 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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
 
  #!/bin/bash
  
  # Copyright 2016-18  Vimal Manohar
  # Apache 2.0
  
  set -e
  set -o pipefail
  
  # This script demonstrates how to use out-of-domain WSJ models to segment long
  # audio recordings of HUB4 with raw unaligned transcripts into short segments
  # with aligned transcripts for training new ASR models. 
  
  # The overall procedure is as follow:
  # 1) Train a GMM on out-of-domain WSJ corpus
  # 2) Decode broadcast news recordings (HUB4) with WSJ GMM and 4-gram biased LM 
  #    trained on the raw unprocessed transcript. 
  # 3) Use the CTM output to segment the recordings keep the best matched
  #    audio and text.
  # 4) Train an in-domain GMM on the above data. 
  # 5) Repeat steps 2, 3 and 4 using the new in-domain GMM.
  # 6) Re-segment the data retaining only the "clean" part of the data.
  
  # See the script steps/cleanup/segment_long_utterances.sh for details about 
  # audio-transcript alignment (Step 2, 3)
  # See the script steps/cleanup/clean_and_segment_data.sh for details about 
  # cleaning up transcripts (Step 6)
  
  # In step 3, if you need to align the full hypothesis of audio with the 
  # reference text as opposed to finding the best matching substring, 
  # then use --align-full-hyp true in the scripts below.
  
  # WSJ models (From step 1)
  # %WER 29.9 | 728 32834 | 72.9 17.8 9.3 2.8 29.9 92.7 | exp/wsj_tri3/decode_nosp_eval97.pem_rescore/score_16_0.0/eval97.pem.ctm.filt.sys
  # %WER 30.8 | 728 32834 | 71.8 18.4 9.8 2.6 30.8 92.3 | exp/wsj_tri3/decode_nosp_eval97.pem/score_17_0.0/eval97.pem.ctm.filt.sys
  
  # In-domain GMM (From step 4) -- 107 hrs
  # %WER 19.1 | 728 32834 | 82.7 12.2 5.1 1.9 19.1 86.4 | exp/tri4_a/decode_nosp_eval97.pem_rescore/score_14_1.0/eval97.pem.ctm.filt.sys
  # %WER 20.4 | 728 32834 | 81.6 13.1 5.3 2.1 20.4 87.4 | exp/tri4_a/decode_nosp_eval97.pem/score_14_0.0/eval97.pem.ctm.filt.sys
  
  # Stage 2 in-domain GMM (From step 5) -- 124 hrs
  # %WER 20.9 | 728 32834 | 81.2 13.6 5.3 2.1 20.9 87.4 | exp/tri4_2a/decode_nosp_eval97.pem/score_14_0.0/eval97.pem.ctm.filt.sys
  # %WER 19.8 | 728 32834 | 82.3 12.9 4.7 2.2 19.8 86.1 | exp/tri4_2a/decode_nosp_eval97.pem_rescore/score_12_0.5/eval97.pem.ctm.filt.sys
  
  # GMM trained on cleaned transcripts (From step 6) -- 120 hrs
  # %WER 18.4 | 728 32834 | 83.6 11.9 4.5 2.1 18.4 84.8 | exp/tri5_2a_cleaned/decode_nosp_eval97.pem_rescore/score_13_0.0/eval97.pem.ctm.filt.sys
  # %WER 19.6 | 728 32834 | 82.5 12.7 4.8 2.2 19.6 86.8 | exp/tri5_2a_cleaned/decode_nosp_eval97.pem/score_13_0.0/eval97.pem.ctm.filt.sys
  
  # Oracle HUB4 transcripts -- 148 hrs
  # %WER 17.8 | 728 32834 | 84.1 11.8 4.1 1.9 17.8 82.8 | exp/tri4/decode_nosp_eval97.pem_rescore/score_13_0.5/eval97.pem.ctm.filt.sys
  # %WER 19.0 | 728 32834 | 83.0 12.7 4.3 2.0 19.0 84.2 | exp/tri4/decode_nosp_eval97.pem/score_13_0.0/eval97.pem.ctm.filt.sys
  
  stage=0
  segment_stage=-8
  nj=40
  reco_nj=80
  stage1_affix=a    # For steps 2, 3 and 4 above
  stage2_affix=2a   # For step 5 above
  
  # WSJ run.sh must be run until the data preparation stage
  wsj_base=../../wsj/s5   # Change this to the WSJ base directory
  
  if [ -f ./path.sh ]; then . ./path.sh; fi
  . ./cmd.sh
  
  . utils/parse_options.sh
  
  if [ ! -f $wsj_base/data/train_si284/wav.scp ]; then
    echo "WSJ data directory $wsj_base/data/train_si284 is not prepared."
    echo "Run the initial stages of WSJ's run.sh"
    exit 0
  fi
  
  if [ $stage -le 0 ]; then
    # We copy the prepared data to the current directory
    utils/copy_data_dir.sh $wsj_base/data/train_si84_2kshort data/train_si84_2kshort
    utils/copy_data_dir.sh $wsj_base/data/train_si84 data/train_si84
    utils/copy_data_dir.sh $wsj_base/data/train_si284 data/train_si284
  fi
  
  ###############################################################################
  ## Simulate unsegmented HUB4 data directory.
  ###############################################################################
  
  if [ $stage -le 1 ]; then
    utils/data/convert_data_dir_to_whole.sh data/train data/train_long
  
    steps/make_mfcc.sh --cmd "$train_cmd --max-jobs-run 40" \
      --nj $reco_nj --write-utt2num-frames true \
      data/train_long exp/make_mfcc/train_long mfcc
    steps/compute_cmvn_stats.sh data/train_long \
      exp/make_mfcc/train_long mfcc
    utils/fix_data_dir.sh data/train_long
  fi
  
  ###############################################################################
  ## Train GMM on out-of-domain WSJ corpus 
  ###############################################################################
  
  if [ $stage -le 2 ]; then
    steps/train_mono.sh --boost-silence 1.25 --nj $nj --cmd "$train_cmd" \
      data/train_si84_2kshort data/lang_nosp exp/wsj_mono0a
  fi
  
  if [ $stage -le 3 ]; then
    steps/align_si.sh --boost-silence 1.25 --nj $nj --cmd "$train_cmd" \
      data/train_si84 data/lang_nosp exp/wsj_mono0a exp/wsj_mono0a_ali_si84
  
    steps/train_deltas.sh --boost-silence 1.25 --cmd "$train_cmd" 2500 15000 \
      data/train_si84 data/lang_nosp exp/wsj_mono0a_ali_si84 exp/wsj_tri1
  fi
  
  if [ $stage -le 4 ]; then
    steps/align_si.sh --nj $nj --cmd "$train_cmd" \
      data/train_si284 data/lang_nosp exp/wsj_tri1 exp/wsj_tri1_ali_si284
  
    steps/train_lda_mllt.sh --cmd "$train_cmd" \
      --splice-opts "--left-context=3 --right-context=3" 4000 42000 \
      data/train_si284 data/lang_nosp exp/wsj_tri1_ali_si284 exp/wsj_tri2
  fi
  
  if [ $stage -le 5 ]; then
    steps/align_si.sh --nj $nj --cmd "$train_cmd" \
      data/train_si284 data/lang_nosp exp/wsj_tri2 exp/wsj_tri2_ali_si284
  
    steps/train_sat.sh --cmd "$train_cmd" \
      4000 42000 \
      data/train_si284 data/lang_nosp exp/wsj_tri2_ali_si284 exp/wsj_tri3
  fi
  
  if [ $stage -le 6 ]; then
    utils/mkgraph.sh data/lang_nosp_test \
      exp/wsj_tri3/{,graph_nosp_test}
  
    for dset in eval97.pem; do
      this_nj=`cat data/$dset/spk2utt | wc -l`
      if [ $this_nj -gt 20 ]; then
        this_nj=20
      fi
      steps/decode_fmllr.sh --nj $this_nj --cmd "$decode_cmd" --num-threads 4 \
        exp/wsj_tri3/graph_nosp_test data/$dset \
        exp/wsj_tri3/decode_nosp_${dset}
      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
        data/lang_nosp_test data/lang_nosp_test_rescore \
        data/${dset} exp/wsj_tri3/decode_nosp_${dset} \
        exp/wsj_tri3/decode_nosp_${dset}_rescore
    done
  fi
  
  ###############################################################################
  # Segment long HUB4 recordings and retrieve transcript using 
  # Smith-Waterman alignment.
  # Use a SAT model trained on train_si284 (wsj_tri3) as seed model for decoding.
  ###############################################################################
  
  if [ $stage -le 7 ]; then
    steps/cleanup/segment_long_utterances.sh --cmd "$train_cmd" \
      --stage $segment_stage --nj $reco_nj \
      --max-bad-proportion 0.5 --align-full-hyp false \
      exp/wsj_tri3 data/lang_nosp data/train_long \
      data/train_reseg_${stage1_affix} exp/segment_long_utts_${stage1_affix}_train
  fi
  
  if [ $stage -le 8 ]; then
    steps/compute_cmvn_stats.sh data/train_reseg_${stage1_affix} \
      exp/make_mfcc/train_reseg_${stage1_affix} mfcc
    utils/fix_data_dir.sh data/train_reseg_${stage1_affix}
  
    utils/data/modify_speaker_info.sh data/train_reseg_${stage1_affix} \
      data/train_reseg_${stage1_affix}_spk30sec
    steps/compute_cmvn_stats.sh data/train_reseg_${stage1_affix}_spk30sec \
      exp/make_mfcc/train_reseg_${stage1_affix}_spk30sec mfcc
    utils/fix_data_dir.sh data/train_reseg_${stage1_affix}_spk30sec
  fi
  
  ###############################################################################
  # Train new in-domain GMM (tri4_a) on retrieved transcripts.
  ###############################################################################
  
  if [ $stage -le 9 ]; then
    steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
      data/train_reseg_${stage1_affix}_spk30sec data/lang_nosp \
      exp/wsj_tri3 exp/wsj_tri3_ali_train_reseg_${stage1_affix}
  
    steps/train_sat.sh --cmd "$train_cmd" 4200 40000 \
      data/train_reseg_${stage1_affix}_spk30sec data/lang_nosp \
      exp/wsj_tri3_ali_train_reseg_${stage1_affix} exp/tri3_${stage1_affix} 
  fi
  
  if [ $stage -le 10 ]; then
    steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
      data/train_reseg_${stage1_affix}_spk30sec data/lang_nosp exp/tri3_${stage1_affix} exp/tri3_${stage1_affix}_ali
  
    steps/train_sat.sh --cmd "$train_cmd" 5000 100000 \
      data/train_reseg_${stage1_affix}_spk30sec data/lang_nosp exp/tri3_${stage1_affix}_ali exp/tri4_${stage1_affix}
  fi
  
  if [ $stage -le 11 ]; then
    utils/mkgraph.sh data/lang_nosp_test exp/tri4_${stage1_affix}/{,graph_nosp_test}
    for dset in eval97.pem; do
      this_nj=`cat data/$dset/spk2utt | wc -l`
      if [ $this_nj -gt 20 ]; then
        this_nj=20
      fi
      steps/decode_fmllr.sh --nj $this_nj --cmd "$decode_cmd" --num-threads 4 \
        exp/tri4_${stage1_affix}/graph_nosp_test data/$dset exp/tri4_${stage1_affix}/decode_nosp_${dset}
      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
        data/lang_nosp_test data/lang_nosp_test_rescore \
        data/${dset} exp/tri4_${stage1_affix}/decode_nosp_${dset} \
        exp/tri4_${stage1_affix}/decode_nosp_${dset}_rescore
    done
  fi
  
  ###############################################################################
  # Segment long HUB4 recordings and retrieve transcript using 
  # Smith-Waterman alignment.
  # Use in-domain SAT model (tri4_a) as seed model for decoding.
  ###############################################################################
  
  if [ $stage -le 12 ]; then
    steps/cleanup/segment_long_utterances.sh --cmd "$train_cmd" \
      --stage $segment_stage --nj $reco_nj \
      --max-bad-proportion 0.5 --align-full-hyp false \
      exp/tri4_${stage1_affix} data/lang_nosp data/train_long \
      data/train_reseg_${stage2_affix} exp/segment_long_utts_${stage2_affix}_train
  fi
  
  if [ $stage -le 13 ]; then
    steps/compute_cmvn_stats.sh data/train_reseg_${stage2_affix} \
      exp/make_mfcc/train_reseg_${stage2_affix} mfcc
    utils/fix_data_dir.sh data/train_reseg_${stage2_affix}
  fi
  
  ###############################################################################
  # Train new in-domain GMM (tri4_2a) on retrieved transcripts.
  ###############################################################################
  
  if [ $stage -le 14 ]; then
    steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
      data/train_reseg_${stage2_affix} data/lang_nosp \
      exp/tri4_${stage1_affix} exp/tri4_${stage1_affix}_ali_train_reseg_${stage2_affix}
  
    steps/train_sat.sh --cmd "$train_cmd" 4200 40000 \
      data/train_reseg_${stage2_affix} data/lang_nosp \
      exp/tri4_${stage1_affix}_ali_train_reseg_${stage2_affix} exp/tri4_${stage2_affix} 
  fi
  
  if [ $stage -le 15 ]; then
    utils/mkgraph.sh data/lang_nosp_test exp/tri4_${stage2_affix}/{,graph_nosp_test}
    for dset in eval97.pem; do
      this_nj=`cat data/$dset/spk2utt | wc -l`
      if [ $this_nj -gt 20 ]; then
        this_nj=20
      fi
      steps/decode_fmllr.sh --nj $this_nj --cmd "$decode_cmd" --num-threads 4 \
        exp/tri4_${stage2_affix}/graph_nosp_test data/$dset exp/tri4_${stage2_affix}/decode_nosp_${dset}
      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
        data/lang_nosp_test data/lang_nosp_test_rescore \
        data/${dset} exp/tri4_${stage2_affix}/decode_nosp_${dset} \
        exp/tri4_${stage2_affix}/decode_nosp_${dset}_rescore
    done
  fi
  
  ###############################################################################
  # Cleanup transcripts
  # Use in-domain SAT model (tri4_2a) as seed model for decoding.
  ###############################################################################
  
  cleanup_stage=-1
  cleanup_affix=cleaned
  srcdir=exp/tri4_${stage2_affix}
  cleaned_data=data/train_reseg_${stage2_affix}_${cleanup_affix}
  dir=${srcdir}_${cleanup_affix}_work
  cleaned_dir=${srcdir}_${cleanup_affix}
  
  if [ $stage -le 16 ]; then
    steps/cleanup/clean_and_segment_data.sh --stage $cleanup_stage --nj 80 \
      --cmd "$train_cmd" \
      data/train_reseg_${stage2_affix} data/lang_nosp \
      $srcdir $dir $cleaned_data
  fi
  
  ###############################################################################
  # Train new in-domain GMM (tri4_2a) on cleaned-up transcripts.
  ###############################################################################
  
  if [ $stage -le 17 ]; then
    steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
      $cleaned_data data/lang_nosp $srcdir ${srcdir}_ali_${cleanup_affix}
  
    steps/train_sat.sh --cmd "$train_cmd" \
      5000 100000 $cleaned_data data/lang_nosp \
      ${srcdir}_ali_${cleanup_affix} exp/tri5_${stage2_affix}_${cleanup_affix}
  fi
  
  if [ $stage -le 18 ]; then
    utils/mkgraph.sh data/lang_nosp_test \
      exp/tri5_${stage2_affix}_${cleanup_affix}/{,graph_nosp_test}
    for dset in eval97.pem; do
      this_nj=`cat data/$dset/spk2utt | wc -l`
      if [ $this_nj -gt 20 ]; then
        this_nj=20
      fi
      steps/decode_fmllr.sh --nj $this_nj --cmd "$decode_cmd" --num-threads 4 \
        exp/tri5_${stage2_affix}_${cleanup_affix}/graph_nosp_test data/$dset \
        exp/tri5_${stage2_affix}_${cleanup_affix}/decode_nosp_${dset}
      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
        data/lang_nosp_test data/lang_nosp_test_rescore \
        data/${dset} exp/tri5_${stage2_affix}_${cleanup_affix}/decode_nosp_${dset} \
        exp/tri5_${stage2_affix}_${cleanup_affix}/decode_nosp_${dset}_rescore
    done
  fi
  
  exit 0