Skip to content

Commit e330320

Browse files
phanisankar-nidadavoludanpovey
phanisankar-nidadavolu
authored and
danpovey
committed
[scripts,egs] Made changes to the augmentation script to make it work for ASR and speaker ID (kaldi-asr#3119)
Now multi-style training with noise and reverberation is an option (instead of speed augmentation). Multi-style training seems to be more robust to unseen/noisy conditions.
1 parent e2dc9c3 commit e330320

File tree

9 files changed

+1216
-172
lines changed

9 files changed

+1216
-172
lines changed

‎egs/swbd/s5c/local/chain/multi_condition/run_tdnn_aug_1a.sh

+281
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,281 @@
1+
#!/bin/bash
2+
3+
# This recipe does multi-style training of TDNN model
4+
5+
# local/chain/compare_wer_general.sh --rt03 tdnn7q_sp tdnn1a_aug
6+
# System tdnn7q_sp tdnn1a_aug
7+
# WER on train_dev(tg) 11.91 12.06
8+
# WER on train_dev(fg) 10.99 10.92
9+
# WER on eval2000(tg) 14.3 14.4
10+
# WER on eval2000(fg) 12.8 12.9
11+
# WER on rt03(tg) 17.2 17.1
12+
# WER on rt03(fg) 15.1 14.8
13+
# Final train prob -0.062 -0.087
14+
# Final valid prob -0.074 -0.105
15+
# Final train prob (xent) -0.933 -1.164
16+
# Final valid prob (xent) -0.9027 -1.2246
17+
# Num-parameters 18693376 18483664
18+
19+
set -e
20+
21+
# configs for 'chain'
22+
stage=0
23+
train_stage=-10
24+
get_egs_stage=-10
25+
num_epochs=3
26+
27+
# Augmentation options
28+
aug_list="reverb babble music noise clean" # Original train dir is referred to as `clean`
29+
num_reverb_copies=1
30+
use_ivectors=true
31+
32+
affix=1a
33+
suffix="_aug"
34+
if [ -e data/rt03 ]; then maybe_rt03=rt03; else maybe_rt03= ; fi
35+
36+
decode_iter=
37+
decode_nj=50
38+
39+
# training options
40+
frames_per_eg=150,110,100
41+
remove_egs=false
42+
common_egs_dir=
43+
xent_regularize=0.1
44+
dropout_schedule='0,[email protected],[email protected],0'
45+
46+
test_online_decoding=false # if true, it will run the last decoding stage.
47+
48+
# End configuration section.
49+
echo "$0 $@" # Print the command line for logging
50+
51+
. ./cmd.sh
52+
. ./path.sh
53+
. ./utils/parse_options.sh
54+
55+
dir=exp/chain/tdnn${affix}${suffix}
56+
57+
if ! cuda-compiled; then
58+
cat <<EOF && exit 1
59+
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
60+
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
61+
where "nvcc" is installed.
62+
EOF
63+
fi
64+
65+
clean_set=train_nodup
66+
clean_ali=tri4_ali_nodup
67+
train_set=$clean_set$suffix # Will be prepared by the script local/nnet3/prepare_multistyle_data.sh
68+
ali_dir=$clean_ali$suffix
69+
treedir=exp/chain/tri5_7d_tree$suffix
70+
lang=data/lang_chain_2y
71+
72+
# First creates augmented data and then extracts features for it data
73+
# The script also creates alignments for aug data by copying clean alignments
74+
local/nnet3/multi_condition/run_aug_common.sh --stage $stage \
75+
--aug-list "$aug_list" --num-reverb-copies $num_reverb_copies \
76+
--use-ivectors "$use_ivectors" \
77+
--train-set $clean_set --clean-ali $clean_ali || exit 1;
78+
79+
if [ $stage -le 11 ]; then
80+
# Get the alignments as lattices (gives the LF-MMI training more freedom).
81+
# use the same num-jobs as the alignments
82+
prefixes=""
83+
include_original=false
84+
for n in $aug_list; do
85+
if [ "$n" == "reverb" ]; then
86+
for i in `seq 1 $num_reverb_copies`; do
87+
prefixes="$prefixes "reverb$i
88+
done
89+
elif [ "$n" != "clean" ]; then
90+
prefixes="$prefixes "$n
91+
else
92+
# The original train directory will not have any prefix
93+
# include_original flag will take care of copying the original lattices
94+
include_original=true
95+
fi
96+
done
97+
nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
98+
steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/${clean_set} \
99+
data/lang exp/tri4 exp/tri4_lats_nodup${suffix}_clean
100+
rm exp/tri4_lats_nodup${suffix}_clean/fsts.*.gz # save space
101+
steps/copy_lat_dir.sh --nj $nj --cmd "$train_cmd" \
102+
--include-original "$include_original" --prefixes "$prefixes" \
103+
data/${train_set} exp/tri4_lats_nodup${suffix}_clean exp/tri4_lats_nodup${suffix} || exit 1;
104+
fi
105+
106+
if [ $stage -le 12 ]; then
107+
# Create a version of the lang/ directory that has one state per phone in the
108+
# topo file. [note, it really has two states.. the first one is only repeated
109+
# once, the second one has zero or more repeats.]
110+
rm -rf $lang
111+
cp -r data/lang $lang
112+
silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
113+
nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
114+
# Use our special topology... note that later on may have to tune this
115+
# topology.
116+
steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
117+
fi
118+
119+
if [ $stage -le 13 ]; then
120+
# Build a tree using our new topology. This is the critically different
121+
# step compared with other recipes.
122+
steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
123+
--context-opts "--context-width=2 --central-position=1" \
124+
--cmd "$train_cmd" 7000 data/$train_set $lang exp/$ali_dir $treedir
125+
fi
126+
127+
if [ $stage -le 14 ]; then
128+
echo "$0: creating neural net configs using the xconfig parser";
129+
130+
num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
131+
learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
132+
affine_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
133+
tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
134+
linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
135+
prefinal_opts="l2-regularize=0.01"
136+
output_opts="l2-regularize=0.002"
137+
138+
mkdir -p $dir/configs
139+
140+
cat <<EOF > $dir/configs/network.xconfig
141+
input dim=100 name=ivector
142+
input dim=40 name=input
143+
144+
# please note that it is important to have input layer with the name=input
145+
# as the layer immediately preceding the fixed-affine-layer to enable
146+
# the use of short notation for the descriptor
147+
fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
148+
149+
# the first splicing is moved before the lda layer, so no splicing here
150+
relu-batchnorm-dropout-layer name=tdnn1 $affine_opts dim=1536
151+
tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
152+
tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
153+
tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
154+
tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=0
155+
tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
156+
tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
157+
tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
158+
tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
159+
tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
160+
tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
161+
tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
162+
tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
163+
tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
164+
tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
165+
linear-component name=prefinal-l dim=256 $linear_opts
166+
167+
prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
168+
output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
169+
170+
prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
171+
output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
172+
EOF
173+
steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
174+
fi
175+
176+
if [ $stage -le 15 ]; then
177+
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
178+
utils/create_split_dir.pl \
179+
/export/b0{5,6,7,8}/$USER/kaldi-data/egs/swbd-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
180+
fi
181+
182+
steps/nnet3/chain/train.py --stage $train_stage \
183+
--cmd "$train_cmd" \
184+
--feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
185+
--feat.cmvn-opts "--norm-means=false --norm-vars=false" \
186+
--chain.xent-regularize $xent_regularize \
187+
--chain.leaky-hmm-coefficient 0.1 \
188+
--chain.l2-regularize 0.0 \
189+
--chain.apply-deriv-weights false \
190+
--chain.lm-opts="--num-extra-lm-states=2000" \
191+
--trainer.dropout-schedule $dropout_schedule \
192+
--trainer.add-option="--optimization.memory-compression-level=2" \
193+
--egs.dir "$common_egs_dir" \
194+
--egs.stage $get_egs_stage \
195+
--egs.opts "--frames-overlap-per-eg 0 --constrained false" \
196+
--egs.chunk-width $frames_per_eg \
197+
--trainer.num-chunk-per-minibatch 64 \
198+
--trainer.frames-per-iter 1500000 \
199+
--trainer.num-epochs $num_epochs \
200+
--trainer.optimization.num-jobs-initial 3 \
201+
--trainer.optimization.num-jobs-final 16 \
202+
--trainer.optimization.initial-effective-lrate 0.00025 \
203+
--trainer.optimization.final-effective-lrate 0.000025 \
204+
--trainer.max-param-change 2.0 \
205+
--cleanup.remove-egs $remove_egs \
206+
--feat-dir data/${train_set}_hires \
207+
--tree-dir $treedir \
208+
--lat-dir exp/tri4_lats_nodup$suffix \
209+
--dir $dir || exit 1;
210+
211+
fi
212+
213+
if [ $stage -le 16 ]; then
214+
# Note: it might appear that this $lang directory is mismatched, and it is as
215+
# far as the 'topo' is concerned, but this script doesn't read the 'topo' from
216+
# the lang directory.
217+
utils/mkgraph.sh --self-loop-scale 1.0 data/lang_sw1_tg $dir $dir/graph_sw1_tg
218+
fi
219+
220+
221+
graph_dir=$dir/graph_sw1_tg
222+
iter_opts=
223+
if [ ! -z $decode_iter ]; then
224+
iter_opts=" --iter $decode_iter "
225+
fi
226+
if [ $stage -le 17 ]; then
227+
rm $dir/.error 2>/dev/null || true
228+
for decode_set in train_dev eval2000 $maybe_rt03; do
229+
(
230+
steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
231+
--nj $decode_nj --cmd "$decode_cmd" $iter_opts \
232+
--online-ivector-dir exp/nnet3/ivectors_${decode_set} \
233+
$graph_dir data/${decode_set}_hires \
234+
$dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
235+
if $has_fisher; then
236+
steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
237+
data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
238+
$dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
239+
fi
240+
) || touch $dir/.error &
241+
done
242+
wait
243+
if [ -f $dir/.error ]; then
244+
echo "$0: something went wrong in decoding"
245+
exit 1
246+
fi
247+
fi
248+
249+
if $test_online_decoding && [ $stage -le 16 ]; then
250+
# note: if the features change (e.g. you add pitch features), you will have to
251+
# change the options of the following command line.
252+
steps/online/nnet3/prepare_online_decoding.sh \
253+
--mfcc-config conf/mfcc_hires.conf \
254+
$lang exp/nnet3/extractor $dir ${dir}_online
255+
256+
rm $dir/.error 2>/dev/null || true
257+
for decode_set in train_dev eval2000 $maybe_rt03; do
258+
(
259+
# note: we just give it "$decode_set" as it only uses the wav.scp, the
260+
# feature type does not matter.
261+
262+
steps/online/nnet3/decode.sh --nj $decode_nj --cmd "$decode_cmd" \
263+
--acwt 1.0 --post-decode-acwt 10.0 \
264+
$graph_dir data/${decode_set}_hires \
265+
${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
266+
if $has_fisher; then
267+
steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
268+
data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
269+
${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
270+
fi
271+
) || touch $dir/.error &
272+
done
273+
wait
274+
if [ -f $dir/.error ]; then
275+
echo "$0: something went wrong in decoding"
276+
exit 1
277+
fi
278+
fi
279+
280+
281+
exit 0;

‎egs/swbd/s5c/local/chain/run_tdnn_aug.sh

+1
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
multi_condition/run_tdnn_aug_1a.sh

0 commit comments

Comments
 (0)