run_ivector_common.sh
6.52 KB
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
#!/bin/bash
set -euo pipefail
# This script is modified based on mini_librispeech/s5/local/nnet3/run_ivector_common.sh
# This script is called from local/nnet3/run_tdnn.sh and
# local/chain/run_tdnn.sh (and may eventually be called by more
# scripts). It contains the common feature preparation and
# iVector-related parts of the script. See those scripts for examples
# of usage.
stage=0
train_set=train
test_sets="dev test"
gmm=tri5a
online=false
nnet3_affix=
. ./cmd.sh
. ./path.sh
. utils/parse_options.sh
gmm_dir=exp/${gmm}
ali_dir=exp/${gmm}_sp_ali
for f in data/${train_set}/feats.scp ${gmm_dir}/final.mdl; do
if [ ! -f $f ]; then
echo "$0: expected file $f to exist"
exit 1
fi
done
online_affix=
if [ $online = true ]; then
online_affix=_online
fi
if [ $stage -le 1 ]; then
# Although the nnet will be trained by high resolution data, we still have to
# perturb the normal data to get the alignment _sp stands for speed-perturbed
echo "$0: preparing directory for low-resolution speed-perturbed data (for alignment)"
utils/data/perturb_data_dir_speed_3way.sh data/${train_set} data/${train_set}_sp
echo "$0: making MFCC features for low-resolution speed-perturbed data"
steps/make_mfcc_pitch.sh --cmd "$train_cmd" --nj 70 data/${train_set}_sp \
exp/make_mfcc/train_sp mfcc_perturbed || exit 1;
steps/compute_cmvn_stats.sh data/${train_set}_sp \
exp/make_mfcc/train_sp mfcc_perturbed || exit 1;
utils/fix_data_dir.sh data/${train_set}_sp
fi
if [ $stage -le 2 ]; then
echo "$0: aligning with the perturbed low-resolution data"
steps/align_fmllr.sh --nj 30 --cmd "$train_cmd" \
data/${train_set}_sp data/lang $gmm_dir $ali_dir || exit 1
fi
if [ $stage -le 3 ]; then
# Create high-resolution MFCC features (with 40 cepstra instead of 13).
# this shows how you can split across multiple file-systems.
echo "$0: creating high-resolution MFCC features"
mfccdir=mfcc_perturbed_hires$online_affix
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/mfcc/aidatatang-$(date +'%m_%d_%H_%M')/s5/$mfccdir/storage $mfccdir/storage
fi
for datadir in ${train_set}_sp ${test_sets}; do
utils/copy_data_dir.sh data/$datadir data/${datadir}_hires$online_affix
done
# do volume-perturbation on the training data prior to extracting hires
# features; this helps make trained nnets more invariant to test data volume.
utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires$online_affix || exit 1;
for datadir in ${train_set}_sp ${test_sets}; do
steps/make_mfcc_pitch$online_affix.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
--cmd "$train_cmd" data/${datadir}_hires$online_affix exp/make_hires/$datadir $mfccdir || exit 1;
steps/compute_cmvn_stats.sh data/${datadir}_hires$online_affix exp/make_hires/$datadir $mfccdir || exit 1;
utils/fix_data_dir.sh data/${datadir}_hires$online_affix || exit 1;
# create MFCC data dir without pitch to extract iVector
utils/data/limit_feature_dim.sh 0:39 data/${datadir}_hires$online_affix data/${datadir}_hires_nopitch || exit 1;
steps/compute_cmvn_stats.sh data/${datadir}_hires_nopitch exp/make_hires/$datadir $mfccdir || exit 1;
done
fi
if [ $stage -le 4 ]; then
echo "$0: computing a subset of data to train the diagonal UBM."
# We'll use about a quarter of the data.
mkdir -p exp/nnet3${nnet3_affix}/diag_ubm
temp_data_root=exp/nnet3${nnet3_affix}/diag_ubm
num_utts_total=$(wc -l <data/${train_set}_sp_hires_nopitch/utt2spk)
num_utts=$[$num_utts_total/4]
utils/data/subset_data_dir.sh data/${train_set}_sp_hires_nopitch \
$num_utts ${temp_data_root}/${train_set}_sp_hires_nopitch_subset
echo "$0: computing a PCA transform from the hires data."
steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
--splice-opts "--left-context=3 --right-context=3" \
--max-utts 10000 --subsample 2 \
${temp_data_root}/${train_set}_sp_hires_nopitch_subset \
exp/nnet3${nnet3_affix}/pca_transform
echo "$0: training the diagonal UBM."
# Use 512 Gaussians in the UBM.
steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj 30 \
--num-frames 700000 \
--num-threads 8 \
${temp_data_root}/${train_set}_sp_hires_nopitch_subset 512 \
exp/nnet3${nnet3_affix}/pca_transform exp/nnet3${nnet3_affix}/diag_ubm
fi
if [ $stage -le 5 ]; then
# Train the iVector extractor. Use all of the speed-perturbed data since iVector extractors
# can be sensitive to the amount of data. The script defaults to an iVector dimension of
# 100.
echo "$0: training the iVector extractor"
steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj 10 \
data/${train_set}_sp_hires_nopitch exp/nnet3${nnet3_affix}/diag_ubm \
exp/nnet3${nnet3_affix}/extractor || exit 1;
fi
train_set=train_sp
if [ $stage -le 6 ]; then
# We extract iVectors on the speed-perturbed training data after combining
# short segments, which will be what we train the system on. With
# --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
# each of these pairs as one speaker; this gives more diversity in iVectors..
# Note that these are extracted 'online'.
# note, we don't encode the 'max2' in the name of the ivectordir even though
# that's the data we extract the ivectors from, as it's still going to be
# valid for the non-'max2' data, the utterance list is the same.
ivectordir=exp/nnet3${nnet3_affix}/ivectors_${train_set}
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $ivectordir/storage ]; then
utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/ivectors/aidatatang-$(date +'%m_%d_%H_%M')/s5/$ivectordir/storage $ivectordir/storage
fi
# having a larger number of speakers is helpful for generalization, and to
# handle per-utterance decoding well (iVector starts at zero).
temp_data_root=${ivectordir}
utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
data/${train_set}_hires_nopitch ${temp_data_root}/${train_set}_sp_hires_nopitch_max2
steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
${temp_data_root}/${train_set}_sp_hires_nopitch_max2 \
exp/nnet3${nnet3_affix}/extractor $ivectordir
# Also extract iVectors for the test data, but in this case we don't need the speed
# perturbation (sp).
for data in $test_sets; do
steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 8 \
data/${data}_hires_nopitch exp/nnet3${nnet3_affix}/extractor \
exp/nnet3${nnet3_affix}/ivectors_${data}
done
fi
exit 0