diff --git a/egs/casia_hwdb/v1/cmd.sh b/egs/casia_hwdb/v1/cmd.sh
new file mode 100755
index 00000000000..3c8eb9f93a5
--- /dev/null
+++ b/egs/casia_hwdb/v1/cmd.sh
@@ -0,0 +1,13 @@
+# you can change cmd.sh depending on what type of queue you are using.
+# If you have no queueing system and want to run on a local machine, you
+# can change all instances 'queue.pl' to run.pl (but be careful and run
+# commands one by one: most recipes will exhaust the memory on your
+# machine).  queue.pl works with GridEngine (qsub).  slurm.pl works
+# with slurm.  Different queues are configured differently, with different
+# queue names and different ways of specifying things like memory;
+# to account for these differences you can create and edit the file
+# conf/queue.conf to match your queue's configuration.  Search for
+# conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
+# or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
+
+export cmd="queue.pl"
diff --git a/egs/casia_hwdb/v1/image b/egs/casia_hwdb/v1/image
new file mode 120000
index 00000000000..6a4b3afeb09
--- /dev/null
+++ b/egs/casia_hwdb/v1/image
@@ -0,0 +1 @@
+../../cifar/v1/image
\ No newline at end of file
diff --git a/egs/casia_hwdb/v1/local/augment_data.sh b/egs/casia_hwdb/v1/local/augment_data.sh
new file mode 100755
index 00000000000..1f13ed15ded
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/augment_data.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+# Copyright   2018 Hossein Hadian
+#             2018 Ashish Arora
+
+# Apache 2.0
+# This script performs data augmentation.
+
+nj=4
+cmd=run.pl
+feat_dim=40
+fliplr=false
+verticle_shift=0
+echo "$0 $@"
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1;
+
+srcdir=$1
+outdir=$2
+datadir=$3
+
+mkdir -p $datadir/augmentations
+echo "copying $srcdir to $datadir/augmentations/aug1, allowed length, creating feats.scp"
+
+for set in aug1; do
+  image/copy_data_dir.sh --spk-prefix $set- --utt-prefix $set- \
+    $srcdir $datadir/augmentations/$set
+  cat $srcdir/allowed_lengths.txt > $datadir/augmentations/$set/allowed_lengths.txt
+  local/extract_features.sh --nj $nj --cmd "$cmd" --feat-dim $feat_dim \
+    --vertical-shift $verticle_shift \
+    --fliplr $fliplr --augment 'random_scale' $datadir/augmentations/$set
+done
+
+echo " combine original data and data from different augmentations"
+utils/combine_data.sh --extra-files images.scp $outdir $srcdir $datadir/augmentations/aug1
+cat $srcdir/allowed_lengths.txt > $outdir/allowed_lengths.txt
diff --git a/egs/casia_hwdb/v1/local/chain/compare_wer.sh b/egs/casia_hwdb/v1/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..ab880c1adb5
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/chain/compare_wer.sh
@@ -0,0 +1,67 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/cnn{1a,1b}
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/cnn{1a,1b}"
+  exit 1
+fi
+
+echo "# $0 $*"
+used_epochs=false
+
+echo -n "# System                     "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+echo -n "# WER                        "
+for x in $*; do
+  wer=$(cat $x/decode_test/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "# CER                        "
+for x in $*; do
+  cer=$(cat $x/decode_test/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
+
+if $used_epochs; then
+  exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
+fi
+
+echo -n "# Final train prob           "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob           "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train prob (xent)    "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob (xent)    "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
diff --git a/egs/casia_hwdb/v1/local/chain/run_cnn_e2eali_1b.sh b/egs/casia_hwdb/v1/local/chain/run_cnn_e2eali_1b.sh
new file mode 100755
index 00000000000..300c8ae8e31
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/chain/run_cnn_e2eali_1b.sh
@@ -0,0 +1,231 @@
+#!/bin/bash
+
+# e2eali_1b is the same as chainali_1a but uses the e2e chain model to get the
+# lattice alignments and to build a tree
+
+# ./local/chain/compare_wer.sh exp_yomdle_chinese/chain/e2e_cnn_1a exp_yomdle_chinese/chain/cnn_e2eali_1b
+# System                      e2e_cnn_1a cnn_e2eali_1b
+# CER                             15.44     13.57
+# Final train prob               0.0616   -0.0512
+# Final valid prob               0.0390   -0.0718
+# Final train prob (xent)                 -0.6199
+# Final valid prob (xent)                 -0.7448
+
+set -e -o pipefail
+
+data_dir=data
+exp_dir=exp
+
+stage=0
+
+nj=30
+train_set=train
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1b  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=4
+# training chunk-options
+chunk_width=340,300,200,100
+num_leaves=2000
+# we don't need extra left/right context for TDNN systems.
+chunk_left_context=0
+chunk_right_context=0
+tdnn_dim=450
+# training options
+srand=0
+remove_egs=true
+lang_test=lang_test
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+e2echain_model_dir=$exp_dir/chain/e2e_cnn_1a
+ali_dir=$exp_dir/chain/e2e_ali_train
+lat_dir=$exp_dir/chain${nnet3_affix}/e2e_${train_set}_lats
+dir=$exp_dir/chain${nnet3_affix}/cnn_e2eali${affix}
+train_data_dir=$data_dir/${train_set}
+tree_dir=$exp_dir/chain${nnet3_affix}/tree_e2e
+
+# the 'lang' directory is created by this script.
+# If you create such a directory with a non-standard topology
+# you should probably name it differently.
+lang=$data_dir/lang_chain
+for f in $train_data_dir/feats.scp $ali_dir/ali.1.gz $ali_dir/final.mdl; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+
+if [ $stage -le 1 ]; then
+  echo "$0: creating lang directory $lang with chain-type topology"
+  cp -r $data_dir/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 2 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/nnet3/align_lats.sh --nj $nj --cmd "$cmd" \
+                            --acoustic-scale 1.0 \
+                            --scale-opts '--transition-scale=1.0 --self-loop-scale=1.0' \
+                            ${train_data_dir} $data_dir/lang $e2echain_model_dir $lat_dir
+  echo "" >$lat_dir/splice_opts
+
+fi
+
+if [ $stage -le 3 ]; then
+  # Build a tree using our new topology.  We know we have alignments for the
+  # speed-perturbed data (local/nnet3/run_ivector_common.sh made them), so use
+  # those.  The num-leaves is always somewhat less than the num-leaves from
+  # the GMM baseline.
+  if [ -f $tree_dir/final.mdl ]; then
+    echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
+  fi
+
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor $frame_subsampling_factor \
+    --alignment-subsampling-factor 1 \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$cmd" $num_leaves ${train_data_dir} \
+    $lang $ali_dir $tree_dir
+fi
+
+
+if [ $stage -le 4 ]; then
+  mkdir -p $dir
+  echo "$0: creating neural net configs using the xconfig parser";
+  num_targets=$(tree-info $tree_dir/tree | grep num-pdfs | awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  cnn_opts="l2-regularize=0.075"
+  tdnn_opts="l2-regularize=0.075"
+  output_opts="l2-regularize=0.1"
+  common1="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=32"
+  common2="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=128"
+  common3="$cnn_opts required-time-offsets= height-offsets=-1,0,1 num-filters-out=512"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=180 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=60 height-out=60 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=60 height-out=60 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn3 height-in=60 height-out=30 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn4 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn5 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn6 height-in=30 height-out=15 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn7 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  conv-relu-batchnorm-layer name=cnn8 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  conv-relu-batchnorm-layer name=cnn9 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  relu-batchnorm-layer name=tdnn1 input=Append(-8,-4,0,4,8) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5 $tdnn_opts
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' mod?els... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn3 dim=$tdnn_dim target-rms=0.5 $tdnn_opts
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5 $output_opts
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+
+if [ $stage -le 5 ]; then
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{3,4,5,6}/$USER/kaldi-data/egs/iam-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/chain/train.py --stage=$train_stage \
+    --cmd="$cmd" \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=1 \
+    --chain.left-tolerance 3 \
+    --chain.right-tolerance 3 \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=6 \
+    --trainer.frames-per-iter=1000000 \
+    --trainer.optimization.num-jobs-initial=4 \
+    --trainer.optimization.num-jobs-final=8 \
+    --trainer.optimization.initial-effective-lrate=0.001 \
+    --trainer.optimization.final-effective-lrate=0.0001 \
+    --trainer.optimization.shrink-value=1.0 \
+    --trainer.num-chunk-per-minibatch=16,8 \
+    --trainer.optimization.momentum=0.0 \
+    --egs.chunk-width=$chunk_width \
+    --egs.chunk-left-context=$chunk_left_context \
+    --egs.chunk-right-context=$chunk_right_context \
+    --egs.chunk-left-context-initial=0 \
+    --egs.chunk-right-context-final=0 \
+    --egs.dir="$common_egs_dir" \
+    --egs.opts="--frames-overlap-per-eg 0 --constrained false" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=wait \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --tree-dir=$tree_dir \
+    --lat-dir=$lat_dir \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 6 ]; then
+  # The reason we are using data/lang here, instead of $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 $data_dir/$lang_test \
+    $dir $dir/graph || exit 1;
+fi
+
+if [ $stage -le 7 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+    --extra-left-context $chunk_left_context \
+    --extra-right-context $chunk_right_context \
+    --extra-left-context-initial 0 \
+    --extra-right-context-final 0 \
+    --frames-per-chunk $frames_per_chunk \
+    --nj $nj --cmd "$cmd" \
+    $dir/graph $data_dir/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/casia_hwdb/v1/local/chain/run_flatstart_cnn1a.sh b/egs/casia_hwdb/v1/local/chain/run_flatstart_cnn1a.sh
new file mode 100755
index 00000000000..023fbff1c14
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/chain/run_flatstart_cnn1a.sh
@@ -0,0 +1,170 @@
+#!/bin/bash
+# Copyright    2017  Hossein Hadian
+
+# This script does end2end chain training (i.e. from scratch)
+
+# ./local/chain/compare_wer.sh exp_yomdle_chinese/chain/e2e_cnn_1a exp_yomdle_chinese/chain/cnn_e2eali_1b
+# System                      e2e_cnn_1a cnn_e2eali_1b
+# CER                             15.44     13.57
+# Final train prob               0.0616   -0.0512
+# Final valid prob               0.0390   -0.0718
+# Final train prob (xent)                 -0.6199
+# Final valid prob (xent)                 -0.7448
+
+set -e
+
+data_dir=data
+exp_dir=exp
+
+# configs for 'chain'
+stage=0
+nj=30
+train_stage=-10
+get_egs_stage=-10
+affix=1a
+
+# training options
+tdnn_dim=450
+num_epochs=4
+num_jobs_initial=4
+num_jobs_final=8
+minibatch_size=150=64,32/300=32,16/600=16,8/1200=8,4
+common_egs_dir=
+l2_regularize=0.00005
+frames_per_iter=1000000
+cmvn_opts="--norm-means=false --norm-vars=false"
+train_set=train
+lang_test=lang_test
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+lang=$data_dir/lang_e2e
+treedir=$exp_dir/chain/e2e_monotree  # it's actually just a trivial tree (no tree building)
+dir=$exp_dir/chain/e2e_cnn_${affix}
+
+if [ $stage -le 0 ]; then
+  # Create a version of the lang/ directory that has one state per phone in the
+  # topo file. [note, it really has two states.. the first one is only repeated
+  # once, the second one has zero or more repeats.]
+  rm -rf $lang
+  cp -r $data_dir/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 1 ]; then
+  steps/nnet3/chain/e2e/prepare_e2e.sh --nj $nj --cmd "$cmd" \
+                                       --shared-phones true \
+                                       --type mono \
+                                       $data_dir/$train_set $lang $treedir
+  $cmd $treedir/log/make_phone_lm.log \
+  cat $data_dir/$train_set/text \| \
+    steps/nnet3/chain/e2e/text_to_phones.py $data_dir/lang \| \
+    utils/sym2int.pl -f 2- $data_dir/lang/phones.txt \| \
+    chain-est-phone-lm --num-extra-lm-states=1500 \
+                       ark:- $treedir/phone_lm.fst
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+  num_targets=$(tree-info $treedir/tree | grep num-pdfs | awk '{print $2}')
+  
+  cnn_opts="l2-regularize=0.075"
+  tdnn_opts="l2-regularize=0.075"
+  output_opts="l2-regularize=0.1"
+
+  common1="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=32"
+  common2="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=128"
+  common3="$cnn_opts required-time-offsets= height-offsets=-1,0,1 num-filters-out=512"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=180 name=input
+  conv-relu-batchnorm-layer name=cnn1 height-in=60 height-out=60 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=60 height-out=30 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn3 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn4 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn5 height-in=30 height-out=15 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn6 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  conv-relu-batchnorm-layer name=cnn7 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  relu-batchnorm-layer name=tdnn1 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5 $output_opts
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs
+fi
+
+if [ $stage -le 3 ]; then
+  # no need to store the egs in a shared storage because we always
+  # remove them. Anyway, it takes only 5 minutes to generate them.
+
+  steps/nnet3/chain/e2e/train_e2e.py --stage $train_stage \
+    --cmd "$cmd" \
+    --feat.cmvn-opts "$cmvn_opts" \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize $l2_regularize \
+    --chain.apply-deriv-weights false \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--num_egs_diagnostic 100 --num_utts_subset 400" \
+    --chain.frame-subsampling-factor 4 \
+    --chain.alignment-subsampling-factor 4 \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter $frames_per_iter \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.momentum 0 \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.shrink-value 1.0 \
+    --trainer.max-param-change 2.0 \
+    --cleanup.remove-egs true \
+    --use-gpu=wait \
+    --feat-dir $data_dir/${train_set} \
+    --tree-dir $treedir \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 4 ]; then
+  # The reason we are using data/lang here, instead of $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 $data_dir/$lang_test \
+    $dir $dir/graph || exit 1;
+fi
+
+if [ $stage -le 5 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+    --nj $nj --cmd "$cmd" \
+    $dir/graph $data_dir/test $dir/decode_test || exit 1;
+fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir
diff --git a/egs/casia_hwdb/v1/local/extract_database.sh b/egs/casia_hwdb/v1/local/extract_database.sh
new file mode 100755
index 00000000000..1af3713d586
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/extract_database.sh
@@ -0,0 +1,28 @@
+#!/bin/bash
+# Copyright 2018 Chun-Chieh Chang
+
+# The original format of the dataset given is GEDI and page images.
+# This script is written to create line images from page images.
+# It also creates csv files from the GEDI files.
+
+database_train=/export/corpora5/handwriting_ocr/CASIA_HWDB/Offline/
+database_competition=/export/corpora5/handwriting_ocr/CASIA_HWDB/Offline/
+cangjie_url=https://raw.githubusercontent.com/wanleung/libcangjie/master/tables/cj5-cc.txt
+download_dir=download
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1; 
+
+mkdir -p ${download_dir}/{Train,Test}
+for task in 0 1 2; do
+    for datasplit in Train Test; do
+        unzip -q -d ${download_dir}/${datasplit} ${database_train}/CASIA-HWDB2.${task}/${datasplit}_Dgr.zip
+    done
+done
+
+unzip -q -d ${download_dir}/Competition ${database_competition}/competition-dgr.zip 
+
+echo "Downloading table for CangJie."
+wget -P $download_dir/ $cangjie_url || exit 1;
+sed -ie '1,8d' $download_dir/cj5-cc.txt
diff --git a/egs/casia_hwdb/v1/local/extract_features.sh b/egs/casia_hwdb/v1/local/extract_features.sh
new file mode 100755
index 00000000000..f75837ae5b3
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/extract_features.sh
@@ -0,0 +1,45 @@
+#!/bin/bash
+# Copyright   2017 Yiwen Shao
+#             2018 Ashish Arora
+
+nj=4
+cmd=run.pl
+feat_dim=40
+fliplr=false
+augment='no_aug'
+num_channels=3
+echo "$0 $@"
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1;
+
+data=$1
+featdir=$data/data
+scp=$data/images.scp
+logdir=$data/log
+
+mkdir -p $logdir
+mkdir -p $featdir
+
+# make $featdir an absolute pathname
+featdir=`perl -e '($dir,$pwd)= @ARGV; if($dir!~m:^/:) { $dir = "$pwd/$dir"; } print $dir; ' $featdir ${PWD}`
+
+for n in $(seq $nj); do
+    split_scps="$split_scps $logdir/images.$n.scp"
+done
+
+# split images.scp
+utils/split_scp.pl $scp $split_scps || exit 1;
+
+$cmd JOB=1:$nj $logdir/extract_features.JOB.log \
+  image/ocr/make_features.py $logdir/images.JOB.scp \
+    --allowed_len_file_path $data/allowed_lengths.txt \
+    --feat-dim $feat_dim --num-channels $num_channels --fliplr $fliplr --augment_type $augment \| \
+    copy-feats --compress=true --compression-method=7 \
+    ark:- ark,scp:$featdir/images.JOB.ark,$featdir/images.JOB.scp
+
+## aggregates the output scp's to get feats.scp
+for n in $(seq $nj); do
+  cat $featdir/images.$n.scp || exit 1;
+done > $data/feats.scp || exit 1
diff --git a/egs/casia_hwdb/v1/local/gen_topo.py b/egs/casia_hwdb/v1/local/gen_topo.py
new file mode 100755
index 00000000000..f64dcc5eec1
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/gen_topo.py
@@ -0,0 +1,91 @@
+#!/usr/bin/env python
+
+# Copyright 2017 (author: Chun-Chieh Chang)
+
+# Generate a topology file.  This allows control of the number of states in the
+# non-silence HMMs, and in the silence HMMs. This is a modified version of
+# 'utils/gen_topo.pl'. The difference is that this creates two topologies for
+# the non-silence HMMs. The number of states for punctuations is different than
+# the number of states for other characters.
+
+from __future__ import print_function
+import argparse
+import string
+
+parser = argparse.ArgumentParser(description="Usage: steps/nnet3/chain/gen_topo.py "
+                                             "<colon-separated-nonsilence-phones> <colon-separated-silence-phones>"
+                                             "e.g.:  steps/nnet3/chain/gen_topo.pl 4:5:6:7:8:9:10 1:2:3\n",
+                                 epilog="See egs/swbd/s5c/local/chain/train_tdnn_a.sh for example of usage.");
+parser.add_argument("num_nonsil_states", type=int, help="number of states for nonsilence phones");
+parser.add_argument("num_sil_states", type=int, help="number of states for silence phones");
+parser.add_argument("num_cj5_states", type=int, help="number of states for punctuation");
+parser.add_argument("nonsilence_phones", type=str,
+                    help="List of non-silence phones as integers, separated by colons, e.g. 4:5:6:7:8:9");
+parser.add_argument("silence_phones", type=str,
+                    help="List of silence phones as integers, separated by colons, e.g. 1:2:3");
+parser.add_argument("phone_list", type=str, help="file containing all phones and their corresponding number.");
+
+args = parser.parse_args()
+
+silence_phones = [ int(x) for x in args.silence_phones.split(":") ]
+nonsilence_phones = [ int(x) for x in args.nonsilence_phones.split(":") ]
+all_phones = silence_phones +  nonsilence_phones
+
+cj5_phones = []
+with open(args.phone_list) as f:
+    for line in f:
+        line = line.strip()
+        phone = line.split(' ')[0]
+        if "cj5" in phone:
+            cj5_phones.append(int(line.split(' ')[1]))
+# For nonsilence phones that are not punctuations
+print("<Topology>")
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in nonsilence_phones if x not in cj5_phones]))
+print("</ForPhones>")
+for x in range(0, args.num_nonsil_states):
+    xp1 = x + 1
+    print("<State> " + str(x) + " <PdfClass> " + str(x) + " <Transition> " + str(x) + " 0.75 <Transition> " + str(xp1) + " 0.25 </State>")
+print("<State> " + str(args.num_nonsil_states) + " </State>")
+print("</TopologyEntry>")
+
+# For nonsilence phones that are cj5
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in nonsilence_phones if x in cj5_phones]))
+print("</ForPhones>")
+for x in range(0, args.num_cj5_states):
+    xp1 = x + 1
+    print("<State> " + str(x) + " <PdfClass> " + str(x) + " <Transition> " + str(x) + " 0.75 <Transition> " + str(xp1) + " 0.25 </State>")
+print("<State> " + str(args.num_cj5_states) + " </State>")
+print("</TopologyEntry>")
+
+# For silence phones
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in silence_phones]))
+print("</ForPhones>")
+if(args.num_sil_states > 1):
+    transp = 1.0 / (args.num_sil_states - 1)
+    
+    state_str = "<State> 0 <PdfClass> 0 "
+    for x in range(0, (args.num_sil_states - 1)):
+        state_str = state_str + "<Transition> " + str(x) + " " + str(transp) + " "
+    state_str = state_str + "</State>"
+    print(state_str)
+
+    for x in range(1, (args.num_sil_states - 1)):
+        state_str = "<State> " + str(x) + " <PdfClass> " + str(x) + " "
+        for y in range(1, args.num_sil_states):
+            state_str = state_str + "<Transition> " + str(y) + " " + str(transp) + " "
+        state_str = state_str + "</State>"
+        print(state_str)
+    second_last = args.num_sil_states - 1
+    print("<State> " + str(second_last) + " <PdfClass> " + str(second_last) + " <Transition> " + str(second_last) + " 0.75 <Transition> " + str(args.num_sil_states) + " 0.25 </State>")
+    print("<State> " + str(args.num_sil_states) + " </State>")
+else:
+    print("<State> 0 <PdfClass> 0 <Transition> 0 0.75 <Transition> 1 0.25 </State>")
+    print("<State> " + str(args.num_sil_states) + " </State>")
+print("</TopologyEntry>")
+print("</Topology>")
diff --git a/egs/casia_hwdb/v1/local/normalize_text.py b/egs/casia_hwdb/v1/local/normalize_text.py
new file mode 100755
index 00000000000..80c4e3ad3ab
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/normalize_text.py
@@ -0,0 +1,17 @@
+#!/usr/bin/env python3
+# Copyright   2018 Chun-Chieh Chang
+
+# This script reads in text and outputs the normalized version
+
+import io
+import re
+import sys
+import unicodedata
+
+sys.stdin = io.TextIOWrapper(sys.stdin.buffer, encoding="utf8")
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf8")
+for line in sys.stdin:
+    line = line.strip()
+    line = unicodedata.normalize('NFC', line)
+    line = re.sub(r'\s', ' ', line)
+    sys.stdout.write(line + '\n')
diff --git a/egs/casia_hwdb/v1/local/prepare_dict.sh b/egs/casia_hwdb/v1/local/prepare_dict.sh
new file mode 100755
index 00000000000..cf2ecb1ce9b
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/prepare_dict.sh
@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+
+# Copyright      2017  Hossein Hadian
+#                2017  Chun Chieh Chang
+#                2017  Ashish Arora
+
+# This script prepares the dictionary.
+
+set -e
+dir=data/local/dict
+data_dir=data
+
+. ./utils/parse_options.sh || exit 1;
+
+base_dir=$(echo "$DIRECTORY" | cut -d "/" -f2)
+
+mkdir -p $dir
+
+cut -d' ' -f1 download/cj5-cc.txt | ./utils/lang/bpe/learn_bpe.py -s 300 > $dir/bpe.out
+cut -d' ' -f1 download/cj5-cc.txt | ./utils/lang/bpe/apply_bpe.py -c $dir/bpe.out | sed 's/@@//g' > $dir/bpe_text
+cut -d' ' -f2- download/cj5-cc.txt | sed 's/ //g' > $dir/ids
+paste -d' ' $dir/bpe_text $dir/ids > $dir/cj5-cc.txt
+local/prepare_lexicon.py --data-dir $data_dir $dir
+
+cut -d' ' -f2- $dir/lexicon.txt | sed 's/SIL//g' | tr ' ' '\n' | sort -u | sed '/^$/d' >$dir/nonsilence_phones.txt || exit 1;
+
+echo '<sil> SIL' >> $dir/lexicon.txt
+
+echo SIL > $dir/silence_phones.txt
+
+echo SIL >$dir/optional_silence.txt
+
+echo -n "" >$dir/extra_questions.txt
diff --git a/egs/casia_hwdb/v1/local/prepare_lexicon.py b/egs/casia_hwdb/v1/local/prepare_lexicon.py
new file mode 100755
index 00000000000..224c199ecef
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/prepare_lexicon.py
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+
+# Copyright  2018  Ashish Arora
+#                  Chun-Chieh Chang
+
+import argparse
+import os
+
+parser = argparse.ArgumentParser(description="""Creates the list of characters and words in lexicon""")
+parser.add_argument('dir', type=str, help='output path')
+parser.add_argument('--data-dir', type=str, default='data', help='Path to text file')
+args = parser.parse_args()
+
+### main ###
+radical = ['日', '月', '金', '木', '水', '火', '土', '竹', '戈', '十', '大', '中', '一', '弓', '人', '心', '手','口','尸','廿','山','女','田','卜']
+lex = {}
+text_path = os.path.join(args.data_dir, 'train', 'text')
+text_fh = open(text_path, 'r', encoding='utf-8')
+
+# Used specially for Chinese.
+# Uses the ChangJie keyboard input method to create subword units for Chinese.
+cj5_table = {}
+with open(os.path.join(args.dir, 'cj5-cc.txt'), 'r', encoding='utf-8') as f:
+    for line in f:
+        line_vect = line.strip().split()
+        if not line_vect[0].startswith('yyy') and not line_vect[0].startswith('z'):
+            cj5_table[line_vect[-1]] = "cj5_" + " cj5_".join(line_vect[:-1])
+#            lex[line_vect[1]] = "cj5_" + " cj5_".join(list(line_vect[0]))
+
+with open(text_path, 'r', encoding='utf-8') as f:
+    for line in f:
+        line_vect = line.strip().split()
+        for i in range(1, len(line_vect)):
+            characters = list(line_vect[i])
+	    # Put SIL instead of "|". Because every "|" in the beginning of the words is for initial-space of that word
+            characters = " ".join([ 'SIL' if char == '|' else char if char in radical else cj5_table[char] if char in cj5_table else char for char in characters])
+            characters = characters.replace('#','<HASH>')
+            lex[line_vect[i]] = characters
+
+with open(os.path.join(args.dir, 'lexicon.txt'), 'w', encoding='utf-8') as fp:
+    for key in sorted(lex):
+        fp.write(key + " " + lex[key] + "\n")
diff --git a/egs/casia_hwdb/v1/local/process_data.py b/egs/casia_hwdb/v1/local/process_data.py
new file mode 100755
index 00000000000..8548ac2c58e
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/process_data.py
@@ -0,0 +1,464 @@
+#!/usr/bin/env python3
+
+# Copyright      2018  Ashish Arora
+#                2018  Chun Chieh Chang
+
+""" This script reads the extracted Farsi OCR (yomdle and slam) database files 
+    and creates the following files (for the data subset selected via --dataset):
+    text, utt2spk, images.scp.
+  Eg. local/process_data.py data/download/ data/local/splits/train.txt data/train
+  Eg. text file: english_phone_books_0001_1 To sum up, then, it would appear that
+      utt2spk file: english_phone_books_0001_0 english_phone_books_0001
+      images.scp file: english_phone_books_0001_0 \
+      data/download/truth_line_image/english_phone_books_0001_0.png
+"""
+
+import argparse
+import numpy as np
+import os
+import re
+import struct
+import sys
+import unicodedata
+from collections import namedtuple
+from math import atan2, cos, sin, pi, degrees, sqrt
+from PIL import Image
+from scipy import misc
+from scipy.spatial import ConvexHull
+
+parser = argparse.ArgumentParser(description="Creates text, utt2spk, and images.scp files")
+parser.add_argument('database_path', type=str, help='Path to data')
+parser.add_argument('out_dir', type=str, help='directory to output files')
+parser.add_argument('--padding', type=int, default=100, help='Padding so BBox does not exceed image area')
+args = parser.parse_args()
+
+"""
+bounding_box is a named tuple which contains:
+             area (float): area of the rectangle
+             length_parallel (float): length of the side that is parallel to unit_vector
+             length_orthogonal (float): length of the side that is orthogonal to unit_vector
+             rectangle_center(int, int): coordinates of the rectangle center
+             (use rectangle_corners to get the corner points of the rectangle)
+             unit_vector (float, float): direction of the length_parallel side.
+             (it's orthogonal vector can be found with the orthogonal_vector function
+             unit_vector_angle (float): angle of the unit vector to be in radians.
+             corner_points [(float, float)]: set that contains the corners of the rectangle
+"""
+
+bounding_box_tuple = namedtuple('bounding_box_tuple', 'area '
+                                        'length_parallel '
+                                        'length_orthogonal '
+                                        'rectangle_center '
+                                        'unit_vector '
+                                        'unit_vector_angle '
+                                        'corner_points')
+
+
+def unit_vector(pt0, pt1):
+    """ Given two points pt0 and pt1, return a unit vector that
+        points in the direction of pt0 to pt1.
+    Returns
+    -------
+    (float, float): unit vector
+    """
+    dis_0_to_1 = sqrt((pt0[0] - pt1[0])**2 + (pt0[1] - pt1[1])**2)
+    return (pt1[0] - pt0[0]) / dis_0_to_1, \
+           (pt1[1] - pt0[1]) / dis_0_to_1
+
+
+def orthogonal_vector(vector):
+    """ Given a vector, returns a orthogonal/perpendicular vector of equal length.
+    Returns
+    ------
+    (float, float): A vector that points in the direction orthogonal to vector.
+    """
+    return -1 * vector[1], vector[0]
+
+def bounding_area(index, hull):
+    """ Given index location in an array and convex hull, it gets two points
+        hull[index] and hull[index+1]. From these two points, it returns a named
+        tuple that mainly contains area of the box that bounds the hull. This
+        bounding box orintation is same as the orientation of the lines formed
+        by the point hull[index] and hull[index+1].
+    Returns
+    -------
+    a named tuple that contains:
+    area: area of the rectangle
+    length_parallel: length of the side that is parallel to unit_vector
+    length_orthogonal: length of the side that is orthogonal to unit_vector
+    rectangle_center: coordinates of the rectangle center
+    unit_vector: direction of the length_parallel side.
+    (it's orthogonal vector can be found with the orthogonal_vector function)
+    """
+    unit_vector_p = unit_vector(hull[index], hull[index+1])
+    unit_vector_o = orthogonal_vector(unit_vector_p)
+
+    dis_p = tuple(np.dot(unit_vector_p, pt) for pt in hull)
+    dis_o = tuple(np.dot(unit_vector_o, pt) for pt in hull)
+
+    min_p = min(dis_p)
+    min_o = min(dis_o)
+    len_p = max(dis_p) - min_p
+    len_o = max(dis_o) - min_o
+
+    return {'area': len_p * len_o,
+            'length_parallel': len_p,
+            'length_orthogonal': len_o,
+            'rectangle_center': (min_p + float(len_p) / 2, min_o + float(len_o) / 2),
+            'unit_vector': unit_vector_p}
+
+def to_xy_coordinates(unit_vector_angle, point):
+    """ Given angle from horizontal axis and a point from origin,
+        returns converted unit vector coordinates in x, y coordinates.
+        angle of unit vector should be in radians.
+    Returns
+    ------
+    (float, float): converted x,y coordinate of the unit vector.
+    """
+    angle_orthogonal = unit_vector_angle + pi / 2
+    return point[0] * cos(unit_vector_angle) + point[1] * cos(angle_orthogonal), \
+           point[0] * sin(unit_vector_angle) + point[1] * sin(angle_orthogonal)
+
+def rotate_points(center_of_rotation, angle, points):
+    """ Rotates a point cloud around the center_of_rotation point by angle
+    input
+    -----
+    center_of_rotation (float, float): angle of unit vector to be in radians.
+    angle (float): angle of rotation to be in radians.
+    points [(float, float)]: Points to be a list or tuple of points. Points to be rotated.
+    Returns
+    ------
+    [(float, float)]: Rotated points around center of rotation by angle
+    """
+    rot_points = []
+    ang = []
+    for pt in points:
+        diff = tuple([pt[d] - center_of_rotation[d] for d in range(2)])
+        diff_angle = atan2(diff[1], diff[0]) + angle
+        ang.append(diff_angle)
+        diff_length = sqrt(sum([d**2 for d in diff]))
+        rot_points.append((center_of_rotation[0] + diff_length * cos(diff_angle),
+                           center_of_rotation[1] + diff_length * sin(diff_angle)))
+
+    return rot_points
+
+def rectangle_corners(rectangle):
+    """ Given rectangle center and its inclination, returns the corner
+        locations of the rectangle.
+    Returns
+    ------
+    [(float, float)]: 4 corner points of rectangle.
+    """
+    corner_points = []
+    for i1 in (.5, -.5):
+        for i2 in (i1, -1 * i1):
+            corner_points.append((rectangle['rectangle_center'][0] + i1 * rectangle['length_parallel'],
+                            rectangle['rectangle_center'][1] + i2 * rectangle['length_orthogonal']))
+
+    return rotate_points(rectangle['rectangle_center'], rectangle['unit_vector_angle'], corner_points)
+
+def get_orientation(origin, p1, p2):
+    """
+    Given origin and two points, return the orientation of the Point p1 with
+    regards to Point p2 using origin.
+    Returns
+    -------
+    integer: Negative if p1 is clockwise of p2.
+    """
+    difference = (
+        ((p2[0] - origin[0]) * (p1[1] - origin[1]))
+        - ((p1[0] - origin[0]) * (p2[1] - origin[1]))
+    )
+    return difference
+
+def compute_hull(points):
+    """
+    Given input list of points, return a list of points that
+    made up the convex hull.
+    Returns
+    -------
+    [(float, float)]: convexhull points
+    """
+    hull_points = []
+    start = points[0]
+    min_x = start[0]
+    for p in points[1:]:
+        if p[0] < min_x:
+            min_x = p[0]
+            start = p
+    point = start
+    hull_points.append(start)
+
+    far_point = None
+    while far_point is not start:
+        p1 = None
+        for p in points:
+            if p is point:
+                continue
+            else:
+                p1 = p
+                break
+
+        far_point = p1
+
+        for p2 in points:
+            if p2 is point or p2 is p1:
+                continue
+            else:
+                direction = get_orientation(point, far_point, p2)
+                if direction > 0:
+                    far_point = p2
+
+        hull_points.append(far_point)
+        point = far_point
+    return hull_points
+
+
+def minimum_bounding_box(points):
+    """ Given a list of 2D points, it returns the minimum area rectangle bounding all
+        the points in the point cloud.
+    Returns
+    ------
+    returns a namedtuple that contains:
+    area: area of the rectangle
+    length_parallel: length of the side that is parallel to unit_vector
+    length_orthogonal: length of the side that is orthogonal to unit_vector
+    rectangle_center: coordinates of the rectangle center
+    unit_vector: direction of the length_parallel side. RADIANS
+    unit_vector_angle: angle of the unit vector
+    corner_points: set that contains the corners of the rectangle
+    """
+    if len(points) <= 2: raise ValueError('More than two points required.')
+
+    hull_ordered = [points[index] for index in ConvexHull(points).vertices]
+    hull_ordered.append(hull_ordered[0])
+    #hull_ordered = compute_hull(points)
+    hull_ordered = tuple(hull_ordered)
+
+    min_rectangle = bounding_area(0, hull_ordered)
+    for i in range(1, len(hull_ordered)-1):
+        rectangle = bounding_area(i, hull_ordered)
+        if rectangle['area'] < min_rectangle['area']:
+            min_rectangle = rectangle
+
+    min_rectangle['unit_vector_angle'] = atan2(min_rectangle['unit_vector'][1], min_rectangle['unit_vector'][0])
+    min_rectangle['rectangle_center'] = to_xy_coordinates(min_rectangle['unit_vector_angle'], min_rectangle['rectangle_center'])
+
+    return bounding_box_tuple(
+        area = min_rectangle['area'],
+        length_parallel = min_rectangle['length_parallel'],
+        length_orthogonal = min_rectangle['length_orthogonal'],
+        rectangle_center = min_rectangle['rectangle_center'],
+        unit_vector = min_rectangle['unit_vector'],
+        unit_vector_angle = min_rectangle['unit_vector_angle'],
+        corner_points = set(rectangle_corners(min_rectangle)))
+
+def get_center(im):
+    """ Given image, returns the location of center pixel
+    Returns
+    -------
+    (int, int): center of the image
+    """
+    center_x = float(im.size[0]) / 2
+    center_y = float(im.size[1]) / 2
+    return int(center_x), int(center_y)
+
+def get_horizontal_angle(unit_vector_angle):
+    """ Given an angle in radians, returns angle of the unit vector in
+        first or fourth quadrant.
+    Returns
+    ------
+    (float): updated angle of the unit vector to be in radians.
+             It is only in first or fourth quadrant.
+    """
+    if unit_vector_angle > pi / 2 and unit_vector_angle <= pi:
+        unit_vector_angle = unit_vector_angle - pi
+    elif unit_vector_angle > -pi and unit_vector_angle < -pi / 2:
+        unit_vector_angle = unit_vector_angle + pi
+
+    return unit_vector_angle
+
+def get_smaller_angle(bounding_box):
+    """ Given a rectangle, returns its smallest absolute angle from horizontal axis.
+    Returns
+    ------
+    (float): smallest angle of the rectangle to be in radians.
+    """
+    unit_vector = bounding_box.unit_vector
+    unit_vector_angle = bounding_box.unit_vector_angle
+    ortho_vector = orthogonal_vector(unit_vector)
+    ortho_vector_angle = atan2(ortho_vector[1], ortho_vector[0])
+
+    unit_vector_angle_updated = get_horizontal_angle(unit_vector_angle)
+    ortho_vector_angle_updated = get_horizontal_angle(ortho_vector_angle)
+
+    if abs(unit_vector_angle_updated) < abs(ortho_vector_angle_updated):
+        return unit_vector_angle_updated
+    else:
+        return ortho_vector_angle_updated
+
+def rotated_points(bounding_box, center):
+    """ Given the rectangle, returns corner points of rotated rectangle.
+        It rotates the rectangle around the center by its smallest angle.
+    Returns
+    -------
+    [(int, int)]: 4 corner points of rectangle.
+    """
+    p1, p2, p3, p4 = bounding_box.corner_points
+    x1, y1 = p1
+    x2, y2 = p2
+    x3, y3 = p3
+    x4, y4 = p4
+    center_x, center_y = center
+    rotation_angle_in_rad = -get_smaller_angle(bounding_box)
+    x_dash_1 = (x1 - center_x) * cos(rotation_angle_in_rad) - (y1 - center_y) * sin(rotation_angle_in_rad) + center_x
+    x_dash_2 = (x2 - center_x) * cos(rotation_angle_in_rad) - (y2 - center_y) * sin(rotation_angle_in_rad) + center_x
+    x_dash_3 = (x3 - center_x) * cos(rotation_angle_in_rad) - (y3 - center_y) * sin(rotation_angle_in_rad) + center_x
+    x_dash_4 = (x4 - center_x) * cos(rotation_angle_in_rad) - (y4 - center_y) * sin(rotation_angle_in_rad) + center_x
+
+    y_dash_1 = (y1 - center_y) * cos(rotation_angle_in_rad) + (x1 - center_x) * sin(rotation_angle_in_rad) + center_y
+    y_dash_2 = (y2 - center_y) * cos(rotation_angle_in_rad) + (x2 - center_x) * sin(rotation_angle_in_rad) + center_y
+    y_dash_3 = (y3 - center_y) * cos(rotation_angle_in_rad) + (x3 - center_x) * sin(rotation_angle_in_rad) + center_y
+    y_dash_4 = (y4 - center_y) * cos(rotation_angle_in_rad) + (x4 - center_x) * sin(rotation_angle_in_rad) + center_y
+    return x_dash_1, y_dash_1, x_dash_2, y_dash_2, x_dash_3, y_dash_3, x_dash_4, y_dash_4
+
+def pad_image(image):
+    """ Given an image, returns a padded image around the border.
+        This routine save the code from crashing if bounding boxes that are
+        slightly outside the page boundary.
+    Returns
+    -------
+    image: page image
+    """
+    offset = int(args.padding // 2)
+    padded_image = Image.new('RGB', (image.size[0] + int(args.padding), image.size[1] + int(args.padding)), "white")
+    padded_image.paste(im = image, box = (offset, offset))
+    return padded_image
+
+def update_minimum_bounding_box_input(bounding_box_input):
+    """ Given list of 2D points, returns list of 2D points shifted by an offset.
+    Returns
+    ------
+    points [(float, float)]: points, a list or tuple of 2D coordinates
+    """
+    updated_minimum_bounding_box_input = []
+    offset = int(args.padding // 2)
+    for point in bounding_box_input:
+        x, y = point
+        new_x = x + offset
+        new_y = y + offset
+        word_coordinate = (new_x, new_y)
+        updated_minimum_bounding_box_input.append(word_coordinate)
+
+    return updated_minimum_bounding_box_input
+
+### main ###
+print("Processing '{}' data...".format(args.out_dir))
+
+text_file = os.path.join(args.out_dir, 'text')
+text_fh = open(text_file, 'w', encoding='utf-8')
+utt2spk_file = os.path.join(args.out_dir, 'utt2spk')
+utt2spk_fh = open(utt2spk_file, 'w', encoding='utf-8')
+image_file = os.path.join(args.out_dir, 'images.scp')
+image_fh = open(image_file, 'w', encoding='utf-8')
+
+for filename in sorted(os.listdir(args.database_path)):
+    if filename.endswith('.dgr'):
+        with open(os.path.join(args.database_path, filename), 'rb')  as f:
+            iHdSize = struct.unpack('i', f.read(4))[0]
+            szFormatCode = struct.unpack(''.join('c' for x in range(0,8)), f.read(8))
+            szFormatCode = "".join([x.decode('utf8') for x in szFormatCode])
+            szIllustr = f.read(iHdSize - 36)
+            szCodeType = struct.unpack(''.join(['c' for x in range(0,20)]), f.read(20))
+            szCodeType = "".join([x.decode('utf8') for x in szCodeType])
+            sCodeLen = struct.unpack('h', f.read(2))[0]
+            sBitApp = struct.unpack('h', f.read(2))[0]
+            iImgHei = struct.unpack('i', f.read(4))[0]
+            iImgWid = struct.unpack('i', f.read(4))[0]
+            pDocImg = Image.new('L', (iImgWid, iImgHei), (255))
+            iLineNum = struct.unpack('i', f.read(4))[0]
+            text_dict = {}
+            image_dict = {}
+            for i in range(0, iLineNum):
+                iWordNum = struct.unpack('i', f.read(4))[0]
+                for j in range(0, iWordNum):
+                    pWordLabel = f.read(sCodeLen).decode('gb18030', errors='ignore')
+                    sTop = struct.unpack('h', f.read(2))[0]
+                    sLeft = struct.unpack('h', f.read(2))[0]
+                    sHei = struct.unpack('h', f.read(2))[0]
+                    sWid = struct.unpack('h', f.read(2))[0]
+                    if i in text_dict:
+                        text_dict[i] += [pWordLabel]
+                    else:
+                        text_dict[i] = [pWordLabel]
+                    if i in image_dict:
+                        image_dict[i] += [[sTop, sLeft, sHei, sWid]]
+                    else:
+                        image_dict[i] = [[sTop, sLeft, sHei, sWid]]
+                    pTmpData = struct.unpack("{}B".format(sHei * sWid), f.read(sHei * sWid))
+                    character = misc.toimage(np.array(pTmpData).reshape(sHei, sWid))
+                    pDocImg.paste(character, (sLeft, sTop))
+            pDocImg.save(os.path.join(args.out_dir, 'data', 'images', os.path.splitext(filename)[0] + '.png'), 'png')
+            
+            im_page = pad_image(pDocImg)
+            for i in range(0, iLineNum):
+                text = ""
+                points = []
+                for j, char in enumerate(text_dict[i]):
+                    text += char
+                    points.append([image_dict[i][j][1], image_dict[i][j][0]])
+                    points.append([image_dict[i][j][1] + image_dict[i][j][3], image_dict[i][j][0]])
+                    points.append([image_dict[i][j][1], image_dict[i][j][0] + image_dict[i][j][2]])
+                    points.append([image_dict[i][j][1] + image_dict[i][j][3], image_dict[i][j][0] + image_dict[i][j][2]])
+                updated_mbb_input = update_minimum_bounding_box_input(points)
+                bounding_box = minimum_bounding_box(updated_mbb_input)
+                p1, p2, p3, p4 = bounding_box.corner_points 
+                x1, y1 = p1
+                x2, y2 = p2
+                x3, y3 = p3
+                x4, y4 = p4
+                min_x = int(min(x1, x2, x3, x4))
+                min_y = int(min(y1, y2, y3, y4))
+                max_x = int(max(x1, x2, x3, x4))
+                max_y = int(max(y1, y2, y3, y4))
+                box = (min_x, min_y, max_x, max_y)
+                region_initial = im_page.crop(box)
+                rot_points = []
+                p1_new = (x1 - min_x, y1 - min_y)
+                p2_new = (x2 - min_x, y2 - min_y)
+                p3_new = (x3 - min_x, y3 - min_y)
+                p4_new = (x4 - min_x, y4 - min_y)
+                rot_points.append(p1_new)
+                rot_points.append(p2_new)
+                rot_points.append(p3_new)
+                rot_points.append(p4_new)
+                cropped_bounding_box = bounding_box_tuple(bounding_box.area,
+                    bounding_box.length_parallel,
+                    bounding_box.length_orthogonal,
+                    bounding_box.length_orthogonal,
+                    bounding_box.unit_vector,
+                    bounding_box.unit_vector_angle,
+                    set(rot_points))
+
+                rotation_angle_in_rad = get_smaller_angle(cropped_bounding_box)
+                img2 = region_initial.rotate(degrees(rotation_angle_in_rad), resample=Image.BICUBIC)
+                x_dash_1, y_dash_1, x_dash_2, y_dash_2, x_dash_3, y_dash_3, x_dash_4, y_dash_4 = rotated_points(
+                    cropped_bounding_box, get_center(region_initial))
+
+
+                min_x = int(min(x_dash_1, x_dash_2, x_dash_3, x_dash_4))
+                min_y = int(min(y_dash_1, y_dash_2, y_dash_3, y_dash_4))
+                max_x = int(max(x_dash_1, x_dash_2, x_dash_3, x_dash_4))
+                max_y = int(max(y_dash_1, y_dash_2, y_dash_3, y_dash_4))
+                box = (min_x, min_y, max_x, max_y)
+                region_final = img2.crop(box)
+                text = text.replace('\x00', '')
+                text = unicodedata.normalize('NFC', text)
+                image_id = os.path.splitext(filename)[0] + '_' + str(i).zfill(3)
+                image_filepath = os.path.join(args.out_dir, 'data', 'images', os.path.splitext(filename)[0] + '_' + str(i).zfill(3) + '.png')
+                writer_id = os.path.splitext(filename)[0].split('-')[0]
+                region_final.save(image_filepath, 'png')
+                
+                text_fh.write(image_id + ' ' + text + '\n')
+                utt2spk_fh.write(image_id + ' ' + writer_id + '\n')
+                image_fh.write(image_id + ' ' + image_filepath + '\n')
diff --git a/egs/casia_hwdb/v1/local/score.sh b/egs/casia_hwdb/v1/local/score.sh
new file mode 100755
index 00000000000..f2405205f02
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/score.sh
@@ -0,0 +1,5 @@
+#!/bin/bash
+
+
+steps/scoring/score_kaldi_wer.sh --max-lmwt 10 "$@"
+steps/scoring/score_kaldi_cer.sh --max-lmwt 10 --stage 2 "$@"
diff --git a/egs/casia_hwdb/v1/local/train_lm.sh b/egs/casia_hwdb/v1/local/train_lm.sh
new file mode 100755
index 00000000000..bc738f217da
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/train_lm.sh
@@ -0,0 +1,110 @@
+#!/bin/bash
+
+# Copyright 2016  Vincent Nguyen
+#           2016  Johns Hopkins University (author: Daniel Povey)
+#           2017  Ashish Arora
+#           2017  Hossein Hadian
+# Apache 2.0
+#
+# This script trains a LM on the YOMDLE training transcriptions.
+# It is based on the example scripts distributed with PocoLM
+
+# It will check if pocolm is installed and if not will proceed with installation
+
+set -e
+stage=0
+dir=data/local/local_lm
+data_dir=data
+
+echo "$0 $@"  # Print the command line for logging
+. ./utils/parse_options.sh || exit 1;
+
+lm_dir=${dir}/data
+
+
+mkdir -p $dir
+. ./path.sh || exit 1; # for KALDI_ROOT
+export PATH=$KALDI_ROOT/tools/pocolm/scripts:$PATH
+( # First make sure the pocolm toolkit is installed.
+ cd $KALDI_ROOT/tools || exit 1;
+ if [ -d pocolm ]; then
+   echo Not installing the pocolm toolkit since it is already there.
+ else
+   echo "$0: Please install the PocoLM toolkit with: "
+   echo " cd ../../../tools; extras/install_pocolm.sh; cd -"
+   exit 1;
+ fi
+) || exit 1;
+
+bypass_metaparam_optim_opt=
+# If you want to bypass the metaparameter optimization steps with specific metaparameters
+# un-comment the following line, and change the numbers to some appropriate values.
+# You can find the values from output log of train_lm.py.
+# These example numbers of metaparameters is for 4-gram model (with min-counts)
+# running with train_lm.py.
+# The dev perplexity should be close to the non-bypassed model.
+#bypass_metaparam_optim_opt=
+# Note: to use these example parameters, you may need to remove the .done files
+# to make sure the make_lm_dir.py be called and tain only 3-gram model
+#for order in 3; do
+#rm -f ${lm_dir}/${num_word}_${order}.pocolm/.done
+
+if [ $stage -le 0 ]; then
+  mkdir -p ${dir}/data
+  mkdir -p ${dir}/data/text
+
+  echo "$0: Getting the Data sources"
+
+  rm ${dir}/data/text/* 2>/dev/null || true
+
+  # Note: the name 'dev' is treated specially by pocolm, it automatically
+  # becomes the dev set.
+  nr=`cat $data_dir/train/text | wc -l`
+  nr_dev=$(($nr / 10 ))
+  nr_train=$(( $nr - $nr_dev ))
+
+  # use the training data as an additional data source.
+  # we can later fold the dev data into this.
+  head -n $nr_train $data_dir/train/text | cut -d " " -f 2- >  ${dir}/data/text/train.txt
+  tail -n $nr_dev $data_dir/train/text | cut -d " " -f 2- > ${dir}/data/text/dev.txt
+
+  # for reporting perplexities, we'll use the "real" dev set.
+  # (the validation data is used as ${dir}/data/text/dev.txt to work
+  # out interpolation weights.)
+  # note, we can't put it in ${dir}/data/text/, because then pocolm would use
+  # it as one of the data sources.
+  cut -d " " -f 2-  < $data_dir/test/text  > ${dir}/data/real_dev_set.txt
+
+  # get the wordlist from MADCAT text
+  cat ${dir}/data/text/train.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | sort | uniq -c | sort -bnr > ${dir}/data/word_count
+  cat ${dir}/data/word_count | awk '{print $2}' > ${dir}/data/wordlist
+fi
+
+order=3
+
+if [ $stage -le 1 ]; then
+  # decide on the vocabulary.
+  # Note: you'd use --wordlist if you had a previously determined word-list
+  # that you wanted to use.
+  # Note: if you have more than one order, use a certain amount of words as the
+  # vocab and want to restrict max memory for 'sort',
+  echo "$0: training the unpruned LM"
+  min_counts='train=1'
+  wordlist=${dir}/data/wordlist
+
+  lm_name="`basename ${wordlist}`_${order}"
+  if [ -n "${min_counts}" ]; then
+    lm_name+="_`echo ${min_counts} | tr -s "[:blank:]" "_" | tr "=" "-"`"
+  fi
+  unpruned_lm_dir=${lm_dir}/${lm_name}.pocolm
+  train_lm.py  --wordlist=${wordlist} --num-splits=5 --warm-start-ratio=1 \
+               --min-counts="$min_counts" \
+               --limit-unk-history=true \
+               ${bypass_metaparam_optim_opt} \
+               ${dir}/data/text ${order} ${lm_dir}/work ${unpruned_lm_dir}
+
+  get_data_prob.py ${dir}/data/real_dev_set.txt ${unpruned_lm_dir} 2>&1 | grep -F '[perplexity'
+
+  mkdir -p ${dir}/data/arpa
+  format_arpa_lm.py ${unpruned_lm_dir} | gzip -c > ${dir}/data/arpa/${order}gram_unpruned.arpa.gz
+fi
diff --git a/egs/casia_hwdb/v1/local/train_lm_lr.sh b/egs/casia_hwdb/v1/local/train_lm_lr.sh
new file mode 100755
index 00000000000..a8b1bfb76a4
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/train_lm_lr.sh
@@ -0,0 +1,117 @@
+#!/bin/bash
+
+# Copyright 2016  Vincent Nguyen
+#           2016  Johns Hopkins University (author: Daniel Povey)
+#           2017  Ashish Arora
+#           2017  Hossein Hadian
+# Apache 2.0
+#
+# This script trains a LM on the YOMDLE+Extra training transcriptions.
+# It is based on the example scripts distributed with PocoLM
+
+# It will check if pocolm is installed and if not will proceed with installation
+
+set -e
+stage=0
+dir=data/local/local_lm
+data_dir=data
+extra_lm=download/extra_lm.txt
+order=3
+
+echo "$0 $@"  # Print the command line for logging
+. ./utils/parse_options.sh || exit 1;
+
+lm_dir=${dir}/data
+
+
+mkdir -p $dir
+. ./path.sh || exit 1; # for KALDI_ROOT
+export PATH=$KALDI_ROOT/tools/pocolm/scripts:$PATH
+( # First make sure the pocolm toolkit is installed.
+ cd $KALDI_ROOT/tools || exit 1;
+ if [ -d pocolm ]; then
+   echo Not installing the pocolm toolkit since it is already there.
+ else
+   echo "$0: Please install the PocoLM toolkit with: "
+   echo " cd ../../../tools; extras/install_pocolm.sh; cd -"
+   exit 1;
+ fi
+) || exit 1;
+
+bypass_metaparam_optim_opt=
+# If you want to bypass the metaparameter optimization steps with specific metaparameters
+# un-comment the following line, and change the numbers to some appropriate values.
+# You can find the values from output log of train_lm.py.
+# These example numbers of metaparameters is for 4-gram model (with min-counts)
+# running with train_lm.py.
+# The dev perplexity should be close to the non-bypassed model.
+#bypass_metaparam_optim_opt=
+# Note: to use these example parameters, you may need to remove the .done files
+# to make sure the make_lm_dir.py be called and tain only 3-gram model
+#for order in 3; do
+#rm -f ${lm_dir}/${num_word}_${order}.pocolm/.done
+
+if [ $stage -le 0 ]; then
+  mkdir -p ${dir}/data
+  mkdir -p ${dir}/data/text
+
+  echo "$0: Getting the Data sources"
+
+  rm ${dir}/data/text/* 2>/dev/null || true
+
+  cat ${extra_lm} | \
+    local/normalize_text.py | \
+    utils/lang/bpe/prepend_words.py | \
+    python3 utils/lang/bpe/apply_bpe.py -c $data_dir/train/bpe.out | \
+    sed 's/@@//g' > ${dir}/data/text/extra_lm.txt
+  
+  # Note: the name 'dev' is treated specially by pocolm, it automatically
+  # becomes the dev set.
+  nr=`cat $data_dir/train/text | wc -l`
+  nr_dev=$(($nr / 10 ))
+  nr_train=$(( $nr - $nr_dev ))
+
+  # use the training data as an additional data source.
+  # we can later fold the dev data into this.
+  head -n $nr_train $data_dir/train/text | cut -d " " -f 2- >  ${dir}/data/text/train.txt
+  tail -n $nr_dev $data_dir/train/text | cut -d " " -f 2- > ${dir}/data/text/dev.txt
+
+  # for reporting perplexities, we'll use the "real" dev set.
+  # (the validation data is used as ${dir}/data/text/dev.txt to work
+  # out interpolation weights.)
+  # note, we can't put it in ${dir}/data/text/, because then pocolm would use
+  # it as one of the data sources.
+  cut -d " " -f 2-  < $data_dir/test/text  > ${dir}/data/real_dev_set.txt
+
+  # get the wordlist from MADCAT text
+  cat ${dir}/data/text/{train,extra_lm}.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | sort | uniq -c | sort -bnr > ${dir}/data/word_count
+  #cat ${dir}/data/text/extra_fa.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | sort | uniq -c | sort -bnr > ${dir}/data/word_count
+  cat ${dir}/data/word_count | awk '{print $2}' > ${dir}/data/wordlist
+fi
+
+if [ $stage -le 1 ]; then
+  # decide on the vocabulary.
+  # Note: you'd use --wordlist if you had a previously determined word-list
+  # that you wanted to use.
+  # Note: if you have more than one order, use a certain amount of words as the
+  # vocab and want to restrict max memory for 'sort',
+  echo "$0: training the unpruned LM"
+  min_counts='extra_lm=10 train=1'
+  wordlist=${dir}/data/wordlist
+
+  lm_name="`basename ${wordlist}`_${order}"
+  if [ -n "${min_counts}" ]; then
+    lm_name+="_`echo ${min_counts} | tr -s "[:blank:]" "_" | tr "=" "-"`"
+  fi
+  unpruned_lm_dir=${lm_dir}/${lm_name}.pocolm
+  train_lm.py  --wordlist=${wordlist} --num-splits=30 --warm-start-ratio=1 \
+               --min-counts="$min_counts" \
+               --limit-unk-history=true \
+               ${bypass_metaparam_optim_opt} \
+               ${dir}/data/text ${order} ${lm_dir}/work ${unpruned_lm_dir}
+
+  get_data_prob.py ${dir}/data/real_dev_set.txt ${unpruned_lm_dir} 2>&1 | grep -F '[perplexity'
+
+  mkdir -p ${dir}/data/arpa
+  format_arpa_lm.py ${unpruned_lm_dir} | gzip -c > ${dir}/data/arpa/${order}gram_unpruned.arpa.gz
+fi
diff --git a/egs/casia_hwdb/v1/local/wer_output_filter b/egs/casia_hwdb/v1/local/wer_output_filter
new file mode 100755
index 00000000000..8702738144f
--- /dev/null
+++ b/egs/casia_hwdb/v1/local/wer_output_filter
@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+# Copyright      2017  Hossein Hadian
+
+# This is a filter used in scoring. It separates all
+# punctuations from words. For e.g. this sentence:
+
+# "They have come!" he said reverently, gripping his
+# hands. "Isn't it a glorious thing! Long awaited."
+
+# is converted to this:
+
+# " They have come ! " he said reverently , gripping his
+# hands . " Isn ' t it a glorious thing ! Long awaited . "
+
+# Sample BPE-based output:
+# |He |ro se |from |his |b re ak f as t - s ch oo l |b en ch
+
+import io
+import sys
+import unicodedata
+
+sys.stdin = io.TextIOWrapper(sys.stdin.buffer, encoding="utf8");
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf8");
+for line in sys.stdin:
+  line = unicodedata.normalize('NFKC', line)
+  words = line.strip().split()
+  uttid = words[0]
+  transcript = ''.join(words[1:])
+  transcript = transcript.replace('|', ' ')
+  print("{} {}".format(uttid, transcript))
diff --git a/egs/casia_hwdb/v1/path.sh b/egs/casia_hwdb/v1/path.sh
new file mode 100644
index 00000000000..2d17b17a84a
--- /dev/null
+++ b/egs/casia_hwdb/v1/path.sh
@@ -0,0 +1,6 @@
+export KALDI_ROOT=`pwd`/../../..
+[ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
+export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$PWD:$PATH
+[ ! -f $KALDI_ROOT/tools/config/common_path.sh ] && echo >&2 "The standard file $KALDI_ROOT/tools/config/common_path.sh is not present -> Exit!" && exit 1
+. $KALDI_ROOT/tools/config/common_path.sh
+export LC_ALL=C
diff --git a/egs/casia_hwdb/v1/run.sh b/egs/casia_hwdb/v1/run.sh
new file mode 100755
index 00000000000..44d1f26117c
--- /dev/null
+++ b/egs/casia_hwdb/v1/run.sh
@@ -0,0 +1,110 @@
+#!/bin/bash
+
+set -e
+stage=0
+nj=60
+
+database_train=/export/corpora5/handwriting_ocr/CASIA_HWDB/Offline/
+database_competition=/export/corpora5/handwriting_ocr/CASIA_HWDB/Offline/
+data_dir=data
+exp_dir=exp
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if [ $stage -le -1 ]; then
+    mkdir download/Train
+    mkdir download/Test
+    mkdir download/Competition
+    local/extract_database.sh --database-train $database_train \
+        --database-competition $database_competition
+fi
+
+if [ $stage -le 0 ]; then
+    mkdir -p data/train/data/images
+    mkdir -p data/test/data/images
+    mkdir -p data/competition/data/images
+    local/process_data.py download/Train data/train
+    local/process_data.py download/Test data/test
+    local/process_data.py download/Competition data/competition
+    image/fix_data_dir.sh ${data_dir}/test
+    image/fix_data_dir.sh ${data_dir}/train
+    image/fix_data_dir.sh ${data_dir}/competition
+fi
+
+mkdir -p $data_dir/{train,test}/data
+if [ $stage -le 1 ]; then
+    echo "$0: Obtaining image groups. calling get_image2num_frames"
+    echo "Date: $(date)."
+    image/get_image2num_frames.py --feat-dim 60 $data_dir/train
+    image/get_allowed_lengths.py --frame-subsampling-factor 4 10 $data_dir/train
+
+    for datasplit in train test competition; do
+        echo "$0: Extracting features and calling compute_cmvn_stats for dataset: $datasplit. "
+        echo "Date: $(date)."
+        local/extract_features.sh --nj $nj --cmd "$cmd" \
+            --feat-dim 60 --num-channels 3 \
+            $data_dir/${datasplit}
+        steps/compute_cmvn_stats.sh $data_dir/${datasplit} || exit 1;
+    done
+
+    echo "$0: Fixing data directory for train dataset"
+    echo "Date: $(date)."
+    utils/fix_data_dir.sh $data_dir/train
+fi
+
+#if [ $stage -le 2 ]; then
+#    for datasplit in train; do
+#        echo "$(date) stage 2: Performing augmentation, it will double training data"
+#        local/augment_data.sh --nj $nj --cmd "$cmd" --feat-dim 60 $data_dir/${datasplit} $data_dir/${datasplit}_aug $data_dir
+#        steps/compute_cmvn_stats.sh $data_dir/${datasplit}_aug || exit 1;
+#    done
+#fi
+
+if [ $stage -le 3 ]; then
+    echo "$0: Preparing dictionary and lang..."
+    if [ ! -f $data_dir/train/bpe.out ]; then
+        cut -d' ' -f2- $data_dir/train/text | utils/lang/bpe/prepend_words.py | python3 utils/lang/bpe/learn_bpe.py -s 700 > $data_dir/train/bpe.out
+        for datasplit in test train; do
+            cut -d' ' -f1 $data_dir/$datasplit/text > $data_dir/$datasplit/ids
+            cut -d' ' -f2- $data_dir/$datasplit/text | utils/lang/bpe/prepend_words.py | python3 utils/lang/bpe/apply_bpe.py -c $data_dir/train/bpe.out | sed 's/@@//g' > $data_dir/$datasplit/bpe_text
+            mv $data_dir/$datasplit/text $data_dir/$datasplit/text.old
+            paste -d' ' $data_dir/$datasplit/ids $data_dir/$datasplit/bpe_text > $data_dir/$datasplit/text
+        done
+    fi
+
+    local/prepare_dict.sh --data-dir $data_dir --dir $data_dir/local/dict
+    # This recipe uses byte-pair encoding, the silences are part of the words' pronunciations.
+    # So we set --sil-prob to 0.0
+    utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 8 --sil-prob 0.0 --position-dependent-phones false \
+        $data_dir/local/dict "<sil>" $data_dir/lang/temp $data_dir/lang
+    utils/lang/bpe/add_final_optional_silence.sh --final-sil-prob 0.5 $data_dir/lang
+fi
+
+if [ $stage -le 4 ]; then
+    echo "$0: Estimating a language model for decoding..."
+    local/train_lm.sh --data-dir $data_dir  --dir $data_dir/local/local_lm
+    utils/format_lm.sh $data_dir/lang $data_dir/local/local_lm/data/arpa/3gram_unpruned.arpa.gz \
+        $data_dir/local/dict/lexicon.txt $data_dir/lang_test
+fi
+
+if [ $stage -le 5 ]; then
+    echo "$0: Calling the flat-start chain recipe..."
+    echo "Date: $(date)." 
+    local/chain/run_flatstart_cnn1a.sh --nj $nj --train-set train --data-dir $data_dir --exp-dir $exp_dir
+fi
+
+if [ $stage -le 6 ]; then
+    echo "$0: Aligning the training data using the e2e chain model..."
+    echo "Date: $(date)."
+    steps/nnet3/align.sh --nj $nj --cmd "$cmd" --use-gpu false \
+        --scale-opts '--transition-scale=1.0 --acoustic-scale=1.0 --self-loop-scale=1.0' \
+        $data_dir/train $data_dir/lang $exp_dir/chain/e2e_cnn_1a $exp_dir/chain/e2e_ali_train
+fi
+
+if [ $stage -le 7 ]; then
+    echo "$0: Building a tree and training a regular chain model using the e2e alignments..."
+    echo "Date: $(date)."
+    local/chain/run_cnn_e2eali_1b.sh --nj $nj --train-set train --data-dir $data_dir --exp-dir $exp_dir
+fi
diff --git a/egs/casia_hwdb/v1/steps b/egs/casia_hwdb/v1/steps
new file mode 120000
index 00000000000..6e99bf5b5ad
--- /dev/null
+++ b/egs/casia_hwdb/v1/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps
\ No newline at end of file
diff --git a/egs/casia_hwdb/v1/utils b/egs/casia_hwdb/v1/utils
new file mode 120000
index 00000000000..b240885218f
--- /dev/null
+++ b/egs/casia_hwdb/v1/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils
\ No newline at end of file
diff --git a/egs/yomdle_zh/v1/local/chain/run_cnn_e2eali_1b.sh b/egs/yomdle_zh/v1/local/chain/run_cnn_e2eali_1b.sh
index 0a4e00d7aed..357ce6a1f8e 100755
--- a/egs/yomdle_zh/v1/local/chain/run_cnn_e2eali_1b.sh
+++ b/egs/yomdle_zh/v1/local/chain/run_cnn_e2eali_1b.sh
@@ -5,11 +5,12 @@
 
 # ./local/chain/compare_wer.sh exp_yomdle_chinese/chain/e2e_cnn_1a exp_yomdle_chinese/chain/cnn_e2eali_1b
 # System                      e2e_cnn_1a cnn_e2eali_1b
-# CER                             15.44     13.57
-# Final train prob               0.0616   -0.0512
-# Final valid prob               0.0390   -0.0718
-# Final train prob (xent)                 -0.6199
-# Final valid prob (xent)                 -0.7448
+# WER                             63.19     53.67
+# CER                             19.01     12.86
+# Final train prob               0.2908   -0.0455
+# Final valid prob               0.2397   -0.0531
+# Final train prob (xent)                 -0.9753
+# Final valid prob (xent)                 -1.0559
 
 set -e -o pipefail
 
@@ -31,7 +32,7 @@ xent_regularize=0.1
 frame_subsampling_factor=4
 # training chunk-options
 chunk_width=340,300,200,100
-num_leaves=1000
+num_leaves=2000
 # we don't need extra left/right context for TDNN systems.
 chunk_left_context=0
 chunk_right_context=0
@@ -75,25 +76,12 @@ done
 
 if [ $stage -le 1 ]; then
   echo "$0: creating lang directory $lang with chain-type topology"
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  if [ -d $lang ]; then
-    if [ $lang/L.fst -nt $data_dir/lang/L.fst ]; then
-      echo "$0: $lang already exists, not overwriting it; continuing"
-    else
-      echo "$0: $lang already exists and seems to be older than data/lang..."
-      echo " ... not sure what to do.  Exiting."
-      exit 1;
-    fi
-  else
-    cp -r $data_dir/lang $lang
-    silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-    nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-    # Use our special topology... note that later on may have to tune this
-    # topology.
-    steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-  fi
+  cp -r $data_dir/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
 fi
 
 if [ $stage -le 2 ]; then
@@ -114,13 +102,12 @@ if [ $stage -le 3 ]; then
   # the GMM baseline.
   if [ -f $tree_dir/final.mdl ]; then
     echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
-    exit 1;
   fi
 
   steps/nnet3/chain/build_tree.sh \
     --frame-subsampling-factor $frame_subsampling_factor \
     --alignment-subsampling-factor 1 \
-    --context-opts "--context-width=3 --central-position=1" \
+    --context-opts "--context-width=2 --central-position=1" \
     --cmd "$cmd" $num_leaves ${train_data_dir} \
     $lang $ali_dir $tree_dir
 fi
@@ -187,7 +174,7 @@ if [ $stage -le 5 ]; then
     --chain.leaky-hmm-coefficient=0.1 \
     --chain.l2-regularize=0.00005 \
     --chain.apply-deriv-weights=false \
-    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=500" \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1500" \
     --chain.frame-subsampling-factor=$frame_subsampling_factor \
     --chain.alignment-subsampling-factor=1 \
     --chain.left-tolerance 3 \
@@ -211,7 +198,7 @@ if [ $stage -le 5 ]; then
     --egs.dir="$common_egs_dir" \
     --egs.opts="--frames-overlap-per-eg 0 --constrained false" \
     --cleanup.remove-egs=$remove_egs \
-    --use-gpu=true \
+    --use-gpu=wait \
     --reporting.email="$reporting_email" \
     --feat-dir=$train_data_dir \
     --tree-dir=$tree_dir \
diff --git a/egs/yomdle_zh/v1/local/chain/run_flatstart_cnn1a.sh b/egs/yomdle_zh/v1/local/chain/run_flatstart_cnn1a.sh
index 88bbd32790c..28ea2863e38 100755
--- a/egs/yomdle_zh/v1/local/chain/run_flatstart_cnn1a.sh
+++ b/egs/yomdle_zh/v1/local/chain/run_flatstart_cnn1a.sh
@@ -5,11 +5,12 @@
 
 # ./local/chain/compare_wer.sh exp_yomdle_chinese/chain/e2e_cnn_1a exp_yomdle_chinese/chain/cnn_e2eali_1b
 # System                      e2e_cnn_1a cnn_e2eali_1b
-# CER                             15.44     13.57
-# Final train prob               0.0616   -0.0512
-# Final valid prob               0.0390   -0.0718
-# Final train prob (xent)                 -0.6199
-# Final valid prob (xent)                 -0.7448
+# WER                             63.19     53.67
+# CER                             19.01     12.86
+# Final train prob               0.2908   -0.0455
+# Final valid prob               0.2397   -0.0531
+# Final train prob (xent)                 -0.9753
+# Final valid prob (xent)                 -1.0559
 
 set -e
 
@@ -77,7 +78,7 @@ if [ $stage -le 1 ]; then
   cat $data_dir/$train_set/text \| \
     steps/nnet3/chain/e2e/text_to_phones.py $data_dir/lang \| \
     utils/sym2int.pl -f 2- $data_dir/lang/phones.txt \| \
-    chain-est-phone-lm --num-extra-lm-states=500 \
+    chain-est-phone-lm --num-extra-lm-states=1500 \
                        ark:- $treedir/phone_lm.fst
 fi
 
@@ -140,6 +141,7 @@ if [ $stage -le 3 ]; then
     --trainer.optimization.shrink-value 1.0 \
     --trainer.max-param-change 2.0 \
     --cleanup.remove-egs true \
+    --use-gpu=wait \
     --feat-dir $data_dir/${train_set} \
     --tree-dir $treedir \
     --dir $dir  || exit 1;
diff --git a/egs/yomdle_zh/v1/local/gen_topo.py b/egs/yomdle_zh/v1/local/gen_topo.py
new file mode 100755
index 00000000000..f64dcc5eec1
--- /dev/null
+++ b/egs/yomdle_zh/v1/local/gen_topo.py
@@ -0,0 +1,91 @@
+#!/usr/bin/env python
+
+# Copyright 2017 (author: Chun-Chieh Chang)
+
+# Generate a topology file.  This allows control of the number of states in the
+# non-silence HMMs, and in the silence HMMs. This is a modified version of
+# 'utils/gen_topo.pl'. The difference is that this creates two topologies for
+# the non-silence HMMs. The number of states for punctuations is different than
+# the number of states for other characters.
+
+from __future__ import print_function
+import argparse
+import string
+
+parser = argparse.ArgumentParser(description="Usage: steps/nnet3/chain/gen_topo.py "
+                                             "<colon-separated-nonsilence-phones> <colon-separated-silence-phones>"
+                                             "e.g.:  steps/nnet3/chain/gen_topo.pl 4:5:6:7:8:9:10 1:2:3\n",
+                                 epilog="See egs/swbd/s5c/local/chain/train_tdnn_a.sh for example of usage.");
+parser.add_argument("num_nonsil_states", type=int, help="number of states for nonsilence phones");
+parser.add_argument("num_sil_states", type=int, help="number of states for silence phones");
+parser.add_argument("num_cj5_states", type=int, help="number of states for punctuation");
+parser.add_argument("nonsilence_phones", type=str,
+                    help="List of non-silence phones as integers, separated by colons, e.g. 4:5:6:7:8:9");
+parser.add_argument("silence_phones", type=str,
+                    help="List of silence phones as integers, separated by colons, e.g. 1:2:3");
+parser.add_argument("phone_list", type=str, help="file containing all phones and their corresponding number.");
+
+args = parser.parse_args()
+
+silence_phones = [ int(x) for x in args.silence_phones.split(":") ]
+nonsilence_phones = [ int(x) for x in args.nonsilence_phones.split(":") ]
+all_phones = silence_phones +  nonsilence_phones
+
+cj5_phones = []
+with open(args.phone_list) as f:
+    for line in f:
+        line = line.strip()
+        phone = line.split(' ')[0]
+        if "cj5" in phone:
+            cj5_phones.append(int(line.split(' ')[1]))
+# For nonsilence phones that are not punctuations
+print("<Topology>")
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in nonsilence_phones if x not in cj5_phones]))
+print("</ForPhones>")
+for x in range(0, args.num_nonsil_states):
+    xp1 = x + 1
+    print("<State> " + str(x) + " <PdfClass> " + str(x) + " <Transition> " + str(x) + " 0.75 <Transition> " + str(xp1) + " 0.25 </State>")
+print("<State> " + str(args.num_nonsil_states) + " </State>")
+print("</TopologyEntry>")
+
+# For nonsilence phones that are cj5
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in nonsilence_phones if x in cj5_phones]))
+print("</ForPhones>")
+for x in range(0, args.num_cj5_states):
+    xp1 = x + 1
+    print("<State> " + str(x) + " <PdfClass> " + str(x) + " <Transition> " + str(x) + " 0.75 <Transition> " + str(xp1) + " 0.25 </State>")
+print("<State> " + str(args.num_cj5_states) + " </State>")
+print("</TopologyEntry>")
+
+# For silence phones
+print("<TopologyEntry>")
+print("<ForPhones>")
+print(" ".join([str(x) for x in silence_phones]))
+print("</ForPhones>")
+if(args.num_sil_states > 1):
+    transp = 1.0 / (args.num_sil_states - 1)
+    
+    state_str = "<State> 0 <PdfClass> 0 "
+    for x in range(0, (args.num_sil_states - 1)):
+        state_str = state_str + "<Transition> " + str(x) + " " + str(transp) + " "
+    state_str = state_str + "</State>"
+    print(state_str)
+
+    for x in range(1, (args.num_sil_states - 1)):
+        state_str = "<State> " + str(x) + " <PdfClass> " + str(x) + " "
+        for y in range(1, args.num_sil_states):
+            state_str = state_str + "<Transition> " + str(y) + " " + str(transp) + " "
+        state_str = state_str + "</State>"
+        print(state_str)
+    second_last = args.num_sil_states - 1
+    print("<State> " + str(second_last) + " <PdfClass> " + str(second_last) + " <Transition> " + str(second_last) + " 0.75 <Transition> " + str(args.num_sil_states) + " 0.25 </State>")
+    print("<State> " + str(args.num_sil_states) + " </State>")
+else:
+    print("<State> 0 <PdfClass> 0 <Transition> 0 0.75 <Transition> 1 0.25 </State>")
+    print("<State> " + str(args.num_sil_states) + " </State>")
+print("</TopologyEntry>")
+print("</Topology>")
diff --git a/egs/yomdle_zh/v1/local/normalize_text.py b/egs/yomdle_zh/v1/local/normalize_text.py
new file mode 100755
index 00000000000..80c4e3ad3ab
--- /dev/null
+++ b/egs/yomdle_zh/v1/local/normalize_text.py
@@ -0,0 +1,17 @@
+#!/usr/bin/env python3
+# Copyright   2018 Chun-Chieh Chang
+
+# This script reads in text and outputs the normalized version
+
+import io
+import re
+import sys
+import unicodedata
+
+sys.stdin = io.TextIOWrapper(sys.stdin.buffer, encoding="utf8")
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf8")
+for line in sys.stdin:
+    line = line.strip()
+    line = unicodedata.normalize('NFC', line)
+    line = re.sub(r'\s', ' ', line)
+    sys.stdout.write(line + '\n')
diff --git a/egs/yomdle_zh/v1/local/prepare_dict.sh b/egs/yomdle_zh/v1/local/prepare_dict.sh
index 65b2e7aa901..cf2ecb1ce9b 100755
--- a/egs/yomdle_zh/v1/local/prepare_dict.sh
+++ b/egs/yomdle_zh/v1/local/prepare_dict.sh
@@ -16,6 +16,10 @@ base_dir=$(echo "$DIRECTORY" | cut -d "/" -f2)
 
 mkdir -p $dir
 
+cut -d' ' -f1 download/cj5-cc.txt | ./utils/lang/bpe/learn_bpe.py -s 300 > $dir/bpe.out
+cut -d' ' -f1 download/cj5-cc.txt | ./utils/lang/bpe/apply_bpe.py -c $dir/bpe.out | sed 's/@@//g' > $dir/bpe_text
+cut -d' ' -f2- download/cj5-cc.txt | sed 's/ //g' > $dir/ids
+paste -d' ' $dir/bpe_text $dir/ids > $dir/cj5-cc.txt
 local/prepare_lexicon.py --data-dir $data_dir $dir
 
 cut -d' ' -f2- $dir/lexicon.txt | sed 's/SIL//g' | tr ' ' '\n' | sort -u | sed '/^$/d' >$dir/nonsilence_phones.txt || exit 1;
diff --git a/egs/yomdle_zh/v1/local/prepare_lexicon.py b/egs/yomdle_zh/v1/local/prepare_lexicon.py
index 3ebb52e38f4..6ac45fd735f 100755
--- a/egs/yomdle_zh/v1/local/prepare_lexicon.py
+++ b/egs/yomdle_zh/v1/local/prepare_lexicon.py
@@ -12,6 +12,7 @@
 args = parser.parse_args()
 
 ### main ###
+radical = ['日', '月', '金', '木', '水', '火', '土', '竹', '戈', '十', '大', '中', '一', '弓', '人', '心', '手','口','尸','廿','山','女','田','卜']
 lex = {}
 text_path = os.path.join(args.data_dir, 'train', 'text')
 text_fh = open(text_path, 'r', encoding='utf-8')
@@ -19,11 +20,11 @@
 # Used specially for Chinese.
 # Uses the ChangJie keyboard input method to create subword units for Chinese.
 cj5_table = {}
-with open('download/cj5-cc.txt', 'r', encoding='utf-8') as f:
+with open(os.path.join(args.dir, 'cj5-cc.txt'), 'r', encoding='utf-8') as f:
     for line in f:
         line_vect = line.strip().split()
-        if not line_vect[0].startswith('yyy') and not line_vect[0].startswith('z'):
-            cj5_table[line_vect[1]] = "cj5_" + " cj5_".join(list(line_vect[0]))
+        if not ''.join(line_vect[:-1]).startswith('yyy') and not ''.join(line_vect[:-1]).startswith('z'):
+            cj5_table[line_vect[-1]] = "cj5_" + " cj5_".join(line_vect[:-1])
 
 with open(text_path, 'r', encoding='utf-8') as f:
     for line in f:
@@ -31,7 +32,7 @@
         for i in range(1, len(line_vect)):
             characters = list(line_vect[i])
 	    # Put SIL instead of "|". Because every "|" in the beginning of the words is for initial-space of that word
-            characters = " ".join([ 'SIL' if char == '|' else cj5_table[char] if char in cj5_table else char for char in characters])
+            characters = " ".join([ 'SIL' if char == '|' else char if char in radical else cj5_table[char] if char in cj5_table else char for char in characters])
             characters = characters.replace('#','<HASH>')
             lex[line_vect[i]] = characters
 
diff --git a/egs/yomdle_zh/v1/local/process_data.py b/egs/yomdle_zh/v1/local/process_data.py
index 8964af8890a..eadb0052705 100755
--- a/egs/yomdle_zh/v1/local/process_data.py
+++ b/egs/yomdle_zh/v1/local/process_data.py
@@ -15,6 +15,7 @@
 
 import argparse
 import os
+import re
 import sys
 import csv
 import itertools
@@ -50,6 +51,7 @@
             image_id = os.path.splitext(row[1])[0]
             image_filepath = os.path.join(args.database_path, 'truth_line_image', row[1])
             text = unicodedata.normalize('NFC', row[11]).replace('\n', '')
+            text = re.sub(r'\s', ' ', text)
             if os.path.isfile(image_filepath) and os.stat(image_filepath).st_size != 0 and text:
                 text_fh.write(image_id + ' ' + text + '\n')
                 utt2spk_fh.write(image_id + ' ' + '_'.join(image_id.split('_')[:-1]) + '\n')
diff --git a/egs/yomdle_zh/v1/local/train_lm_lr.sh b/egs/yomdle_zh/v1/local/train_lm_lr.sh
index b95e6474b18..a8b1bfb76a4 100755
--- a/egs/yomdle_zh/v1/local/train_lm_lr.sh
+++ b/egs/yomdle_zh/v1/local/train_lm_lr.sh
@@ -59,7 +59,11 @@ if [ $stage -le 0 ]; then
 
   rm ${dir}/data/text/* 2>/dev/null || true
 
-  cat ${extra_lm} | utils/lang/bpe/prepend_words.py | python3 utils/lang/bpe/apply_bpe.py -c $data_dir/train/bpe.out | sed 's/@@//g' > ${dir}/data/text/extra_lm.txt
+  cat ${extra_lm} | \
+    local/normalize_text.py | \
+    utils/lang/bpe/prepend_words.py | \
+    python3 utils/lang/bpe/apply_bpe.py -c $data_dir/train/bpe.out | \
+    sed 's/@@//g' > ${dir}/data/text/extra_lm.txt
   
   # Note: the name 'dev' is treated specially by pocolm, it automatically
   # becomes the dev set.
diff --git a/egs/yomdle_zh/v1/local/wer_output_filter b/egs/yomdle_zh/v1/local/wer_output_filter
index 08d5563bca4..449484e1061 100755
--- a/egs/yomdle_zh/v1/local/wer_output_filter
+++ b/egs/yomdle_zh/v1/local/wer_output_filter
@@ -21,8 +21,8 @@ while (<>) {
     # Normalize dashes to regular hyphen
     $s =~ s/[\x{2010}\x{2011}\x{2012}\x{2013}\x{2014}\x{2015}]+/\x{002D}/g;
     # Normalize various parenthesis to regular parenthesis
-    $s =~ s/\x{UFF09}/\x{0029}/g;
-    $s =~ s/\x{UFF08}/\x{0028}/g;
+    $s =~ s/\x{FF09}/\x{0029}/g;
+    $s =~ s/\x{FF08}/\x{0028}/g;
     
     # Convert various presentation forms to base form
     $s =~ s/[\x{FED1}\x{FED3}\x{FED4}\x{FED2}]+/\x{0641}/g;
diff --git a/egs/yomdle_zh/v1/run.sh b/egs/yomdle_zh/v1/run.sh
index 128f15694cc..eb8e9e11927 100755
--- a/egs/yomdle_zh/v1/run.sh
+++ b/egs/yomdle_zh/v1/run.sh
@@ -77,8 +77,11 @@ if [ $stage -le 3 ]; then
     local/prepare_dict.sh --data-dir $data_dir --dir $data_dir/local/dict
     # This recipe uses byte-pair encoding, the silences are part of the words' pronunciations.
     # So we set --sil-prob to 0.0
-    utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 8 --sil-prob 0.0 --position-dependent-phones false \
+    utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 6 --sil-prob 0.0 --position-dependent-phones false \
         $data_dir/local/dict "<sil>" $data_dir/lang/temp $data_dir/lang
+    silphonelist=`cat $data_dir/lang/phones/silence.csl`
+    nonsilphonelist=`cat $data_dir/lang/phones/nonsilence.csl`
+    local/gen_topo.py 8 4 10 $nonsilphonelist $silphonelist $data_dir/lang/phones.txt > $data_dir/lang/topo 
     utils/lang/bpe/add_final_optional_silence.sh --final-sil-prob 0.5 $data_dir/lang
 fi