diff --git a/egs/iam/README.txt b/egs/iam/README.txt
new file mode 100644
index 00000000000..daeb67af541
--- /dev/null
+++ b/egs/iam/README.txt
@@ -0,0 +1,4 @@
+
+This directory contains example scripts for handwriting recognition on
+the IAM dataset:
+http://www.fki.inf.unibe.ch/databases/iam-handwriting-database
diff --git a/egs/iam/v1/cmd.sh b/egs/iam/v1/cmd.sh
new file mode 100644
index 00000000000..3c8eb9f93a5
--- /dev/null
+++ b/egs/iam/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/iam/v1/image b/egs/iam/v1/image
new file mode 120000
index 00000000000..1668ee99922
--- /dev/null
+++ b/egs/iam/v1/image
@@ -0,0 +1 @@
+../../cifar/v1/image/
\ No newline at end of file
diff --git a/egs/iam/v1/local/chain/compare_wer.sh b/egs/iam/v1/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..4eb665fc702
--- /dev/null
+++ b/egs/iam/v1/local/chain/compare_wer.sh
@@ -0,0 +1,59 @@
+#!/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
+
+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/iam/v1/local/chain/run_cnn_1a.sh b/egs/iam/v1/local/chain/run_cnn_1a.sh
new file mode 100755
index 00000000000..3b1571091c1
--- /dev/null
+++ b/egs/iam/v1/local/chain/run_cnn_1a.sh
@@ -0,0 +1,235 @@
+#!/bin/bash
+
+# Copyright    2017 Hossein Hadian
+#              2017 Chun Chieh Chang
+#              2017 Ashish Arora
+
+# steps/info/chain_dir_info.pl exp/chain/cnn_1a/
+# exp/chain/cnn_1a/: num-iters=21 nj=2..4 num-params=4.4M dim=40->364 combine=-0.021->-0.015 xent:train/valid[13,20,final]=(-1.05,-0.701,-0.591/-1.30,-1.08,-1.00) logprob:train/valid[13,20,final]=(-0.061,-0.034,-0.030/-0.107,-0.101,-0.098)
+
+# cat exp/chain/cnn_1a/decode_test/scoring_kaldi/best_*
+# %WER 5.94 [ 3913 / 65921, 645 ins, 1466 del, 1802 sub ] exp/chain/cnn_1a/decode_test//cer_11_0.0
+# %WER 9.13 [ 1692 / 18542, 162 ins, 487 del, 1043 sub ] exp/chain/cnn_1a/decode_test/wer_11_0.0
+
+set -e -o pipefail
+
+stage=0
+
+nj=30
+train_set=train
+gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
+                # should have alignments for the specified training data.
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+ali=tri3_ali
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=4
+alignment_subsampling_factor=1
+# training chunk-options
+chunk_width=340,300,200,100
+num_leaves=500
+# 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=false
+lang_test=lang_test
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${ali}
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_lats
+dir=exp/chain${nnet3_affix}/cnn${affix}
+train_data_dir=data/${train_set}
+tree_dir=exp/chain${nnet3_affix}/tree
+
+# 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/lang_chain
+
+for f in $train_data_dir/feats.scp \
+    $train_data_dir/feats.scp $gmm_dir/final.mdl \
+    $ali_dir/ali.1.gz $gmm_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"
+  # 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 [ $lang/L.fst -nt data/$lang_test/L.fst ]; then
+      echo "$0: $lang already exists, not overwriting it; continuing"
+    else
+      echo "$0: $lang already exists and seems to be older than data/lang..."
+      echo " ... not sure what to do.  Exiting."
+      exit 1;
+    fi
+  else
+    cp -r data/$lang_test $lang
+    silphonelist=$(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
+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/align_fmllr_lats.sh --nj $nj --cmd "$cmd" ${train_data_dir} \
+    data/$lang_test $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+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."
+     exit 1;
+  fi
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor $frame_subsampling_factor \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$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)
+  common1="height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=40 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn4 height-in=20 height-out=10 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  relu-batchnorm-layer name=tdnn1 input=Append(-4,-2,0,2,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn4 input=Append(-4,0,4) dim=$tdnn_dim
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' mod?els... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn4 dim=$tdnn_dim target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=$frame_subsampling_factor \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=1000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=4 \
+    --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=64,32 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$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 $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/$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/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/iam/v1/local/chain/run_cnn_chainali_1a.sh b/egs/iam/v1/local/chain/run_cnn_chainali_1a.sh
new file mode 100755
index 00000000000..2c8b6c91e5a
--- /dev/null
+++ b/egs/iam/v1/local/chain/run_cnn_chainali_1a.sh
@@ -0,0 +1,244 @@
+#!/bin/bash
+
+# chainali_1a is as 1a except it uses chain alignments (using 1a system) instead of gmm alignments
+
+# ./local/chain/compare_wer.sh exp/chain/cnn_chainali_1a/ exp/chain/cnn_1a/
+# System                      cnn_chainali_1a    cnn_1a
+# WER                             6.69     9.13
+# Final train prob              -0.0128   -0.0297
+# Final valid prob              -0.0447   -0.0975
+# Final train prob (xent)       -0.6448   -0.5915
+# Final valid prob (xent)       -0.9924   -1.0022
+
+# steps/info/chain_dir_info.pl exp/chain/cnn_chainali_1a/
+# exp/chain/cnn_chainali_1a/: num-iters=21 nj=2..4 num-params=4.4M dim=40->364 combine=-0.002->0.000 xent:train/valid[13,20,final]=(-0.929,-0.711,-0.645/-1.16,-1.04,-0.992) logprob:train/valid[13,20,final]=(-0.029,-0.016,-0.013/-0.051,-0.047,-0.045)
+
+# cat exp/chain/cnn_chainali_1a/decode_test/scoring_kaldi/best_*
+# %WER 3.94 [ 2600 / 65921, 549 ins, 837 del, 1214 sub ] exp/chain/cnn_chainali_1a/decode_test/cer_15_0.0
+# %WER 6.69 [ 1241 / 18542, 135 ins, 358 del, 748 sub ] exp/chain/cnn_chainali_1a/decode_test/wer_15_0.5
+
+set -e -o pipefail
+
+stage=0
+
+nj=30
+train_set=train
+gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
+                # should have alignments for the specified training data.
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+ali=tri3_ali
+chain_model_dir=exp/chain${nnet3_affix}/cnn${affix}
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=4
+alignment_subsampling_factor=1
+# training chunk-options
+chunk_width=340,300,200,100
+num_leaves=500
+# 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=false
+lang_test=lang_test
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${ali}
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_lats_chain
+gmm_lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_lats
+dir=exp/chain${nnet3_affix}/cnn_chainali${affix}
+train_data_dir=data/${train_set}
+tree_dir=exp/chain${nnet3_affix}/tree_chain
+
+# 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/lang_chain
+for f in $train_data_dir/feats.scp \
+    $train_data_dir/feats.scp $gmm_dir/final.mdl \
+    $ali_dir/ali.1.gz $gmm_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"
+  # 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 [ $lang/L.fst -nt data/$lang_test/L.fst ]; then
+      echo "$0: $lang already exists, not overwriting it; continuing"
+    else
+      echo "$0: $lang already exists and seems to be older than data/lang..."
+      echo " ... not sure what to do.  Exiting."
+      exit 1;
+    fi
+  else
+    cp -r data/$lang_test $lang
+    silphonelist=$(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
+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" \
+                            --scale-opts '--transition-scale=1.0 --self-loop-scale=1.0' \
+                            ${train_data_dir} data/$lang_test $chain_model_dir $lat_dir
+  cp $gmm_lat_dir/splice_opts $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."
+     exit 1;
+  fi
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor $frame_subsampling_factor \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$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)
+  common1="height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common3="height-offsets=-1,0,1 num-filters-out=70"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=40 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn4 height-in=20 height-out=10 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  relu-batchnorm-layer name=tdnn1 input=Append(-4,-2,0,2,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn4 input=Append(-4,0,4) dim=$tdnn_dim
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' mod?els... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn4 dim=$tdnn_dim target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=$alignment_subsampling_factor \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=1000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=4 \
+    --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=64,32 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$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 $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/$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/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/iam/v1/local/chain/run_cnn_chainali_1b.sh b/egs/iam/v1/local/chain/run_cnn_chainali_1b.sh
new file mode 100755
index 00000000000..ddf596a6126
--- /dev/null
+++ b/egs/iam/v1/local/chain/run_cnn_chainali_1b.sh
@@ -0,0 +1,245 @@
+#!/bin/bash
+
+# chainali_1b is as chainali_1a except it has 3 more cnn layers and 1 less tdnn layer.
+# ./local/chain/compare_wer.sh exp/chain/cnn_chainali_1a/ exp/chain/cnn_chainali_1b/
+# System                      cnn_chainali_1a cnn_chainali_1b
+# WER                              6.69     6.25
+# Final train prob              -0.0132   -0.0041
+# Final valid prob              -0.0509   -0.0337
+# Final train prob (xent)       -0.6393   -0.6287
+# Final valid prob (xent)       -1.0116   -0.9064
+
+# steps/info/chain_dir_info.pl exp/chain/chainali_cnn_1b/
+# exp/chain/chainali_cnn_1b/: num-iters=21 nj=2..4 num-params=4.0M dim=40->364 combine=-0.009->-0.005 xent:train/valid[13,20,final]=(-1.47,-0.728,-0.623/-1.69,-1.02,-0.940) logprob:train/valid[13,20,final]=(-0.068,-0.030,-0.011/-0.086,-0.056,-0.038)
+
+# cat exp/chain/cnn_chainali_1b/decode_test/scoring_kaldi/best_*
+# %WER 3.94 [ 2600 / 65921, 415 ins, 1285 del, 900 sub ] exp/chain/cnn_chainali_1b/decode_test/cer_10_0.0
+# %WER 6.25 [ 1158 / 18542, 103 ins, 469 del, 586 sub ] exp/chain/cnn_chainali_1b/decode_test/wer_12_0.0
+
+set -e -o pipefail
+
+stage=0
+
+nj=30
+train_set=train
+gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
+                # should have alignments for the specified training data.
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1b  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+ali=tri3_ali
+chain_model_dir=exp/chain${nnet3_affix}/cnn${affix}
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=4
+alignment_subsampling_factor=1
+# training chunk-options
+chunk_width=340,300,200,100
+num_leaves=500
+# 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=false
+lang_test=lang_test
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${ali}
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_lats_chain
+gmm_lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_lats
+dir=exp/chain${nnet3_affix}/cnn_chainali${affix}
+train_data_dir=data/${train_set}
+tree_dir=exp/chain${nnet3_affix}/tree_chain
+
+# 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/lang_chain
+for f in $train_data_dir/feats.scp \
+    $train_data_dir/feats.scp $gmm_dir/final.mdl \
+    $ali_dir/ali.1.gz $gmm_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"
+  # 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 [ $lang/L.fst -nt data/$lang_test/L.fst ]; then
+      echo "$0: $lang already exists, not overwriting it; continuing"
+    else
+      echo "$0: $lang already exists and seems to be older than data/lang..."
+      echo " ... not sure what to do.  Exiting."
+      exit 1;
+    fi
+  else
+    cp -r data/$lang_test $lang
+    silphonelist=$(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
+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" \
+                            --scale-opts '--transition-scale=1.0 --self-loop-scale=1.0' \
+                            ${train_data_dir} data/$lang_test $chain_model_dir $lat_dir
+  cp $gmm_lat_dir/splice_opts $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."
+     exit 1;
+  fi
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor $frame_subsampling_factor \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$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)
+  common1="required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common3="required-time-offsets= height-offsets=-1,0,1 num-filters-out=70"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=40 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn4 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn5 height-in=20 height-out=10 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn6 height-in=10 height-out=10 time-offsets=-1,0,1 $common3
+  conv-relu-batchnorm-layer name=cnn7 height-in=10 height-out=10 time-offsets=-1,0,1 $common3
+  relu-batchnorm-layer name=tdnn1 input=Append(-4,-2,0,2,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # 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=$tdnn_dim target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=$alignment_subsampling_factor \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=1000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=4 \
+    --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=64,32 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$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 $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/$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/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/iam/v1/local/make_features.py b/egs/iam/v1/local/make_features.py
new file mode 100755
index 00000000000..b998464953f
--- /dev/null
+++ b/egs/iam/v1/local/make_features.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+
+""" This script converts images to Kaldi-format feature matrices. The input to
+    this script is the path to a data directory, e.g. "data/train". This script
+    reads the images listed in images.scp and writes them to standard output
+    (by default) as Kaldi-formatted matrices (in text form). It also scales the
+    images so they have the same height (via --feat-dim). It can optionally pad
+    the images (on left/right sides) with white pixels.
+
+    eg. local/make_features.py data/train --feat-dim 40
+"""
+
+import argparse
+import os
+import sys
+import numpy as np
+from scipy import misc
+
+parser = argparse.ArgumentParser(description="""Converts images (in 'dir'/images.scp) to features and
+                                                writes them to standard output in text format.""")
+parser.add_argument('dir', type=str,
+                    help='Source data directory (containing images.scp)')
+parser.add_argument('--out-ark', type=str, default='-',
+                    help='Where to write the output feature file')
+parser.add_argument('--feat-dim', type=int, default=40,
+                    help='Size to scale the height of all images')
+parser.add_argument('--padding', type=int, default=5,
+                    help='Number of white pixels to pad on the left'
+                    'and right side of the image.')
+args = parser.parse_args()
+
+
+def write_kaldi_matrix(file_handle, matrix, key):
+    file_handle.write(key + " [ ")
+    num_rows = len(matrix)
+    if num_rows == 0:
+        raise Exception("Matrix is empty")
+    num_cols = len(matrix[0])
+
+    for row_index in range(len(matrix)):
+        if num_cols != len(matrix[row_index]):
+            raise Exception("All the rows of a matrix are expected to "
+                            "have the same length")
+        file_handle.write(" ".join(map(lambda x: str(x), matrix[row_index])))
+        if row_index != num_rows - 1:
+            file_handle.write("\n")
+    file_handle.write(" ]\n")
+
+def get_scaled_image(im):
+    scale_size = args.feat_dim
+    sx = im.shape[1]
+    sy = im.shape[0]
+    scale = (1.0 * scale_size) / sy
+    nx = int(scale_size)
+    ny = int(scale * sx)
+    im = misc.imresize(im, (nx, ny))
+    padding_x = args.padding
+    padding_y = im.shape[0]
+    im_pad = np.concatenate((255 * np.ones((padding_y, padding_x),
+                                           dtype=int), im), axis=1)
+    im_pad1 = np.concatenate((im_pad, 255 * np.ones((padding_y, padding_x),
+                                                    dtype=int)), axis=1)
+    return im_pad1
+
+### main ###
+data_list_path = os.path.join(args.dir,'images.scp')
+
+if args.out_ark == '-':
+    out_fh = sys.stdout
+else:
+    out_fh = open(args.out_ark,'wb')
+
+with open(data_list_path) as f:
+    for line in f:
+        line = line.strip()
+        line_vect = line.split(' ')
+        image_id = line_vect[0]
+        image_path = line_vect[1]
+        im = misc.imread(image_path)
+        im_scale = get_scaled_image(im)
+
+        data = np.transpose(im_scale, (1, 0))
+        data = np.divide(data, 255.0)
+        write_kaldi_matrix(out_fh, data, image_id)
diff --git a/egs/iam/v1/local/prepare_data.sh b/egs/iam/v1/local/prepare_data.sh
new file mode 100755
index 00000000000..1350c5841df
--- /dev/null
+++ b/egs/iam/v1/local/prepare_data.sh
@@ -0,0 +1,149 @@
+#!/bin/bash
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+#                2017  Hossein Hadian
+# Apache 2.0
+
+# This script downloads the IAM handwriting database and prepares the training
+# and test data (i.e text, images.scp, utt2spk and spk2utt) by calling process_data.py.
+# It also downloads the LOB and Brown text corpora. It downloads the database files
+# only if they do not already exist in download directory.
+
+#  Eg. local/prepare_data.sh
+#  Eg. text file: 000_a01-000u-00 A MOVE to stop Mr. Gaitskell from
+#      utt2spk file: 000_a01-000u-00 000
+#      images.scp file: 000_a01-000u-00 data/local/lines/a01/a01-000u/a01-000u-00.png
+#      spk2utt file: 000 000_a01-000u-00 000_a01-000u-01 000_a01-000u-02 000_a01-000u-03
+
+stage=0
+download_dir=data/download
+username=
+password=       # username and password for downloading the IAM database
+                # if you have not already downloaded the database, please
+                # register at http://www.fki.inf.unibe.ch/databases/iam-handwriting-database
+                # and provide this script with your username and password.
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1;
+
+if [[ ! -f $download_dir/lines.tgz && -z $username ]]; then
+  echo "$0: Warning: Couldn't find lines.tgz in $download_dir. Unless the extracted dataset files"
+  echo "exist in your data/local directory this script will fail because the required files"
+  echo "can't be downloaded automatically (it needs registration)."
+  echo "Please register at http://www.fki.inf.unibe.ch/databases/iam-handwriting-database"
+  echo "... and then call this script again with --username <username> --password <password>"
+  echo ""
+fi
+
+lines=data/local/lines
+xml=data/local/xml
+ascii=data/local/ascii
+bcorpus=data/local/browncorpus
+lobcorpus=data/local/lobcorpus
+data_split_info=data/local/largeWriterIndependentTextLineRecognitionTask
+lines_url=http://www.fki.inf.unibe.ch/DBs/iamDB/data/lines/lines.tgz
+xml_url=http://www.fki.inf.unibe.ch/DBs/iamDB/data/xml/xml.tgz
+data_split_info_url=http://www.fki.inf.unibe.ch/DBs/iamDB/tasks/largeWriterIndependentTextLineRecognitionTask.zip
+ascii_url=http://www.fki.inf.unibe.ch/DBs/iamDB/data/ascii/ascii.tgz
+brown_corpus_url=http://www.sls.hawaii.edu/bley-vroman/brown.txt
+lob_corpus_url=http://ota.ox.ac.uk/text/0167.zip
+mkdir -p $download_dir data/local
+
+# download and extact images and transcription
+if [ -d $lines ]; then
+  echo "$0: Not downloading lines images as it is already there."
+else
+  if [ ! -f $download_dir/lines.tgz ]; then
+    echo "$0: Trying to download lines images..."
+    wget -P $download_dir --user "$username" --password "$password" $lines_url || exit 1;
+  fi
+  mkdir -p $lines
+  tar -xzf $download_dir/lines.tgz -C $lines || exit 1;
+  echo "$0: Done downloading and extracting lines images"
+fi
+
+if [ -d $xml ]; then
+  echo "$0: Not downloading transcriptions as it is already there."
+else
+  if [ ! -f $download_dir/xml.tgz ]; then
+    echo "$0: Trying to download transcriptions..."
+    wget -P $download_dir --user "$username" --password "$password" $xml_url || exit 1;
+  fi
+  mkdir -p $xml
+  tar -xzf $download_dir/xml.tgz -C $xml || exit 1;
+  echo "$0: Done downloading and extracting transcriptions."
+fi
+
+if [ -d $data_split_info ]; then
+  echo "$0: Not downloading data split information as it is already there."
+else
+  if [ ! -f $download_dir/largeWriterIndependentTextLineRecognitionTask.zip ]; then
+    echo "$0: Trying to download training and testing data split information..."
+    wget -P $download_dir --user "$username" --password "$password" $data_split_info_url || exit 1;
+  fi
+  mkdir -p $data_split_info
+  unzip $download_dir/largeWriterIndependentTextLineRecognitionTask.zip -d $data_split_info || exit 1;
+  echo "$0: Done downloading and extracting training and testing data split information"
+fi
+
+if [ -d $ascii ]; then
+  echo "$0: Not downloading ascii.tgz as it is already there."
+else
+  if [ ! -f $download_dir/ascii.tgz ]; then
+    echo "$0: trying to download ascii.tgz..."
+    wget -P $download_dir --user "$username" --password "$password" $ascii_url || exit 1;
+  fi
+  mkdir -p $ascii
+  tar -xzf $download_dir/ascii.tgz -C $ascii || exit 1;
+  echo "$0: Done downloading and extracting ascii.tgz"
+fi
+
+if [ -d $lobcorpus ]; then
+  echo "$0: Not downloading the LOB text corpus as it is already there."
+else
+  if [ ! -f $lobcorpus/0167.zip ]; then
+    echo "$0: Downloading the LOB text corpus ..."
+    mkdir -p $lobcorpus
+    wget -P $lobcorpus/ $lob_corpus_url || exit 1;
+  fi
+  unzip $lobcorpus/0167.zip -d $lobcorpus || exit 1;
+  echo "$0: Done downloading and extracting LOB corpus"
+fi
+
+if [ -d $bcorpus ]; then
+  echo "$0: Not downloading the Brown corpus as it is already there."
+else
+  if [ ! -f $bcorpus/brown.txt ]; then
+    mkdir -p $bcorpus
+    echo "$0: Downloading the Brown text corpus..."
+    wget -P $bcorpus $brown_corpus_url || exit 1;
+  fi
+  echo "$0: Done downloading the Brown text corpus"
+fi
+
+mkdir -p data/{train,test,val}
+file_name=largeWriterIndependentTextLineRecognitionTask
+
+train_old="data/local/$file_name/trainset.txt"
+test_old="data/local/$file_name/testset.txt"
+val1_old="data/local/$file_name/validationset1.txt"
+val2_old="data/local/$file_name/validationset2.txt"
+
+train_new="data/local/train.uttlist"
+test_new="data/local/test.uttlist"
+val_new="data/local/validation.uttlist"
+
+cat $train_old > $train_new
+cat $test_old > $test_new
+cat $val1_old $val2_old > $val_new
+
+if [ $stage -le 0 ]; then
+  local/process_data.py data/local data/train --dataset train || exit 1
+  local/process_data.py data/local data/test --dataset test || exit 1
+  local/process_data.py data/local data/val --dataset validation || exit 1
+
+  utils/utt2spk_to_spk2utt.pl data/train/utt2spk > data/train/spk2utt
+  utils/utt2spk_to_spk2utt.pl data/test/utt2spk > data/test/spk2utt
+fi
diff --git a/egs/iam/v1/local/prepare_dict.sh b/egs/iam/v1/local/prepare_dict.sh
new file mode 100755
index 00000000000..77a46df384f
--- /dev/null
+++ b/egs/iam/v1/local/prepare_dict.sh
@@ -0,0 +1,49 @@
+#!/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
+mkdir -p $dir
+
+# First get the set of all letters that occur in data/train/text
+cat data/train/text | \
+  perl -ne '@A = split; shift @A; for(@A) {print join("\n", split(//)), "\n";}' | \
+  sort -u > $dir/nonsilence_phones.txt
+
+# Now list all the unique words (that use only the above letters)
+# in data/train/text and LOB+Brown corpora with their comprising
+# letters as their transcription. (Letter # is replaced with <HASH>)
+
+export letters=$(cat $dir/nonsilence_phones.txt | tr -d "\n")
+
+cut -d' ' -f2- data/train/text | \
+  cat data/local/lobcorpus/0167/download/LOB_COCOA/lob.txt \
+      data/local/browncorpus/brown.txt - | \
+  perl -e '$letters=$ENV{letters};
+while(<>){ @A = split;
+  foreach(@A) {
+    if(! $seen{$_} && $_ =~ m/^[$letters]+$/){
+      $seen{$_} = 1;
+      $trans = join(" ", split(//));
+      $trans =~ s/#/<HASH>/g;
+      print "$_ $trans\n";
+    }
+  }
+}' | sort > $dir/lexicon.txt
+
+
+sed -i "s/#/<HASH>/" $dir/nonsilence_phones.txt
+
+echo '<sil> SIL' >> $dir/lexicon.txt
+echo '<unk> 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/iam/v1/local/process_data.py b/egs/iam/v1/local/process_data.py
new file mode 100755
index 00000000000..fa5eb484707
--- /dev/null
+++ b/egs/iam/v1/local/process_data.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+
+""" This script reads the extracted IAM database files and creates
+    the following files (for the data subset selected via --dataset):
+    text, utt2spk, images.scp.
+
+  Eg. local/process_data.py data/local data/train data --dataset train
+  Eg. text file: 000_a01-000u-00 A MOVE to stop Mr. Gaitskell from
+      utt2spk file: 000_a01-000u-00 000
+      images.scp file: 000_a01-000u-00 data/local/lines/a01/a01-000u/a01-000u-00.png
+"""
+
+import argparse
+import os
+import sys
+import xml.dom.minidom as minidom
+
+parser = argparse.ArgumentParser(description="""Creates text, utt2spk
+                                                and images.scp files.""")
+parser.add_argument('database_path', type=str,
+                    help='Path to the downloaded (and extracted) IAM data')
+parser.add_argument('out_dir', type=str,
+                    help='Where to write output files.')
+parser.add_argument('--dataset', type=str, default='train',
+                    choices=['train', 'test','validation'],
+                    help='Subset of data to process.')
+args = parser.parse_args()
+
+text_file = os.path.join(args.out_dir + '/', 'text')
+text_fh = open(text_file, 'w')
+
+utt2spk_file = os.path.join(args.out_dir + '/', 'utt2spk')
+utt2spk_fh = open(utt2spk_file, 'w')
+
+image_file = os.path.join(args.out_dir + '/', 'images.scp')
+image_fh = open(image_file, 'w')
+
+dataset_path = os.path.join(args.database_path,
+                            args.dataset + '.uttlist')
+
+text_file_path = os.path.join(args.database_path,
+                              'ascii','lines.txt')
+text_dict = {}
+def process_text_file_for_word_model():
+  with open (text_file_path, 'rt') as in_file:
+    for line in in_file:
+      if line[0]=='#':
+        continue
+      line = line.strip()
+      utt_id = line.split(' ')[0]
+      text_vect = line.split(' ')[8:]
+      text = "".join(text_vect)
+      text = text.replace("|", " ")
+      text_dict[utt_id] = text
+
+print("Processing '{}' data...".format(args.dataset))
+process_text_file_for_word_model()
+
+with open(dataset_path) as f:
+  for line in f:
+    line = line.strip()
+    line_vect = line.split('-')
+    xml_file = line_vect[0] + '-' + line_vect[1]
+    xml_path = os.path.join(args.database_path, 'xml', xml_file + '.xml')
+    img_num = line[-3:]
+    doc = minidom.parse(xml_path)
+
+    form_elements = doc.getElementsByTagName('form')[0]
+    writer_id = form_elements.getAttribute('writer-id')
+    outerfolder = form_elements.getAttribute('id')[0:3]
+    innerfolder = form_elements.getAttribute('id')
+    lines_path = os.path.join(args.database_path, 'lines',
+                              outerfolder, innerfolder, innerfolder)
+    image_file_path = lines_path + img_num + '.png'
+    text =  text_dict[line]
+    utt_id = writer_id + '_' + line
+    text_fh.write(utt_id + ' ' + text + '\n')
+    utt2spk_fh.write(utt_id + ' ' + writer_id + '\n')
+    image_fh.write(utt_id + ' ' + image_file_path + '\n')
diff --git a/egs/iam/v1/local/score.sh b/egs/iam/v1/local/score.sh
new file mode 100755
index 00000000000..31564d25326
--- /dev/null
+++ b/egs/iam/v1/local/score.sh
@@ -0,0 +1,5 @@
+#!/bin/bash
+
+
+steps/scoring/score_kaldi_wer.sh "$@"
+steps/scoring/score_kaldi_cer.sh --stage 2 "$@"
diff --git a/egs/iam/v1/local/train_lm.sh b/egs/iam/v1/local/train_lm.sh
new file mode 100755
index 00000000000..aa4303d6a28
--- /dev/null
+++ b/egs/iam/v1/local/train_lm.sh
@@ -0,0 +1,139 @@
+#!/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 an LM on the LOB+Brown text data and IAM 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
+
+echo "$0 $@"  # Print the command line for logging
+. ./utils/parse_options.sh || exit 1;
+
+dir=data/local/local_lm
+lm_dir=${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
+
+  # Using LOB and brown corpus.
+  cat data/local/lobcorpus/0167/download/LOB_COCOA/lob.txt > ${dir}/data/text/text.txt
+  cat data/local/browncorpus/brown.txt >> ${dir}/data/text/text.txt
+
+  # use the validation data as the dev set.
+  # Note: the name 'dev' is treated specially by pocolm, it automatically
+  # becomes the dev set.
+
+  cat data/val/text | cut -d " " -f 2-  > ${dir}/data/text/dev.txt
+
+  # use the training data as an additional data source.
+  # we can later fold the dev data into this.
+  cat data/train/text | cut -d " " -f 2- >  ${dir}/data/text/iam.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/test/text  > ${dir}/data/real_dev_set.txt
+
+  # get the wordlist from IAM text
+  cat ${dir}/data/text/{iam,text}.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | 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 "$0: training the unpruned LM"
+  min_counts='train=2 iam=1'
+  wordlist=${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=10 --warm-start-ratio=20  \
+               --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'
+  #log-prob: -5.05603614242 [perplexity = 156.967086371] over 19477.0 words
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: pruning the LM (to larger size)"
+  # Using 1 million n-grams for a big LM for rescoring purposes.
+  size=1000000
+  prune_lm_dir.py --target-num-ngrams=$size --initial-threshold=0.02 ${unpruned_lm_dir} ${dir}/data/lm_${order}_prune_big
+
+  get_data_prob.py ${dir}/data/real_dev_set.txt ${dir}/data/lm_${order}_prune_big 2>&1 | grep -F '[perplexity'
+  # get_data_prob.py: log-prob of data/local/local_lm/data/real_dev_set.txt given model data/local/local_lm/data/lm_3_prune_big was -5.06654404785 per word [perplexity = 158.625177948] over 19477.0 words
+  # current results, after adding --limit-unk-history=true:
+
+
+  mkdir -p ${dir}/data/arpa
+  format_arpa_lm.py ${dir}/data/lm_${order}_prune_big | gzip -c > ${dir}/data/arpa/${order}gram_big.arpa.gz
+fi
+
+if [ $stage -le 3 ]; then
+  echo "$0: pruning the LM (to smaller size)"
+  # Using 500,000 n-grams for a smaller LM for graph building.  Prune from the
+  # bigger-pruned LM, it'll be faster.
+  size=500000
+  prune_lm_dir.py --target-num-ngrams=$size ${dir}/data/lm_${order}_prune_big ${dir}/data/lm_${order}_prune_small
+
+  get_data_prob.py ${dir}/data/real_dev_set.txt ${dir}/data/lm_${order}_prune_small 2>&1 | grep -F '[perplexity'
+  # get_data_prob.py: log-prob of data/local/local_lm/data/real_dev_set.txt given model data/local/local_lm/data/lm_3_prune_small was -5.24719139498 per word [perplexity = 190.031793995] over 19477.0 words
+  # current results, after adding --limit-unk-history=true (needed for modeling OOVs and not blowing up LG.fst):
+
+
+  format_arpa_lm.py ${dir}/data/lm_${order}_prune_small | gzip -c > ${dir}/data/arpa/${order}gram_small.arpa.gz
+fi
diff --git a/egs/iam/v1/local/wer_output_filter b/egs/iam/v1/local/wer_output_filter
new file mode 100755
index 00000000000..162482539ed
--- /dev/null
+++ b/egs/iam/v1/local/wer_output_filter
@@ -0,0 +1,27 @@
+#!/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 . "
+
+import sys
+import re
+
+punctuations = "!(),.?;:'-\""
+escaped_punctuations = re.escape(punctuations)
+
+for line in sys.stdin:
+  words = line.strip().split()
+  uttid = words[0]
+  transcript = ' '.join(words[1:])
+  split_transcript = " ".join(re.split("([{}])".format(escaped_punctuations),
+                                       transcript)).strip()
+  print("{} {}".format(uttid, split_transcript))
diff --git a/egs/iam/v1/path.sh b/egs/iam/v1/path.sh
new file mode 100755
index 00000000000..2d17b17a84a
--- /dev/null
+++ b/egs/iam/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/iam/v1/run.sh b/egs/iam/v1/run.sh
new file mode 100755
index 00000000000..c8ebb9ae649
--- /dev/null
+++ b/egs/iam/v1/run.sh
@@ -0,0 +1,122 @@
+#!/bin/bash
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+#                2017  Hossein Hadian
+
+set -e
+stage=0
+nj=20
+
+# iam_database points to the database path on the JHU grid. If you have not
+# already downloaded the database you can set it to a local directory
+# like "data/download" and follow the instructions
+# in "local/prepare_data.sh" to download the database:
+iam_database=/export/corpora5/handwriting_ocr/IAM
+
+. ./cmd.sh ## You'll want to change cmd.sh to something that will work on your system.
+           ## This relates to the queue.
+. ./path.sh
+. ./utils/parse_options.sh  # e.g. this parses the above options
+                            # if supplied.
+
+if [ $stage -le 0 ]; then
+  echo "$0: Preparing data..."
+  local/prepare_data.sh --download-dir "$iam_database"
+fi
+mkdir -p data/{train,test}/data
+
+if [ $stage -le 1 ]; then
+  echo "$0: Preparing the test and train feature files..."
+  for dataset in train test; do
+    local/make_features.py data/$dataset --feat-dim 40 | \
+      copy-feats --compress=true --compression-method=7 \
+                 ark:- ark,scp:data/$dataset/data/images.ark,data/$dataset/feats.scp
+    steps/compute_cmvn_stats.sh data/$dataset
+  done
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: Preparing dictionary and lang..."
+  local/prepare_dict.sh
+  utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 8 --sil-prob 0.95 \
+                        data/local/dict "<unk>" data/lang/temp data/lang
+fi
+
+if [ $stage -le 3 ]; then
+  echo "$0: Estimating a language model for decoding..."
+  local/train_lm.sh
+  utils/format_lm.sh data/lang data/local/local_lm/data/arpa/3gram_big.arpa.gz \
+                     data/local/dict/lexicon.txt data/lang_test
+fi
+
+if [ $stage -le 4 ]; then
+  steps/train_mono.sh --nj $nj --cmd $cmd --totgauss 10000 data/train \
+    data/lang exp/mono
+fi
+
+if [ $stage -le 5 ]; then
+  utils/mkgraph.sh --mono data/lang_test exp/mono exp/mono/graph
+
+  steps/decode.sh --nj $nj --cmd $cmd exp/mono/graph data/test \
+    exp/mono/decode_test
+fi
+
+if [ $stage -le 6 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd data/train data/lang \
+    exp/mono exp/mono_ali
+
+  steps/train_deltas.sh --cmd $cmd 500 20000 data/train data/lang \
+    exp/mono_ali exp/tri
+fi
+
+if [ $stage -le 7 ]; then
+  utils/mkgraph.sh data/lang_test exp/tri exp/tri/graph
+
+  steps/decode.sh --nj $nj --cmd $cmd exp/tri/graph data/test \
+    exp/tri/decode_test
+fi
+
+if [ $stage -le 8 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd data/train data/lang \
+    exp/tri exp/tri_ali
+
+  steps/train_lda_mllt.sh --cmd $cmd \
+    --splice-opts "--left-context=3 --right-context=3" 500 20000 \
+    data/train data/lang exp/tri_ali exp/tri2
+fi
+
+if [ $stage -le 9 ]; then
+  utils/mkgraph.sh data/lang_test exp/tri2 exp/tri2/graph
+
+  steps/decode.sh --nj $nj --cmd $cmd exp/tri2/graph \
+    data/test exp/tri2/decode_test
+fi
+
+if [ $stage -le 10 ]; then
+  steps/align_fmllr.sh --nj $nj --cmd $cmd --use-graphs true \
+    data/train data/lang exp/tri2 exp/tri2_ali
+
+  steps/train_sat.sh --cmd $cmd 500 20000 \
+    data/train data/lang exp/tri2_ali exp/tri3
+fi
+
+if [ $stage -le 11 ]; then
+  utils/mkgraph.sh data/lang_test exp/tri3 exp/tri3/graph
+
+  steps/decode_fmllr.sh --nj $nj --cmd $cmd exp/tri3/graph \
+    data/test exp/tri3/decode_test
+fi
+
+if [ $stage -le 12 ]; then
+  steps/align_fmllr.sh --nj $nj --cmd $cmd --use-graphs true \
+    data/train data/lang exp/tri3 exp/tri3_ali
+fi
+
+if [ $stage -le 13 ]; then
+  local/chain/run_cnn_1a.sh
+fi
+
+if [ $stage -le 14 ]; then
+  local/chain/run_cnn_chainali_1b.sh --chain-model-dir exp/chain/cnn_1a --stage 2
+fi
diff --git a/egs/iam/v1/steps b/egs/iam/v1/steps
new file mode 120000
index 00000000000..6e99bf5b5ad
--- /dev/null
+++ b/egs/iam/v1/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps
\ No newline at end of file
diff --git a/egs/iam/v1/utils b/egs/iam/v1/utils
new file mode 120000
index 00000000000..b240885218f
--- /dev/null
+++ b/egs/iam/v1/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils
\ No newline at end of file
diff --git a/egs/uw3/README.txt b/egs/uw3/README.txt
new file mode 100644
index 00000000000..b02d00ff541
--- /dev/null
+++ b/egs/uw3/README.txt
@@ -0,0 +1,4 @@
+
+This directory contains example scripts for optical character recognition
+(i.e. OCR) on the UW3 dataset (it's a printed English OCR corpus):
+http://isis-data.science.uva.nl/events/dlia//datasets/uwash3.html
diff --git a/egs/uw3/v1/cmd.sh b/egs/uw3/v1/cmd.sh
new file mode 100644
index 00000000000..3c8eb9f93a5
--- /dev/null
+++ b/egs/uw3/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/uw3/v1/image b/egs/uw3/v1/image
new file mode 120000
index 00000000000..1668ee99922
--- /dev/null
+++ b/egs/uw3/v1/image
@@ -0,0 +1 @@
+../../cifar/v1/image/
\ No newline at end of file
diff --git a/egs/uw3/v1/local/chain/compare_wer.sh b/egs/uw3/v1/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..1a40523355a
--- /dev/null
+++ b/egs/uw3/v1/local/chain/compare_wer.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/cnn{1a,1b}
+
+# ./local/chain/compare_wer.sh exp/chain/cnn1a
+# System                          cnn1a
+# WER                              0.61
+# CER                              0.15
+# Final train prob              -0.0377
+# Final valid prob              -0.0380
+# Final train prob (xent)       -0.0830
+# Final valid prob (xent)       -0.0838
+
+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" $prob
+done
+echo
diff --git a/egs/uw3/v1/local/chain/run_cnn_1a.sh b/egs/uw3/v1/local/chain/run_cnn_1a.sh
new file mode 100755
index 00000000000..ad7421e1261
--- /dev/null
+++ b/egs/uw3/v1/local/chain/run_cnn_1a.sh
@@ -0,0 +1,234 @@
+#!/bin/bash
+
+# Copyright    2017 Hossein Hadian
+#              2017 Chun Chieh Chang
+#              2017 Ashish Arora
+
+# steps/info/chain_dir_info.pl exp/chain/cnn1a/
+# exp/chain/cnn1a/: num-iters=153 nj=3..10 num-params=3.6M dim=40->268 combine=-0.034->-0.034 xent:train/valid[101,152,final]=(-0.097,-0.186,-0.092/-0.101,-0.212,-0.098) logprob:train/valid[101,152,final]=(-0.035,-0.067,-0.035/-0.036,-0.082,-0.035)
+
+# cat exp/chain/cnn1a/decode_test/scoring_kaldi/best_*
+# %WER 0.19 [ 366 / 188135, 110 ins, 123 del, 133 sub ] exp/chain/cnn1a/decode_test/cer_7_0.5
+# %WER 1.00 [ 357 / 35571, 104 ins, 26 del, 227 sub ] exp/chain/cnn1a/decode_test/wer_5_1.0
+
+
+set -e -o pipefail
+
+stage=0
+nj=30
+
+# affix for exp dirs, e.g. it was _cleaned in tedlium.
+nnet3_affix=
+
+affix=1a
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=5
+
+# training chunk-options
+chunk_width=340,300,200,100
+
+# we don't need extra left/right context for TDNN systems.
+chunk_left_context=0
+chunk_right_context=0
+
+# training options
+srand=0
+remove_egs=false
+
+gmm_dir=exp/tri2
+ali_dir=exp/tri2_ali
+lat_dir=exp/chain${nnet3_affix}/tri2_train_lats
+dir=exp/chain${nnet3_affix}/cnn${affix}
+train_data_dir=data/train
+lores_train_data_dir=$train_data_dir  # for the start, use the same data for gmm and chain
+gmm_lang=data/lang
+lang_test=data/lang_unk
+tree_dir=exp/chain${nnet3_affix}/tree${affix}
+
+# 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/lang_chain
+
+# 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
+
+for f in $train_data_dir/feats.scp \
+    $lores_train_data_dir/feats.scp $gmm_dir/final.mdl \
+    $ali_dir/ali.1.gz $gmm_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"
+  # 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 $lang_test/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 $lang_test $lang
+    silphonelist=$(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
+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/align_fmllr_lats.sh --nj $nj --cmd "$cmd" ${lores_train_data_dir} \
+    $lang_test $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+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: $tree_dir/final.mdl already exists, refusing to overwrite it."
+     exit 1;
+  fi
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor 3 \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$cmd" 300 ${lores_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)
+  common1="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=12"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=40 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-2,-1,0,1,2 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=40 time-offsets=-2,-1,0,1,2 $common1
+  relu-batchnorm-layer name=tdnn1 input=Append(-2,-1,0,1,2) dim=450
+  relu-batchnorm-layer name=tdnn2 input=Append(-5,0,5) dim=450
+  relu-batchnorm-layer name=tdnn3 input=Append(-5,0,5) dim=450
+  relu-batchnorm-layer name=tdnn4 input=Append(-5,0,5) dim=450
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=450 target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # 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' models... 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=tdnn4 dim=450 target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=$frame_subsampling_factor \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=1500000 \
+    --trainer.optimization.num-jobs-initial=3 \
+    --trainer.optimization.num-jobs-final=10 \
+    --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=128,64 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$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 $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/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/uw3/v1/local/make_features.py b/egs/uw3/v1/local/make_features.py
new file mode 100755
index 00000000000..dd0a30a19d7
--- /dev/null
+++ b/egs/uw3/v1/local/make_features.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+
+# Copyright      2017  Chun Chieh Chang
+
+""" This script converts images to Kaldi-format feature matrices. The input to
+    this script is the path to a data directory, e.g. "data/train". This script
+    reads the images listed in images.scp and writes them to standard output
+    (by default) as Kaldi-formatted matrices (in text form). It also scales the
+    images so they have the same height (via --feat-dim). It can optionally pad
+    the images (on left/right sides) with white pixels.
+
+    eg. local/make_features.py data/train --feat-dim 40
+"""
+
+import argparse
+import os
+import sys
+import numpy as np
+from scipy import misc
+from scipy import ndimage
+
+from signal import signal, SIGPIPE, SIG_DFL
+signal(SIGPIPE,SIG_DFL)
+
+parser = argparse.ArgumentParser(description="""Converts images (in 'dir'/images.scp) to features and
+                                                writes them to standard output in text format.""")
+parser.add_argument('dir', type=str, help='data directory (should contain images.scp)')
+parser.add_argument('--out-ark', type=str, default='-', help='where to write the output feature file.')
+parser.add_argument('--feat-dim', type=int, default=40,
+                    help='size to scale the height of all images (i.e. the dimension of the resulting features)')
+parser.add_argument('--pad', type=bool, default=False, help='pad the left and right of the images with 10 white pixels.')
+
+args = parser.parse_args()
+
+
+def write_kaldi_matrix(file_handle, matrix, key):
+    file_handle.write(key + " [ ")
+    num_rows = len(matrix)
+    if num_rows == 0:
+        raise Exception("Matrix is empty")
+    num_cols = len(matrix[0])
+    for row_index in range(len(matrix)):
+        if num_cols != len(matrix[row_index]):
+            raise Exception("All the rows of a matrix are expected to "
+                            "have the same length")
+        file_handle.write(" ".join(map(lambda x: str(x), matrix[row_index])))
+        if row_index != num_rows - 1:
+            file_handle.write("\n")
+    file_handle.write(" ]\n")
+
+def get_scaled_image(im):
+    scale_size = args.feat_dim
+    sx = im.shape[1]
+    sy = im.shape[0]
+    # Some Images are rotated
+    if sy > sx:
+        im = np.rot90(im, k = -1)
+        sx = im.shape[1]
+        sy = im.shape[0]
+
+    scale = (1.0 * scale_size) / sy
+    nx = int(scale_size)
+    ny = int(scale * sx)
+    im = misc.imresize(im, (nx, ny))
+
+    noise = np.random.normal(2, 1,(nx, ny))
+    im = im - noise
+
+    return im
+
+### main ###
+data_list_path = os.path.join(args.dir,'images.scp')
+
+if args.out_ark == '-':
+    out_fh = sys.stdout
+else:
+    out_fh = open(args.out_ark,'wb')
+
+with open(data_list_path) as f:
+    for line in f:
+        line = line.strip()
+        line_vect = line.split(' ')
+        image_id = line_vect[0]
+        image_path = line_vect[1]
+
+        im = misc.imread(image_path, flatten = True)
+        im_scale = get_scaled_image(im)
+
+        if args.pad:
+            pad = np.ones((args.feat_dim, 10)) * 255
+            im_data = np.hstack((pad, im_scale, pad))
+        else:
+            im_data = im_scale
+
+        data = np.transpose(im_data, (1, 0))
+        data = np.divide(data, 255.0)
+        write_kaldi_matrix(out_fh, data, image_id)
diff --git a/egs/uw3/v1/local/prepare_data.sh b/egs/uw3/v1/local/prepare_data.sh
new file mode 100755
index 00000000000..47f62e4335a
--- /dev/null
+++ b/egs/uw3/v1/local/prepare_data.sh
@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# Copyright 2017 Chun Chieh Chang
+
+# This script downloads the UW3 dataset (if not already downloaded)
+# and prepares the "train" and "test" data subsets.
+
+set -e
+download_dir=data/download
+
+. ./cmd.sh
+if [ -f path.sh ]; then . ./path.sh; fi
+. parse_options.sh || exit 1;
+
+# Download dir
+download_url=http://www.tmbdev.net/ocrdata/uw3-lines-book.tgz
+data_dir=data/local/extracted_corpus
+
+mkdir -p $download_dir
+mkdir -p $data_dir
+
+if [ -d $data_dir/book ]; then
+  echo "$0: Not downloading dataset as it is already downloaded."
+else
+  if [ ! -f $download_dir/uw3-lines-book.tgz ]; then
+    echo "$0: Downloading dataset..."
+    wget -P $download_dir $download_url || exit 1;
+  fi
+  echo "$0: Extracting..."
+  tar -xzf $download_dir/uw3-lines-book.tgz -C $data_dir/ || exit 1;
+  echo "$0: Done downloading/extracting the datset."
+fi
+
+mkdir -p data/train
+mkdir -p data/test
+echo "$0: Preparing the test and train subsets..."
+local/process_data.py $data_dir/book data || exit 1
+
+utils/utt2spk_to_spk2utt.pl data/train/utt2spk > data/train/spk2utt
+utils/utt2spk_to_spk2utt.pl data/test/utt2spk > data/test/spk2utt
diff --git a/egs/uw3/v1/local/prepare_dict.sh b/egs/uw3/v1/local/prepare_dict.sh
new file mode 100755
index 00000000000..72c9b50e5ec
--- /dev/null
+++ b/egs/uw3/v1/local/prepare_dict.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Hossein Hadian
+
+set -e
+dir=data/local/dict
+
+mkdir -p $dir
+
+cut -d' ' -f2- data/train/text | tr -cs '[a-z][A-Z][0-9][:punct:]' '\n' | sort -u | \
+  awk '{len=split($0,chars,""); printf($0);
+       for (i=0;i<=len;i++) {
+         if(chars[i]=="#") {chars[i]="<HASH>"}
+         printf(chars[i]" ")
+       };
+       printf("\n")};' | \
+  sed 's/.$//' > $dir/lexicon.txt;
+
+cut -d' ' -f2- $dir/lexicon.txt | tr ' ' '\n' | sort -u >$dir/nonsilence_phones.txt
+
+echo '<sil> SIL' >> $dir/lexicon.txt
+echo '<unk> 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/uw3/v1/local/process_data.py b/egs/uw3/v1/local/process_data.py
new file mode 100755
index 00000000000..f5b37b04c2f
--- /dev/null
+++ b/egs/uw3/v1/local/process_data.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+
+# Copyright       2017  Chun Chieh Chang
+
+# This script goes through the downloaded UW3 dataset and creates data files "text",
+# "utt2spk", and "images.scp" for the train and test subsets in data/train and data/test.
+
+# text - matches the transcriptions with the image id
+# utt2spk - matches the image id's with the speaker/writer names
+# images.scp - matches the image is's with the actual image file
+
+import argparse
+import os
+import random
+
+parser = argparse.ArgumentParser(description="""Creates data/train and data/test.""")
+parser.add_argument('database_path', type=str, help='path to downloaded (and extracted) UW3 corpus')
+parser.add_argument('out_dir', type=str, default='data',
+                    help='where to create the train and test data directories')
+args = parser.parse_args()
+
+### main ###
+train_text_file = os.path.join(args.out_dir, 'train', 'text')
+train_text_fh = open(train_text_file, 'w+')
+train_utt2spk_file = os.path.join(args.out_dir, 'train', 'utt2spk')
+train_utt2spk_fh = open(train_utt2spk_file, 'w+')
+train_image_file = os.path.join(args.out_dir, 'train', 'images.scp')
+train_image_fh = open(train_image_file, 'w+')
+
+test_text_file = os.path.join(args.out_dir, 'test', 'text')
+test_text_fh = open(test_text_file, 'w+')
+test_utt2spk_file = os.path.join(args.out_dir, 'test', 'utt2spk')
+test_utt2spk_fh = open(test_utt2spk_file, 'w+')
+test_image_file = os.path.join(args.out_dir, 'test', 'images.scp')
+test_image_fh = open(test_image_file, 'w+')
+
+random.seed(0)
+page_count = 0
+for page in sorted(os.listdir(args.database_path)):
+  page_path = os.path.join(args.database_path, page)
+  page_count = page_count + 1
+  for line in sorted(os.listdir(page_path)):
+    if line.endswith('.txt'):
+      text_path = os.path.join(args.database_path, page, line)
+      image_name = line.split('.')[0]
+      image_path = os.path.join(args.database_path, page, image_name + '.png')
+      utt_id = page + '_' + image_name
+      gt_fh = open(text_path, 'r')
+      text = gt_fh.readlines()[0].strip()
+
+      # The UW3 dataset doesn't have established training and testing splits
+      # The dataset is randomly split train 95% and test 5%
+      coin = random.randint(0, 20)
+      if coin >= 1:
+        train_text_fh.write(utt_id + ' ' + text + '\n')
+        train_utt2spk_fh.write(utt_id + ' ' + str(page_count) + '\n')
+        train_image_fh.write(utt_id + ' ' + image_path + '\n')
+      elif coin < 1:
+        test_text_fh.write(utt_id + ' ' + text + '\n')
+        test_utt2spk_fh.write(utt_id + ' ' + str(page_count) + '\n')
+        test_image_fh.write(utt_id + ' ' + image_path + '\n')
diff --git a/egs/uw3/v1/local/score.sh b/egs/uw3/v1/local/score.sh
new file mode 100755
index 00000000000..9ea4701a833
--- /dev/null
+++ b/egs/uw3/v1/local/score.sh
@@ -0,0 +1,156 @@
+#!/bin/bash
+# Copyright 2012-2014  Johns Hopkins University (Author: Daniel Povey, Yenda Trmal)
+# Apache 2.0
+
+# This script is like steps/scoring/score_kaldi_wer.sh except it transcribes the <unk>'s
+# using local/unk_arc_post_to_transcription.py and also it calls
+# steps/scoring/score_kaldi_cer.sh at the end.
+
+[ -f ./path.sh ] && . ./path.sh
+
+# begin configuration section.
+cmd=run.pl
+stage=0
+decode_mbr=false
+stats=true
+beam=6
+word_ins_penalty=0.0,0.5,1.0
+min_lmwt=3
+max_lmwt=13
+iter=final
+#end configuration section.
+
+echo "$0 $@"  # Print the command line for logging
+[ -f ./path.sh ] && . ./path.sh
+. parse_options.sh || exit 1;
+
+if [ $# -ne 3 ]; then
+  echo "Usage: $0 [--cmd (run.pl|queue.pl...)] <data-dir> <lang-dir|graph-dir> <decode-dir>"
+  echo " Options:"
+  echo "    --cmd (run.pl|queue.pl...)      # specify how to run the sub-processes."
+  echo "    --stage (0|1|2)                 # start scoring script from part-way through."
+  echo "    --decode_mbr (true/false)       # maximum bayes risk decoding (confusion network)."
+  echo "    --min_lmwt <int>                # minumum LM-weight for lattice rescoring "
+  echo "    --max_lmwt <int>                # maximum LM-weight for lattice rescoring "
+  exit 1;
+fi
+
+data=$1
+lang_or_graph=$2
+dir=$3
+model_path=`echo $dir |xargs dirname`
+symtab=$lang_or_graph/words.txt
+
+for f in $symtab $dir/lat.1.gz $data/text; do
+  [ ! -f $f ] && echo "score.sh: no such file $f" && exit 1;
+done
+
+
+ref_filtering_cmd="cat"
+[ -x local/wer_output_filter ] && ref_filtering_cmd="local/wer_output_filter"
+[ -x local/wer_ref_filter ] && ref_filtering_cmd="local/wer_ref_filter"
+hyp_filtering_cmd="cat"
+[ -x local/wer_output_filter ] && hyp_filtering_cmd="local/wer_output_filter"
+[ -x local/wer_hyp_filter ] && hyp_filtering_cmd="local/wer_hyp_filter"
+
+
+if $decode_mbr ; then
+  echo "$0: scoring with MBR, word insertion penalty=$word_ins_penalty"
+else
+  echo "$0: scoring with word insertion penalty=$word_ins_penalty"
+fi
+
+
+mkdir -p $dir/scoring_kaldi
+cat $data/text | $ref_filtering_cmd > $dir/scoring_kaldi/test_filt.txt || exit 1;
+if [ $stage -le 0 ]; then
+
+  for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+    mkdir -p $dir/scoring_kaldi/penalty_$wip/log
+
+    if $decode_mbr ; then
+      $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/best_path.LMWT.log \
+        acwt=\`perl -e \"print 1.0/LMWT\"\`\; \
+        lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \
+        lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \
+        lattice-prune --beam=$beam ark:- ark:- \| \
+        lattice-mbr-decode  --word-symbol-table=$symtab \
+        ark:- ark,t:- \| \
+        utils/int2sym.pl -f 2- $symtab \| \
+        $hyp_filtering_cmd '>' $dir/scoring_kaldi/penalty_$wip/LMWT.txt || exit 1;
+
+    else
+      $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/best_path.LMWT.log \
+        lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \
+        lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \
+        lattice-1best ark:- ark:- \| \
+        lattice-align-words $lang_or_graph/phones/word_boundary.int $model_path/final.mdl ark:- ark:- \| \
+        lattice-arc-post $model_path/final.mdl ark:- - \| \
+        local/unk_arc_post_to_transcription.py $lang_or_graph/phones.txt $lang_or_graph/words.txt data/lang_unk/oov.int \| \
+        $hyp_filtering_cmd '>' $dir/scoring_kaldi/penalty_$wip/LMWT.txt || exit 1;
+    fi
+
+    $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/score.LMWT.log \
+      cat $dir/scoring_kaldi/penalty_$wip/LMWT.txt \| \
+      tr '[:upper:]' '[:lower:]' \| \
+      compute-wer --text --mode=present \
+      "ark:cat $dir/scoring_kaldi/test_filt.txt| tr '[:upper:]' '[:lower:]' |" ark,p:- ">&" $dir/wer_LMWT_$wip || exit 1;
+
+  done
+fi
+
+
+
+if [ $stage -le 1 ]; then
+
+  for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+    for lmwt in $(seq $min_lmwt $max_lmwt); do
+      # adding /dev/null to the command list below forces grep to output the filename
+      grep WER $dir/wer_${lmwt}_${wip} /dev/null
+    done
+  done | utils/best_wer.sh  >& $dir/scoring_kaldi/best_wer || exit 1
+
+  best_wer_file=$(awk '{print $NF}' $dir/scoring_kaldi/best_wer)
+  best_wip=$(echo $best_wer_file | awk -F_ '{print $NF}')
+  best_lmwt=$(echo $best_wer_file | awk -F_ '{N=NF-1; print $N}')
+
+  if [ -z "$best_lmwt" ]; then
+    echo "$0: we could not get the details of the best WER from the file $dir/wer_*.  Probably something went wrong."
+    exit 1;
+  fi
+
+  if $stats; then
+    mkdir -p $dir/scoring_kaldi/wer_details
+    echo $best_lmwt > $dir/scoring_kaldi/wer_details/lmwt # record best language model weight
+    echo $best_wip > $dir/scoring_kaldi/wer_details/wip # record best word insertion penalty
+
+    $cmd $dir/scoring_kaldi/log/stats1.log \
+      cat $dir/scoring_kaldi/penalty_$best_wip/$best_lmwt.txt \| \
+      align-text --special-symbol="'***'" ark:$dir/scoring_kaldi/test_filt.txt ark:- ark,t:- \|  \
+      utils/scoring/wer_per_utt_details.pl --special-symbol "'***'" \| tee $dir/scoring_kaldi/wer_details/per_utt \|\
+       utils/scoring/wer_per_spk_details.pl $data/utt2spk \> $dir/scoring_kaldi/wer_details/per_spk || exit 1;
+
+    $cmd $dir/scoring_kaldi/log/stats2.log \
+      cat $dir/scoring_kaldi/wer_details/per_utt \| \
+      utils/scoring/wer_ops_details.pl --special-symbol "'***'" \| \
+      sort -b -i -k 1,1 -k 4,4rn -k 2,2 -k 3,3 \> $dir/scoring_kaldi/wer_details/ops || exit 1;
+
+    $cmd $dir/scoring_kaldi/log/wer_bootci.log \
+      compute-wer-bootci --mode=present \
+        ark:$dir/scoring_kaldi/test_filt.txt ark:$dir/scoring_kaldi/penalty_$best_wip/$best_lmwt.txt \
+        '>' $dir/scoring_kaldi/wer_details/wer_bootci || exit 1;
+
+  fi
+fi
+
+steps/scoring/score_kaldi_cer.sh --cmd "$cmd" --stage 2 $data $lang_or_graph $dir
+
+# If we got here, the scoring was successful.
+# As a  small aid to prevent confusion, we remove all wer_{?,??} files;
+# these originate from the previous version of the scoring files
+# i keep both statement here because it could lead to confusion about
+# the capabilities of the script (we don't do cer in the script)
+rm $dir/wer_{?,??} 2>/dev/null
+rm $dir/cer_{?,??} 2>/dev/null
+
+exit 0;
diff --git a/egs/uw3/v1/local/train_lm.sh b/egs/uw3/v1/local/train_lm.sh
new file mode 100755
index 00000000000..39eb051d273
--- /dev/null
+++ b/egs/uw3/v1/local/train_lm.sh
@@ -0,0 +1,102 @@
+#!/bin/bash
+
+# Copyright 2016  Vincent Nguyen
+#           2016  Johns Hopkins University (author: Daniel Povey)
+# Apache 2.0
+#
+#
+# This script trains an LM on the UW3 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
+
+echo "$0 $@"  # Print the command line for logging
+. utils/parse_options.sh || exit 1;
+
+dir=data/local/local_lm
+lm_dir=${dir}/data
+
+. ./path.sh
+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;
+
+num_dev_sentences=4500
+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
+
+  head -n $num_dev_sentences < data/train/text | cut -d " " -f 2-  > ${dir}/data/text/dev.txt
+  tail -n +$[$num_dev_sentences+1] < data/train/text | cut -d " " -f 2- >  ${dir}/data/text/uw3.txt
+
+  # for reporting perplexities, we'll use the "real" dev set.
+  # (a subset of the training data is used as ${dir}/data/text/uw3.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/test/text  > ${dir}/data/real_dev_set.txt
+
+  # get wordlist
+  cat ${dir}/data/text/uw3.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | 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 "$0: training the unpruned LM"
+  min_counts='train=2 uw3=1'
+  wordlist=${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=10 --warm-start-ratio=20  \
+               --limit-unk-history=true \
+               --fold-dev-into=uw3 ${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'
+
+  # No need for pruning as the training data is quite small (total # of
+  # n-grams is 685k). Write the arpa:
+  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/uw3/v1/local/unk_arc_post_to_transcription.py b/egs/uw3/v1/local/unk_arc_post_to_transcription.py
new file mode 100755
index 00000000000..c86d35e4b8a
--- /dev/null
+++ b/egs/uw3/v1/local/unk_arc_post_to_transcription.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python
+
+# Copyright     2017  Ashish Arora
+
+import argparse
+import sys
+
+parser = argparse.ArgumentParser(description="""uses phones to convert unk to word""")
+parser.add_argument('phones', type=str, help='phones and phonesID')
+parser.add_argument('words', type=str, help='word and wordID')
+parser.add_argument('unk', type=str, default='-', help='location of unk file')
+parser.add_argument('--input-ark', type=str, default='-', help='where to read the input data')
+parser.add_argument('--out-ark', type=str, default='-', help='where to write the output data')
+args = parser.parse_args()
+### main ###
+phone_fh = open(args.phones, 'r')
+word_fh = open(args.words, 'r')
+unk_fh = open(args.unk,'r')
+if args.input_ark == '-':
+    input_fh = sys.stdin
+else:
+    input_fh = open(args.input_ark,'r')
+if args.out_ark == '-':
+    out_fh = sys.stdout
+else:
+    out_fh = open(args.out_ark,'wb')
+
+phone_dict = dict()# stores phoneID and phone mapping
+phone_data_vect = phone_fh.read().strip().split("\n")
+for key_val in phone_data_vect:
+  key_val = key_val.split(" ")
+  phone_dict[key_val[1]] = key_val[0]
+word_dict = dict()
+word_data_vect = word_fh.read().strip().split("\n")
+for key_val in word_data_vect:
+  key_val = key_val.split(" ")
+  word_dict[key_val[1]] = key_val[0]
+unk_val = unk_fh.read().strip().split(" ")[0]
+
+utt_word_dict = dict()
+utt_phone_dict = dict()# stores utteranceID and phoneID
+unk_word_dict = dict()
+count=0
+for line in input_fh:
+  line_vect = line.strip().split("\t")
+  if len(line_vect) < 6:
+    print "IndexError"
+    print line_vect
+    continue
+  uttID = line_vect[0]
+  word = line_vect[4]
+  phones = line_vect[5]
+  if uttID in utt_word_dict.keys():
+    utt_word_dict[uttID][count] = word
+    utt_phone_dict[uttID][count] = phones
+  else:
+    count = 0
+    utt_word_dict[uttID] = dict()
+    utt_phone_dict[uttID] = dict()
+    utt_word_dict[uttID][count] = word
+    utt_phone_dict[uttID][count] = phones
+  if word == unk_val: # get character sequence for unk
+    phone_key_vect = phones.split(" ")
+    phone_val_vect = list()
+    for pkey in phone_key_vect:
+      phone_val_vect.append(phone_dict[pkey])
+    phone_2_word = list()
+    for phone_val in phone_val_vect:
+      phone_2_word.append(phone_val.split('_')[0])
+    phone_2_word = ''.join(phone_2_word)
+    utt_word_dict[uttID][count] = phone_2_word
+  else:
+    if word == '0':
+      word_val = ' '
+    else:
+      word_val = word_dict[word]
+    utt_word_dict[uttID][count] = word_val
+  count += 1
+
+transcription = ""
+for key in sorted(utt_word_dict.iterkeys()):
+  transcription = key
+  for index in sorted(utt_word_dict[key].iterkeys()):
+    value = utt_word_dict[key][index]
+    transcription = transcription + " " + value
+  out_fh.write(transcription + '\n')
diff --git a/egs/uw3/v1/path.sh b/egs/uw3/v1/path.sh
new file mode 100755
index 00000000000..2d17b17a84a
--- /dev/null
+++ b/egs/uw3/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/uw3/v1/run.sh b/egs/uw3/v1/run.sh
new file mode 100755
index 00000000000..68c51fa4690
--- /dev/null
+++ b/egs/uw3/v1/run.sh
@@ -0,0 +1,106 @@
+#!/bin/bash
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+#                2017  Hossein Hadian
+
+
+set -e
+stage=0
+nj=30
+
+# This is the database path on the JHU grid. You may set this
+# to data/download, in which case the script will automatically download
+# the database:
+uw3_database=/export/a10/corpora5/handwriting_ocr/UW3/
+
+. ./path.sh
+. ./cmd.sh ## You'll want to change cmd.sh to something that will work on your system.
+           ## This relates to the queue.
+. utils/parse_options.sh  # e.g. this parses the --stage option if supplied.
+
+
+if [ $stage -le 0 ]; then
+  # Data preparation
+  local/prepare_data.sh --download-dir "$uw3_database"
+fi
+
+mkdir -p data/{train,test}/data
+if [ $stage -le 1 ]; then
+  echo "$0: Preparing feature files for the test and training data..."
+  for f in train test; do
+    local/make_features.py --feat-dim 40 --pad true data/$f | \
+      copy-feats --compress=true --compression-method=7 \
+      ark:- ark,scp:data/$f/data/images.ark,data/$f/feats.scp || exit 1
+
+    steps/compute_cmvn_stats.sh data/$f || exit 1;
+  done
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: Preparing dictionary and lang..."
+  local/prepare_dict.sh
+  utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 8 \
+    data/local/dict "<unk>" data/lang/temp data/lang
+fi
+
+if [ $stage -le 3 ]; then
+  echo "$0: Estimating a language model for decoding..."
+  local/train_lm.sh
+  utils/format_lm.sh data/lang data/local/local_lm/data/arpa/3gram_unpruned.arpa.gz \
+                     data/local/dict/lexicon.txt data/lang_test
+
+  echo "$0: Preparing the unk model for open-vocab decoding..."
+  utils/lang/make_unk_lm.sh --ngram-order 4 --num-extra-ngrams 7500 \
+                            data/local/dict exp/unk_lang_model
+  utils/prepare_lang.sh --num-sil-states 4 --num-nonsil-states 8 \
+                        --unk-fst exp/unk_lang_model/unk_fst.txt \
+                        data/local/dict "<unk>" data/lang_unk/temp data/lang_unk
+  cp data/lang_test/G.fst data/lang_unk/G.fst
+fi
+
+if [ $stage -le 4 ]; then
+  steps/train_mono.sh --nj $nj --cmd $cmd \
+    data/train data/lang exp/mono
+fi
+
+if [ $stage -le 5 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd \
+    data/train data/lang exp/mono exp/mono_ali
+  steps/train_deltas.sh --cmd $cmd 500 20000 \
+    data/train data/lang exp/mono_ali exp/tri
+fi
+
+if [ $stage -le 6 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd \
+    data/train data/lang exp/tri exp/tri_ali
+  steps/train_lda_mllt.sh --cmd $cmd --splice-opts "--left-context=3 --right-context=3" 500 20000 \
+    data/train data/lang exp/tri_ali exp/tri2
+fi
+
+if [ $stage -le 7 ]; then
+  utils/mkgraph.sh --mono data/lang_test exp/mono exp/mono/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    exp/mono/graph data/test exp/mono/decode_test
+fi
+
+if [ $stage -le 8 ]; then
+  utils/mkgraph.sh data/lang_test exp/tri exp/tri/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    exp/tri/graph data/test exp/tri/decode_test
+fi
+
+if [ $stage -le 9 ]; then
+  utils/mkgraph.sh data/lang_test exp/tri2 exp/tri2/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    exp/tri2/graph data/test exp/tri2/decode_test
+fi
+
+if [ $stage -le 10 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd --use-graphs true \
+    data/train data/lang exp/tri2 exp/tri2_ali
+fi
+
+if [ $stage -le 11 ]; then
+  local/chain/run_cnn_1a.sh
+fi
diff --git a/egs/uw3/v1/steps b/egs/uw3/v1/steps
new file mode 120000
index 00000000000..6e99bf5b5ad
--- /dev/null
+++ b/egs/uw3/v1/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps
\ No newline at end of file
diff --git a/egs/uw3/v1/utils b/egs/uw3/v1/utils
new file mode 120000
index 00000000000..b240885218f
--- /dev/null
+++ b/egs/uw3/v1/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils
\ No newline at end of file