diff --git a/egs/iam/s5/cmd.sh b/egs/iam/s5/cmd.sh
new file mode 100644
index 00000000000..466a51def1b
--- /dev/null
+++ b/egs/iam/s5/cmd.sh
@@ -0,0 +1,15 @@
+# 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"
+export train_cmd="queue.pl"
+export decode_cmd="queue.pl --mem 4G"
diff --git a/egs/iam/s5/image b/egs/iam/s5/image
new file mode 120000
index 00000000000..1668ee99922
--- /dev/null
+++ b/egs/iam/s5/image
@@ -0,0 +1 @@
+../../cifar/v1/image/
\ No newline at end of file
diff --git a/egs/iam/s5/local/augment_and_make_feature_vect.py b/egs/iam/s5/local/augment_and_make_feature_vect.py
new file mode 100755
index 00000000000..b1c179d71ed
--- /dev/null
+++ b/egs/iam/s5/local/augment_and_make_feature_vect.py
@@ -0,0 +1,288 @@
+#!/usr/bin/env python
+import random
+import argparse
+import os
+import sys
+import scipy.io as sio
+import numpy as np
+from scipy import misc
+from scipy.ndimage.interpolation import affine_transform
+import math
+from signal import signal, SIGPIPE, SIG_DFL
+signal(SIGPIPE, SIG_DFL)
+
+parser = argparse.ArgumentParser(
+    description="""Generates and saves the feature vectors""")
+parser.add_argument(
+    'dir', type=str, help='directory of images.scp and is also output directory')
+parser.add_argument('--seg', type=str, default='1',
+                    help='JOB number of images.JOB.scp if run in parallel mode')
+parser.add_argument('--out-ark', type=str, default='-',
+                    help='where to write the output feature file')
+parser.add_argument('--scale-size', type=int, default=40,
+                    help='size to scale the height of all images')
+parser.add_argument('--padding', type=int, default=5,
+                    help='size to scale the height of all images')
+parser.add_argument('--vertical-shift', type=int, default=10,
+                    help='total number of padding pixel per column')
+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.scale_size
+    sx = im.shape[1]  # width
+    sy = im.shape[0]  # height
+    scale = (1.0 * scale_size) / sy
+    nx = int(scale_size)
+    ny = int(scale * sx)
+    im = misc.imresize(im, (nx, ny))
+    padding_x = max(5, int((args.padding / 100) * im.shape[1]))
+    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
+
+
+def contrast_normalization(im, low_pct, high_pct):
+    element_number = im.size
+    rows = im.shape[0]
+    cols = im.shape[1]
+    im_contrast = np.zeros(shape=im.shape)
+    low_index = int(low_pct * element_number)
+    high_index = int(high_pct * element_number)
+    sorted_im = np.sort(im, axis=None)
+    low_thred = sorted_im[low_index]
+    high_thred = sorted_im[high_index]
+    for i in range(rows):
+        for j in range(cols):
+            if im[i, j] > high_thred:
+                im_contrast[i, j] = 255  # lightest to white
+            elif im[i, j] < low_thred:
+                im_contrast[i, j] = 0  # darkest to black
+            else:
+                # linear normalization
+                im_contrast[i, j] = (im[i, j] - low_thred) * \
+                    255 / (high_thred - low_thred)
+    return im_contrast
+
+
+def geometric_moment(frame, p, q):
+    m = 0
+    for i in range(frame.shape[1]):
+        for j in range(frame.shape[0]):
+            m += (i ** p) * (j ** q) * frame[i][i]
+    return m
+
+
+def central_moment(frame, p, q):
+    u = 0
+    x_bar = geometric_moment(frame, 1, 0) / \
+        geometric_moment(frame, 0, 0)  # m10/m00
+    y_bar = geometric_moment(frame, 0, 1) / \
+        geometric_moment(frame, 0, 0)  # m01/m00
+    for i in range(frame.shape[1]):
+        for j in range(frame.shape[0]):
+            u += ((i - x_bar)**p) * ((j - y_bar)**q) * frame[i][j]
+    return u
+
+
+def height_normalization(frame, w, h):
+    frame_normalized = np.zeros(shape=(h, w))
+    alpha = 4
+    x_bar = geometric_moment(frame, 1, 0) / \
+        geometric_moment(frame, 0, 0)  # m10/m00
+    y_bar = geometric_moment(frame, 0, 1) / \
+        geometric_moment(frame, 0, 0)  # m01/m00
+    sigma_x = (alpha * ((central_moment(frame, 2, 0) /
+                         geometric_moment(frame, 0, 0)) ** .5))  # alpha * sqrt(u20/m00)
+    sigma_y = (alpha * ((central_moment(frame, 0, 2) /
+                         geometric_moment(frame, 0, 0)) ** .5))  # alpha * sqrt(u02/m00)
+    for x in range(w):
+        for y in range(h):
+            i = int((x / w - 0.5) * sigma_x + x_bar)
+            j = int((y / h - 0.5) * sigma_y + y_bar)
+            frame_normalized[x][y] = frame[i][j]
+    return frame_normalized
+
+
+def find_slant(im):
+    rows = im.shape[0]
+    cols = im.shape[1]
+    sum_max = 0
+    slant_degree = 0
+    for shear_degree in range(-45, 45, 5):
+        sum = 0
+        shear_rad = shear_degree / 360.0 * 2 * math.pi
+        shear_matrix = np.array([[1, 0],
+                                 [np.tan(shear_rad), 1]])
+        sheared_im = affine_transform(im, shear_matrix, cval=255.0)
+        for j in range(cols):
+            foreground = (sheared_im[:, j] < 100)
+            number = np.sum(foreground)
+            # print(number)
+            if number != 0:
+                start_point = -1
+                end_point = -1
+                start_point = 0
+                for i in range(rows):
+                    if foreground[i] == 1:
+                        start_point = i
+                        break
+                for i in range(rows - 1, -1, -1):
+                    if foreground[i] == 1:
+                        end_point = i
+                        break
+                length = end_point - start_point + 1
+                #print(number, length)
+                if length == number:
+                    sum = sum + number * number
+        #print(shear_degree, sum)
+        if sum > sum_max:
+            sum_max = sum
+            slant_degree = shear_degree
+    return slant_degree
+
+
+def deslant(im, shear):
+    padding_x = int(abs(np.tan(shear)) * im.shape[0])
+    padding_y = im.shape[0]
+    if shear > 0:
+        im_pad = np.concatenate(
+            (255 * np.ones((padding_y, padding_x), dtype=int), im), axis=1)
+    else:
+        im_pad = np.concatenate(
+            (im, 255 * np.ones((padding_y, padding_x), dtype=int)), axis=1)
+
+    shear_matrix = np.array([[1, 0],
+                             [np.tan(shear), 1]])
+    # sheared_im = affine_transform(image, shear_matrix, output_shape=(
+    # im.shape[0], im.shape[1] + abs(int(im.shape[0] * np.tan(shear)))), cval=128.0)
+    sheared_im = affine_transform(im_pad, shear_matrix, cval=255.0)
+    return sheared_im
+
+
+def vertical_shift(im, mode='mid'):
+    total = args.vertical_shift
+    if mode == 'mid':
+        top = total / 2
+        bottom = total - top
+    elif mode == 'top':  # more padding on top
+        top = random.randint(total / 2, total)
+        bottom = total - top
+    elif mode == 'bottom':  # more padding on bottom
+        top = random.randint(0, total / 2)
+        bottom = total - top
+    width = im.shape[1]
+    im_pad = np.concatenate(
+        (255 * np.ones((top, width), dtype=int), im), axis=0)
+    im_pad = np.concatenate(
+        (im_pad, 255 * np.ones((bottom, width), dtype=int)), axis=0)
+    return im_pad
+
+
+def image_augment(im, out_fh, image_id):
+    random.seed(1)
+    shift_setting = ['mid', 'top', 'bottom']
+    image_shift_id = []
+    for i in range(3):
+        image_shift_id.append(image_id + '_shift' + str(i + 1))
+        im_shift = vertical_shift(im, shift_setting[i])
+        im_scaled = get_scaled_image(im_shift)
+        data = np.transpose(im_scaled, (1, 0))
+        data = np.divide(data, 255.0)
+        new_scp_list.append(image_id + '_shift' + str(i + 1))
+        write_kaldi_matrix(out_fh, data, image_shift_id[i])
+
+
+# main #
+new_scp_list = list()
+text_file = os.path.join(args.dir, 'backup', 'text')
+text_dict = dict()  # stores imageID and text
+
+with open(text_file) as text_fh:
+    for uttID_text in text_fh:
+        uttID_text = uttID_text.strip()
+        uttID_text_vect = uttID_text.split(" ")
+        uttID = uttID_text_vect[0]
+        imageID = uttID.split("_")[1]
+        text_vect = uttID_text_vect[1:]
+        text = " ".join(text_vect)
+        text_dict[imageID] = text
+
+utt2spk_file = os.path.join(args.dir, 'backup', 'utt2spk')
+uttID_spk_dict = dict()  # stores imageID and speaker
+
+with open(utt2spk_file) as utt2spk_fh:
+    for uttID_spk in utt2spk_fh:
+        uttID_spk = uttID_spk.strip()
+        uttID_spk_vect = uttID_spk.split(" ")
+        uttID = uttID_spk_vect[0]
+        imageID = uttID.split("_")[1]
+        spk = uttID_spk_vect[1]
+        uttID_spk_dict[imageID] = spk
+
+image_file = os.path.join(args.dir, 'backup', 'images.scp')
+uttID_path_dict = dict()  # stores imageID and image path
+
+with open(image_file) as image_fh:
+    for uttID_path in image_fh:
+        uttID_path = uttID_path.strip()
+        uttID_path_vect = uttID_path.split(" ")
+        uttID = uttID_path_vect[0]
+        imageID = uttID.split("_")[1]
+        path = uttID_path_vect[1]
+        uttID_path_dict[imageID] = path
+
+scp_name = 'images.scp'
+data_list_path = os.path.join(args.dir, 'backup', scp_name)
+
+if args.out_ark == '-':
+    out_fh = sys.stdout
+else:
+    out_fh = open(args.out_ark, 'wb')
+
+text_file = os.path.join(args.dir, 'text')
+text_fh = open(text_file, 'w+')
+
+utt2spk_file = os.path.join(args.dir, 'utt2spk')
+utt2spk_fh = open(utt2spk_file, 'w+')
+
+image_file = os.path.join(args.dir, 'images.scp')
+image_fh = open(image_file, 'w+')
+
+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_contrast = contrast_normalization(im, 0.05, 0.2)
+        #shear = (find_slant(im_contrast) / 360.0) * 2 * math.pi
+        im_scaled = get_scaled_image(im)
+        image_augment(im_scaled, out_fh, image_id)
+
+for uttID in new_scp_list:
+    imageID = uttID.split("_")[1]
+    text_fh.write(uttID + ' ' + text_dict[imageID] + '\n')
+    utt2spk_fh.write(uttID + ' ' + uttID_spk_dict[imageID] + '\n')
+    image_fh.write(uttID + ' ' + uttID_path_dict[imageID] + '\n')
diff --git a/egs/iam/s5/local/chain/align_nnet3_lats.sh b/egs/iam/s5/local/chain/align_nnet3_lats.sh
new file mode 100755
index 00000000000..cd81000ac37
--- /dev/null
+++ b/egs/iam/s5/local/chain/align_nnet3_lats.sh
@@ -0,0 +1,105 @@
+#!/bin/bash
+#
+# Copyright 2012-2015  Johns Hopkins University (Author: Daniel Povey)
+#           2017       Hossein Hadian
+# Apache 2.0
+
+# Begin configuration section.
+stage=0
+nj=4
+cmd=run.pl
+# Begin configuration.
+scale_opts="--transition-scale=1.0 --self-loop-scale=1.0"
+acoustic_scale=1.0
+beam=10
+final_beam=20  # For the lattice-generation phase there is no retry-beam.  This
+               # is a limitation of gmm-latgen-faster.  We just use an
+               # intermediate beam.  We'll lose a little data and it will be
+               # slightly slower.  (however, the min-active of 200 that
+               # gmm-latgen-faster defaults to may help.)
+post_decode_acwt=12.0  # can be used in 'chain' systems to scale acoustics by 10 so the
+                      # regular scoring script works.
+frames_per_chunk=50
+extra_left_context=0
+extra_right_context=0
+extra_left_context_initial=-1
+extra_right_context_final=-1
+
+# End configuration options.
+
+echo "$0 $@"  # Print the command line for logging
+
+[ -f path.sh ] && . ./path.sh # source the path.
+. parse_options.sh || exit 1;
+
+if [ $# != 4 ]; then
+   echo "usage: steps/align_nnet3_lats.sh <data-dir> <lang-dir> <src-dir> <align-dir>"
+   echo "e.g.:  steps/align_nnet3_lats.sh data/train data/lang exp/tri1 exp/tri1_lats"
+   echo "main options (for others, see top of script file)"
+   echo "  --config <config-file>                           # config containing options"
+   echo "  --nj <nj>                                        # number of parallel jobs"
+   echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
+   exit 1;
+fi
+
+data=$1
+lang=$2
+srcdir=$3
+dir=$4
+
+oov=`cat $lang/oov.int` || exit 1;
+silphonelist=`cat $lang/phones/silence.csl` || exit 1;
+sdata=$data/split$nj
+
+mkdir -p $dir/log
+echo $nj > $dir/num_jobs
+[[ -d $sdata && $data/feats.scp -ot $sdata ]] || split_data.sh $data $nj || exit 1;
+
+utils/lang/check_phones_compatible.sh $lang/phones.txt $srcdir/phones.txt || exit 1;
+cp $lang/phones.txt $dir || exit 1;
+
+cp $srcdir/{tree,final.mdl} $dir || exit 1;
+cmvn_opts=`cat $srcdir/cmvn_opts 2>/dev/null`
+cp $srcdir/cmvn_opts $dir 2>/dev/null # cmn/cmvn option.
+
+feats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- |"
+
+tra="ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt $sdata/JOB/text|";
+frame_subsampling_opt=
+if [ -f $srcdir/frame_subsampling_factor ]; then
+  # e.g. for 'chain' systems
+  frame_subsampling_factor=$(cat $srcdir/frame_subsampling_factor)
+  frame_subsampling_opt="--frame-subsampling-factor=$frame_subsampling_factor"
+  cp $srcdir/frame_subsampling_factor $dir
+fi
+
+model=$srcdir/final.mdl
+if [ "$post_decode_acwt" == 1.0 ]; then
+  lat_wspecifier="ark:|gzip -c >$dir/lat.JOB.gz"
+else
+  lat_wspecifier="ark:|lattice-scale --acoustic-scale=$post_decode_acwt ark:- ark:- | gzip -c >$dir/lat.JOB.gz"
+fi
+
+if [ $stage -le 1 ]; then
+  echo "$0: generating lattices containing alternate pronunciations."
+  $cmd JOB=1:$nj $dir/log/generate_lattices.JOB.log \
+    compile-train-graphs --read-disambig-syms=$lang/phones/disambig.int \
+      $scale_opts $dir/tree $dir/final.mdl  $lang/L.fst "$tra" ark:- \| \
+      nnet3-latgen-faster $frame_subsampling_opt \
+        --frames-per-chunk=$frames_per_chunk \
+        --extra-left-context=$extra_left_context \
+        --extra-right-context=$extra_right_context \
+        --extra-left-context-initial=$extra_left_context_initial \
+        --extra-right-context-final=$extra_right_context_final \
+        --minimize=false --acoustic-scale=$acoustic_scale --beam=$final_beam \
+        --lattice-beam=$final_beam --allow-partial=false \
+        --word-determinize=false --word-symbol-table=$lang/words.txt \
+        $model ark:- "$feats" \
+        "$lat_wspecifier" || exit 1;
+fi
+
+echo "$0: done generating lattices from training transcripts."
+
+utils/summarize_warnings.pl $dir/log
+
+exit 0;
diff --git a/egs/iam/s5/local/chain/compare_wer.sh b/egs/iam/s5/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..8cd85ecdf01
--- /dev/null
+++ b/egs/iam/s5/local/chain/compare_wer.sh
@@ -0,0 +1,56 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/cnn{1a,1b}
+
+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/s5/local/chain/run_cnn_1a.sh b/egs/iam/s5/local/chain/run_cnn_1a.sh
new file mode 100755
index 00000000000..02df03d1cd6
--- /dev/null
+++ b/egs/iam/s5/local/chain/run_cnn_1a.sh
@@ -0,0 +1,231 @@
+#!/bin/bash
+
+# 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->380 combine=-0.033->-0.025 xent:train/valid[13,20,final]=(-1.07,-1.31,-0.560/-1.30,-1.70,-0.978) logprob:train/valid[13,20,final]=(-0.064,-0.119,-0.011/-0.115,-0.208,-0.096)
+
+# head exp/chain/cnn_1a/decode_test/scoring_kaldi/best_wer
+#%WER 18.34 [ 3231 / 17616, 348 ins, 693 del, 2190 sub ] exp/chain/cnn_1a/decode_test/wer_8_1.0
+
+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=32
+chunk_right_context=32
+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/${train_set}
+lores_train_data_dir=$train_data_dir  # for the start, use the same data for gmm and chain
+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 \
+    $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 [ $lang/L.fst -nt data/$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 $nj --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/$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 $frame_subsampling_factor \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$train_cmd" $num_leaves ${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=36"
+  common2="required-time-offsets=0 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="$decode_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 32 \
+    --extra-right-context-final 32 \
+    --frames-per-chunk $frames_per_chunk \
+    --nj $nj --cmd "$decode_cmd" \
+    $dir/graph data/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/iam/s5/local/chain/run_cnn_chainali_1a.sh b/egs/iam/s5/local/chain/run_cnn_chainali_1a.sh
new file mode 100755
index 00000000000..0f0296d7af6
--- /dev/null
+++ b/egs/iam/s5/local/chain/run_cnn_chainali_1a.sh
@@ -0,0 +1,236 @@
+#!/bin/bash
+
+# chainali_1a uses chain model for lattice instead of gmm-hmm model. It has more cnn layers as compared to 1a
+# (18.34% -> 13.68%)
+
+# steps/info/chain_dir_info.pl exp/chain/cnn1a_chainali/
+# exp/chain/cnn_chainali_1a/: num-iters=21 nj=2..4 num-params=3.8M dim=40->380 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
+
+# head exp/chain/cnn_chainali_1a/decode_test/scoring_kaldi/best_wer
+# %WER 13.68 [ 2410 / 17616, 243 ins, 633 del, 1534 sub ] exp/chain/cnn_chainali_1a/decode_test/wer_8_1.0
+
+set -e -o pipefail
+
+stage=0
+
+nj=30
+train_set=train
+gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
+                # should have alignments for the specified training data.
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+ali=tri3_ali
+chain_model_dir=exp/chain${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=32
+chunk_right_context=32
+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}
+lores_train_data_dir=$train_data_dir  # for the start, use the same data for gmm and chain
+tree_dir=exp/chain${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 \
+    $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 [ $lang/L.fst -nt data/$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
+  local/chain/align_nnet3_lats.sh --nj $nj --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/$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 $frame_subsampling_factor \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$train_cmd" $num_leaves ${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=36"
+  common2="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common3="required-time-offsets=0 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="$decode_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 32 \
+    --extra-right-context-final 32 \
+    --frames-per-chunk $frames_per_chunk \
+    --nj $nj --cmd "$decode_cmd" \
+    $dir/graph data/test $dir/decode_test || exit 1;
+fi
diff --git a/egs/iam/s5/local/iam_train_lm.sh b/egs/iam/s5/local/iam_train_lm.sh
new file mode 100755
index 00000000000..e7ba2295b13
--- /dev/null
+++ b/egs/iam/s5/local/iam_train_lm.sh
@@ -0,0 +1,140 @@
+#!/bin/bash
+
+# Copyright 2016  Vincent Nguyen
+#           2016  Johns Hopkins University (author: Daniel Povey)
+# Apache 2.0
+#
+# This script trains a LM on the Cantab-Tedlium text data and tedlium acoustic training data.
+# It is based on the example scripts distributed with PocoLM
+
+# It will first check if pocolm is installed and if not will process with installation
+# It will then get the source data from the pre-downloaded Cantab-Tedlium files
+# and the pre-prepared data/train text source.
+
+
+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=6161
+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.
+  head -86858 data/download/lobcorpus/0167/download/LOB_COCOA/lob.txt > ${dir}/data/text/text.txt
+  cat data/download/browncorpus/brown.txt >> ${dir}/data/text/text.txt
+  #cat test_words.txt  >> ${dir}/data/text/text.txt
+  # use a subset of the annotated training 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
+  # .. and the rest of the training data as an additional data source.
+  # we can later fold the dev data into this.
+  head -n $[$num_dev_sentences] < data/train/text | cut -d " " -f 2- >  ${dir}/data/text/ted.txt
+
+  # for reporting perplexities, we'll use the "real" dev set.
+  # (a subset of the training data is used as ${dir}/data/text/ted.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 ${dir}/data/text/text.txt | tr '[:space:]' '[\n*]' | grep -v "^\s*$" | 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 "$0: training the unpruned LM"
+  min_counts='train=2 ted=1'
+  wordlist=${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=ted ${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/s5/local/make_feature_vect.py b/egs/iam/s5/local/make_feature_vect.py
new file mode 100755
index 00000000000..dd35f1b14c7
--- /dev/null
+++ b/egs/iam/s5/local/make_feature_vect.py
@@ -0,0 +1,71 @@
+#!/usr/bin/env python
+
+import argparse
+import os
+import sys
+import scipy.io as sio
+import numpy as np
+from scipy import misc
+
+from signal import signal, SIGPIPE, SIG_DFL
+signal(SIGPIPE,SIG_DFL)
+
+parser = argparse.ArgumentParser(description="""Generates and saves the feature vectors""")
+parser.add_argument('dir', type=str, help='directory of images.scp and is also output directory')
+parser.add_argument('--out-ark', type=str, default='-', help='where to write the output feature file')
+parser.add_argument('--scale-size', type=int, default=40, help='size to scale the height of all images')
+parser.add_argument('--padding', type=int, default=5, help='size to scale the height of all images')
+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.scale_size
+    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 = max(5,int((args.padding/100)*im.shape[1]))
+    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/s5/local/prepare_data.sh b/egs/iam/s5/local/prepare_data.sh
new file mode 100755
index 00000000000..2c79e67cc89
--- /dev/null
+++ b/egs/iam/s5/local/prepare_data.sh
@@ -0,0 +1,135 @@
+#!/bin/bash
+# This script loads the IAM handwritten dataset
+
+stage=0
+nj=20
+dir=data
+
+. ./cmd.sh
+if [ -f path.sh ]; then . ./path.sh; fi
+. parse_options.sh || exit 1;
+
+#download dir
+add_val_data_train=false
+dl_dir=data/download
+lines=$dl_dir/lines
+xml=$dl_dir/xml
+ascii=$dl_dir/ascii
+bcorpus=$dl_dir/browncorpus
+lobcorpus=$dl_dir/lobcorpus
+data_split_info=$dl_dir/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 $dl_dir
+#download and extact images and transcription
+if [ -d $lines ]; then
+  echo Not downloading lines images as it is already there.
+else
+  if [ ! -f $dl_dir/lines.tgz ]; then
+    echo Downloading lines images...
+    wget -P $dl_dir --user userjh --password password $lines_url || exit 1;
+  fi
+  mkdir -p $lines
+  tar -xvzf $dl_dir/lines.tgz -C $lines || exit 1;
+  echo Done downloading and extracting lines images
+fi
+
+if [ -d $xml ]; then
+  echo Not downloading transcription as it is already there.
+else
+  if [ ! -f $dl_dir/xml.tgz ]; then
+    echo Downloading transcription ...
+    wget -P $dl_dir --user userjh --password password $xml_url || exit 1;
+  fi
+  mkdir -p $xml
+  tar -xvzf $dl_dir/xml.tgz -C $xml || exit 1;
+  echo Done downloading and extracting transcription
+fi
+
+if [ -d $data_split_info ]; then
+  echo Not downloading data split, training and testing split, information as it is already there.
+else
+  if [ ! -f $dl_dir/largeWriterIndependentTextLineRecognitionTask.zip ]; then
+    echo Downloading training and testing data Split Information ...
+    wget -P $dl_dir --user userjh --password password $data_split_info_url || exit 1;
+  fi
+  mkdir -p $data_split_info
+  unzip $dl_dir/largeWriterIndependentTextLineRecognitionTask.zip -d $data_split_info || exit 1;
+  echo Done downloading and extracting training and testing data Split Information
+fi
+
+if [ -d $ascii ]; then
+  echo Not downloading ascii folder as it is already there.
+else
+  if [ ! -f $dl_dir/ascii.tgz ]; then
+    echo Downloading ascii folder ...
+    wget -P $dl_dir --user userjh --password password $ascii_url || exit 1;
+  fi
+  mkdir -p $ascii
+  tar -xvzf $dl_dir/ascii.tgz -C $ascii || exit 1;
+  echo Done downloading and extracting ascii folder
+fi
+
+if [ -d $lobcorpus ]; then
+  echo Not downloading lob corpus as it is already there.
+else
+  if [ ! -f $dl_dir/0167.zip ]; then
+    echo Downloading lob corpus ...
+    wget -P $dl_dir $lob_corpus_url || exit 1;
+  fi
+  mkdir -p $lobcorpus
+  unzip $dl_dir/0167.zip -d $lobcorpus || exit 1;
+  echo Done downloading and extracting lob corpus
+fi
+
+if [ -d $bcorpus ]; then
+  echo Not downloading brown corpus as it is already there.
+else
+  if [ ! -f $bcorpus/brown.txt ]; then
+    mkdir -p $bcorpus
+    echo Downloading brown corpus ...
+    wget -P $bcorpus $brown_corpus_url || exit 1;
+  fi
+  echo Done downloading brown corpus
+fi
+
+mkdir -p $dir/{train,test,val}
+file_name=largeWriterIndependentTextLineRecognitionTask
+testset=testset.txt
+trainset=trainset.txt
+val1=validationset1.txt
+val2=validationset2.txt
+test_path="$dl_dir/$file_name/$testset"
+train_path="$dl_dir/$file_name/$trainset"
+val1_path="$dl_dir/$file_name/$val1"
+val2_path="$dl_dir/$file_name/$val2"
+
+new_train_set=new_trainset.txt
+new_test_set=new_testset.txt
+new_val_set=new_valset.txt
+new_train_path="$dir/$new_train_set"
+new_test_path="$dir/$new_test_set"
+new_val_path="$dir/$new_val_set"
+
+if [ $add_val_data_train = true ]; then
+ cat $train_path $val1_path $val2_path > $new_train_path
+ cat $test_path > $new_test_path
+ cat $val1_path $val2_path > $new_val_path
+else
+ cat $train_path > $new_train_path
+ cat $test_path > $new_test_path
+ cat $val1_path $val2_path > $new_val_path
+fi
+
+if [ $stage -le 0 ]; then
+  local/process_data.py $dl_dir $dir/train $dir --dataset new_trainset --model_type word || exit 1
+  local/process_data.py $dl_dir $dir/test $dir --dataset new_testset --model_type word || exit 1
+  local/process_data.py $dl_dir $dir/val $dir --dataset new_valset --model_type word || exit 1
+
+  utils/utt2spk_to_spk2utt.pl $dir/train/utt2spk > $dir/train/spk2utt
+  utils/utt2spk_to_spk2utt.pl $dir/test/utt2spk > $dir/test/spk2utt
+fi
diff --git a/egs/iam/s5/local/prepare_dict.sh b/egs/iam/s5/local/prepare_dict.sh
new file mode 100755
index 00000000000..5d13bdc5d8b
--- /dev/null
+++ b/egs/iam/s5/local/prepare_dict.sh
@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+train_text=$1
+test_text=$2
+dir=$3
+
+mkdir -p $dir
+
+local/prepare_lexicon.py $train_text $test_text $dir
+
+cut -d' ' -f2- $dir/lexicon.txt | tr ' ' '\n' | sort -u >$dir/nonsilence_phones.txt || exit 1;
+
+( echo '<sil> SIL'; ) >> $dir/lexicon.txt || exit 1;
+( echo '<unk> SIL'; ) >> $dir/lexicon.txt || exit 1;
+
+( echo SIL ) > $dir/silence_phones.txt
+
+echo SIL >$dir/optional_silence.txt
+
+echo -n "" >$dir/extra_questions.txt
diff --git a/egs/iam/s5/local/prepare_lexicon.py b/egs/iam/s5/local/prepare_lexicon.py
new file mode 100755
index 00000000000..86298c45733
--- /dev/null
+++ b/egs/iam/s5/local/prepare_lexicon.py
@@ -0,0 +1,46 @@
+#!/usr/bin/env python
+
+import argparse
+import os
+import sys
+
+parser = argparse.ArgumentParser(description="""Creates the list of characters and words in lexicon""")
+parser.add_argument('database_path', type=str, help='path to train text file')
+parser.add_argument('test_text', type=str, help='path to test text file to include it in lexicon')
+parser.add_argument('dir', type=str, help='output path')
+args = parser.parse_args()
+
+### main ###
+char = {}
+lex = {}
+
+text_path = os.path.join(args.database_path,'text')
+with open(text_path) as f:
+  for line in f:
+    line = line.strip()
+    line_vect = line.split(' ')
+    for i in range(1,len(line_vect)):
+      characters = list(line_vect[i])
+      entry = " ".join(characters)
+      entry = entry.replace("#", "<HASH>")
+      if line_vect[i]:
+        lex[line_vect[i]] = entry
+
+if args.test_text > 1:
+  text_path = os.path.join(args.test_text,'text')
+  with open(text_path) as f:
+    for line in f:
+      line = line.strip()
+      line_vect = line.split(' ')
+      for i in range(1,len(line_vect)):
+        characters = list(line_vect[i])
+        entry = " ".join(characters)
+        entry = entry.replace("#", "<HASH>")
+        if line_vect[i]:
+          lex[line_vect[i]] = entry
+
+
+lex_file = os.path.join(args.dir, 'lexicon.txt')
+lex_fh = open(lex_file, 'w+')
+for key in sorted(lex):
+  lex_fh.write(key + " " + lex[key] + "\n")
diff --git a/egs/iam/s5/local/prepare_lm.sh b/egs/iam/s5/local/prepare_lm.sh
new file mode 100755
index 00000000000..59178b881fc
--- /dev/null
+++ b/egs/iam/s5/local/prepare_lm.sh
@@ -0,0 +1,62 @@
+#!/bin/bash
+
+# Copyright 2015     Hossein Hadian
+# Apache 2.0
+
+
+ngram=2
+
+[ -f ./path.sh ] && . ./path.sh  # source the path.
+. parse_options.sh || exit 1
+
+# This is only called for decoding purposes. This is not necessary for training.
+if [ $# != 2 ]; then
+  echo "Usage: $0 [--ngram <n-gram-order>] <lm-src> <out-lang-dir>"
+  echo "<lm-src> can either be an ARPA file or transcriptions file, e.g. train/text"
+  exit 1;
+fi
+
+lmsrc=$1
+lang=$2
+
+tmpdir=$lang/temp
+mkdir -p $tmpdir
+
+if [[ $lmsrc == *.arpa ]]; then
+  echo "$0: Using arpa LM: "$lmsrc
+  arpa=$lmsrc
+else
+  echo "$0: No Arpa file provided..Creating a $ngram-gram LM from provided text: "$lmsrc;
+  if [ -z $IRSTLM ]; then
+    export IRSTLM=$PWD/../../../tools/irstlm
+    export PATH=$PWD/../../../tools/irstlm/bin:$PATH
+    echo "$0: Assuming irstlm is at: "$IRSTLM
+  fi
+  ! which build-lm.sh 2>/dev/null && \
+    echo "$0: IRSTLM does not seem to be installed (build-lm.sh not on your path): " && \
+    echo "go to <kaldi-root>/tools and try 'make irstlm_tgt'" && exit 1;
+
+  cut -d' ' -f2- $lmsrc | sed -e 's:^:<s> :' -e 's:$: </s>:' \
+                              > $tmpdir/tmp_lm_train
+  build-lm.sh -k 1 -i $tmpdir/tmp_lm_train -n $ngram -o $tmpdir/tmp_lm.ilm.gz
+
+  compile-lm $tmpdir/tmp_lm.ilm.gz -t=yes /dev/stdout | \
+    grep -v unk > $tmpdir/lm_phone_bg.arpa
+
+  arpa=$tmpdir/lm_phone_bg.arpa
+fi
+
+cat $arpa | utils/find_arpa_oovs.pl $lang/words.txt  > $tmpdir/oovs.txt
+
+cat $arpa | \
+    grep -v '<s> <s>' | \
+    grep -v '</s> <s>' | \
+    grep -v '</s> </s>' | \
+    arpa2fst - | fstprint | \
+    utils/remove_oovs.pl $tmpdir/oovs.txt | \
+    utils/eps2disambig.pl | utils/s2eps.pl | \
+    fstcompile --isymbols=$lang/words.txt --osymbols=$lang/words.txt \
+               --keep_isymbols=false --keep_osymbols=false | \
+    fstrmepsilon | fstarcsort --sort_type=ilabel >$lang/G.fst || exit 1;
+
+echo "$0: Done preparing the LM"
diff --git a/egs/iam/s5/local/process_data.py b/egs/iam/s5/local/process_data.py
new file mode 100755
index 00000000000..f9838d34563
--- /dev/null
+++ b/egs/iam/s5/local/process_data.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python
+
+import argparse
+import os
+import sys
+import numpy as np
+from scipy import misc
+import xml.dom.minidom as minidom
+
+parser = argparse.ArgumentParser(description="""Creates text utt2spk 
+                                                and image file """)
+parser.add_argument('database_path', type=str,
+                    help='path to downloaded iam data')
+parser.add_argument('out_dir', type=str,
+                    help='where to write output files')
+parser.add_argument('dataset_dir', type=str,
+                    help='directory containing dataset')
+parser.add_argument('--dataset', type=str, default='new_trainset',
+                    choices=['new_trainset', 'new_testset','new_valset'],
+                    help='choose new_trainset, testset')
+parser.add_argument('--model_type', type=str,default='word',
+                    choices=['word', 'character'],
+                    help='word model or character model')
+args = parser.parse_args()
+
+### main ###
+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.dataset_dir,
+                            args.dataset + '.txt')
+
+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()
+      line_vect = line.split(' ')
+      text_vect = line.split(' ')[8:]
+      text = "".join(text_vect)
+      text = text.replace("|", " ")
+      text_dict[line_vect[0]] = text
+
+def process_text_file_for_char_model():
+  with open (text_file_path, 'rt') as in_file:
+    for line in in_file:
+      if line[0]=='#':
+        continue
+      line = line.strip()
+      line_vect = line.split(' ')
+      text_vect = line.split(' ')[8:]
+      text = "".join(text_vect)
+      characters = list(text)
+      spaced_characters = " ".join(characters)
+      spaced_characters = spaced_characters.replace("|", "SIL")
+      spaced_characters = "SIL " + spaced_characters
+      spaced_characters = spaced_characters + " SIL"
+      text_dict[line_vect[0]] = spaced_characters
+
+
+if args.model_type=='word':
+  print 'processing word model'
+  process_text_file_for_word_model()
+else:
+  print 'processing char model'
+  process_text_file_for_char_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/s5/local/run_unk_model.sh b/egs/iam/s5/local/run_unk_model.sh
new file mode 100755
index 00000000000..b8ed377370a
--- /dev/null
+++ b/egs/iam/s5/local/run_unk_model.sh
@@ -0,0 +1,42 @@
+#!/bin/bash
+
+
+utils/lang/make_unk_lm.sh --ngram-order 4 --num-extra-ngrams 7500 data/train/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/train/dict "<unk>" data/lang/temp data/lang_unk
+
+# note: it's important that the LM we built in data/lang/G.fst was created using
+# pocolm with the option --limit-unk-history=true (see ted_train_lm.sh).  This
+# keeps the graph compact after adding the unk model (we only have to add one
+# copy of it).
+
+cp data/lang_test_corpus/G.fst data/lang_unk/G.fst
+#utils/mkgraph.sh data/lang_unk exp/tri3 exp/tri3/graph_unk
+
+. ./cmd.sh
+
+## Caution: if you use this unk-model stuff, be sure that the scoring script
+## does not use lattice-align-words-lexicon, because it's not compatible with
+## the unk-model.  Instead you should use lattice-align-words (of course, this
+## only works if you have position-dependent phones).
+
+#decode_=30
+#for dset in dev test; do
+#    steps/decode_fmllr.sh --nj $decode_nj --cmd "$decode_cmd"  --num-threads 4 \
+#      exp/tri3/graph_unk data/${dset} exp/tri3/decode_${dset}_unk
+#    steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" data/lang data/lang_rescore \
+#       data/${dset} exp/tri3/decode_${dset}_unk exp/tri3/decode_${dset}_unk_rescore
+#done
+#
+#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 $nj --cmd "$decode_cmd" \
+#    $dir/graph data/test $dir/lang_test_corpus_c || exit 1
+#
+# # for x in exp/tri3/decode*; do grep Sum $x/*/*ys | utils/best_wer.sh ; done | grep -v old | grep -v si
diff --git a/egs/iam/s5/local/score.sh b/egs/iam/s5/local/score.sh
new file mode 100755
index 00000000000..518bd6d0459
--- /dev/null
+++ b/egs/iam/s5/local/score.sh
@@ -0,0 +1,154 @@
+#!/bin/bash
+# Copyright 2012-2014  Johns Hopkins University (Author: Daniel Povey, Yenda Trmal)
+# Apache 2.0
+
+# See the script steps/scoring/score_kaldi_cer.sh in case you need to evalutate CER
+
+[ -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=7
+max_lmwt=17
+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/iam/s5/local/unk_arc_post_to_transcription.py b/egs/iam/s5/local/unk_arc_post_to_transcription.py
new file mode 100755
index 00000000000..c27bf226cf9
--- /dev/null
+++ b/egs/iam/s5/local/unk_arc_post_to_transcription.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python
+
+import argparse
+import os
+import sys
+import numpy as np
+from scipy import misc
+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/iam/s5/path.sh b/egs/iam/s5/path.sh
new file mode 100755
index 00000000000..2d17b17a84a
--- /dev/null
+++ b/egs/iam/s5/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/s5/run.sh b/egs/iam/s5/run.sh
new file mode 100755
index 00000000000..d1eeda7e0d1
--- /dev/null
+++ b/egs/iam/s5/run.sh
@@ -0,0 +1,171 @@
+#!/bin/bash
+
+stage=0
+nj=20
+color=1
+data_dir=data
+exp_dir=exp
+augment=false
+. ./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
+  local/prepare_data.sh --nj $nj --dir $data_dir
+fi
+mkdir -p $data_dir/{train,test}/data
+
+if [ $stage -le 1 ]; then
+  local/make_feature_vect.py $data_dir/test --scale-size 40 | \
+    copy-feats --compress=true --compression-method=7 \
+    ark:- ark,scp:$data_dir/test/data/images.ark,$data_dir/test/feats.scp || exit 1
+  steps/compute_cmvn_stats.sh $data_dir/test || exit 1;
+
+  if [ $augment = true ]; then
+    # create a backup directory to store text, utt2spk and image.scp file
+    mkdir -p $data_dir/train/backup
+    mv $data_dir/train/text $data_dir/train/utt2spk $data_dir/train/images.scp $data_dir/train/backup/
+    local/augment_and_make_feature_vect.py $data_dir/train --scale-size 40 --vertical-shift 10 | \
+      copy-feats --compress=true --compression-method=7 \
+      ark:- ark,scp:$data_dir/train/data/images.ark,$data_dir/train/feats.scp || exit 1
+    utils/utt2spk_to_spk2utt.pl $data_dir/train/utt2spk > $data_dir/train/spk2utt
+  else
+    local/make_feature_vect.py $data_dir/train --scale-size 40 | \
+      copy-feats --compress=true --compression-method=7 \
+      ark:- ark,scp:$data_dir/train/data/images.ark,$data_dir/train/feats.scp || exit 1
+  fi
+  steps/compute_cmvn_stats.sh $data_dir/train || exit 1;
+fi
+
+numSilStates=4
+numStates=8
+
+if [ $stage -le 2 ]; then
+  local/prepare_dict.sh $data_dir/train/ $data_dir/test/ $data_dir/train/dict
+  utils/prepare_lang.sh --num-sil-states $numSilStates --num-nonsil-states $numStates \
+    $data_dir/train/dict "<unk>" $data_dir/lang/temp $data_dir/lang
+fi
+
+if [ $stage -le 3 ]; then
+  local/iam_train_lm.sh
+  cp -R $data_dir/lang -T $data_dir/lang_test_corpus
+  gunzip -k -f data/local/local_lm/data/arpa/3gram_big.arpa.gz
+  local/prepare_lm.sh data/local/local_lm/data/arpa/3gram_big.arpa $data_dir/lang_test_corpus || exit 1;
+  local/run_unk_model.sh
+fi
+
+num_gauss=10000
+numLeavesTri=500
+numGaussTri=20000
+
+if [ $stage -le 4 ]; then
+  steps/train_mono.sh --nj $nj --cmd $cmd \
+    --totgauss $num_gauss \
+    $data_dir/train \
+    $data_dir/lang \
+    $exp_dir/mono
+fi
+
+if [ $stage -le 5 ]; then
+  utils/mkgraph.sh --mono $data_dir/lang_test_corpus \
+    $exp_dir/mono \
+    $exp_dir/mono/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    $exp_dir/mono/graph \
+    $data_dir/test \
+    $exp_dir/mono/decode_test
+fi
+
+if [ $stage -le 6 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/mono \
+    $exp_dir/mono_ali
+  steps/train_deltas.sh --cmd $cmd \
+    $numLeavesTri $numGaussTri $data_dir/train $data_dir/lang \
+    $exp_dir/mono_ali \
+    $exp_dir/tri
+fi
+
+if [ $stage -le 7 ]; then
+  utils/mkgraph.sh $data_dir/lang_test_corpus \
+    $exp_dir/tri \
+    $exp_dir/tri/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    $exp_dir/tri/graph \
+    $data_dir/test \
+    $exp_dir/tri/decode_test
+fi
+
+if [ $stage -le 8 ]; then
+  steps/align_si.sh --nj $nj --cmd $cmd \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/tri \
+    $exp_dir/tri_ali
+  steps/train_lda_mllt.sh --cmd $cmd \
+    --splice-opts "--left-context=3 --right-context=3" \
+    $numLeavesTri $numGaussTri \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/tri_ali $exp_dir/tri2
+fi
+
+if [ $stage -le 9 ]; then
+  utils/mkgraph.sh $data_dir/lang_test_corpus \
+    $exp_dir/tri2 \
+    $exp_dir/tri2/graph
+  steps/decode.sh --nj $nj --cmd $cmd \
+    $exp_dir/tri2/graph \
+    $data_dir/test \
+    $exp_dir/tri2/decode_test
+fi
+
+if [ $stage -le 10 ]; then
+  steps/align_fmllr.sh --nj $nj --cmd $cmd \
+    --use-graphs true \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/tri2 \
+    $exp_dir/tri2_ali
+  steps/train_sat.sh --cmd $cmd \
+    $numLeavesTri $numGaussTri \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/tri2_ali $exp_dir/tri3
+fi
+
+if [ $stage -le 11 ]; then
+  utils/mkgraph.sh $data_dir/lang_test_corpus \
+    $exp_dir/tri3 \
+    $exp_dir/tri3/graph
+  steps/decode_fmllr.sh --nj $nj --cmd $cmd \
+    $exp_dir/tri3/graph \
+    $data_dir/test \
+    $exp_dir/tri3/decode_test
+fi
+
+if [ $stage -le 12 ]; then
+  steps/align_fmllr.sh --nj $nj --cmd $cmd \
+    --use-graphs true \
+    $data_dir/train $data_dir/lang \
+    $exp_dir/tri3 \
+    $exp_dir/tri3_ali
+fi
+
+affix=_1a
+nnet3_affix=
+if [ $stage -le 13 ]; then
+  local/chain/run_cnn_1a.sh --stage 0 \
+   --gmm tri3 \
+   --ali tri3_ali \
+   --affix $affix \
+   --nnet3_affix $nnet3_affix \
+   --lang_test lang_unk
+fi
+
+if [ $stage -le 14 ]; then
+  local/chain/run_cnn_chainali_1a.sh --stage 0 \
+   --gmm tri3 \
+   --ali tri3_ali \
+   --affix $affix \
+   --nnet3_affix $nnet3_affix \
+   --chain_model_dir $exp_dir/chain${nnet3_affix}/cnn${affix} \
+   --lang_test lang_unk
+fi
diff --git a/egs/iam/s5/steps b/egs/iam/s5/steps
new file mode 120000
index 00000000000..6e99bf5b5ad
--- /dev/null
+++ b/egs/iam/s5/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps
\ No newline at end of file
diff --git a/egs/iam/s5/utils b/egs/iam/s5/utils
new file mode 120000
index 00000000000..b240885218f
--- /dev/null
+++ b/egs/iam/s5/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils
\ No newline at end of file