diff --git a/egs/cmu_cslu_kids/README b/egs/cmu_cslu_kids/README
new file mode 100644
index 00000000000..0b8512e2487
--- /dev/null
+++ b/egs/cmu_cslu_kids/README
@@ -0,0 +1,21 @@
+This is an ASR recipe for children speech using cmu_kids and cslu_kids.
+Both of the corpora can be found on LDC:
+    - cmu_kids : https://catalog.ldc.upenn.edu/LDC97S63
+    - cslu_kids: https://catalog.ldc.upenn.edu/LDC2007S18
+
+To run this recipe, you'll need a copy of both corpora:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>
+
+By default, this recipe will download an LM pretrained on LibriSpeech from 
+lm_url=www.openslr.org/resources/11. If you already have a copy of this LM 
+and do not wish to redownload, you can specify the LM path using the --lm_src option:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>\
+        --lm_src <path_to_librispeech_lm>
+
+This recipe will also download and clean CMU_Dict by default. If you have a clean copy 
+already, or wish to use your own dictionary, simply copy your version of the dict to 
+        data/local/dict
+
+To run extra features for triphone models or VLTN, set the following options true:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>\
+        --vtln true --extra_features true
diff --git a/egs/cmu_cslu_kids/s5/cmd.sh b/egs/cmu_cslu_kids/s5/cmd.sh
new file mode 100644
index 00000000000..179307556d5
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/cmd.sh
@@ -0,0 +1,23 @@
+# 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 train_cmd=queue.pl
+export decode_cmd="queue.pl --mem 2G"
+# the use of cuda_cmd is deprecated, used only in 'nnet1',
+export cuda_cmd="queue.pl --gpu 1"
+
+if [[ "$(hostname -f)" == "*.fit.vutbr.cz" ]]; then
+  queue_conf=$HOME/queue_conf/default.conf # see example /homes/kazi/iveselyk/queue_conf/default.conf,
+  export train_cmd="queue.pl --config $queue_conf --mem 2G --matylda 0.2"
+  export decode_cmd="queue.pl --config $queue_conf --mem 3G --matylda 0.1"
+  export cuda_cmd="queue.pl --config $queue_conf --gpu 1 --mem 10G --tmp 40G"
+fi
diff --git a/egs/cmu_cslu_kids/s5/conf/decode.config b/egs/cmu_cslu_kids/s5/conf/decode.config
new file mode 100644
index 00000000000..10b0eee900b
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/decode.config
@@ -0,0 +1,4 @@
+# Use wider-than-normal decoding beams for RM.
+first_beam=16.0
+beam=20.0
+lattice_beam=10.0
diff --git a/egs/cmu_cslu_kids/s5/conf/decode_dnn.config b/egs/cmu_cslu_kids/s5/conf/decode_dnn.config
new file mode 100644
index 00000000000..e7cfca74763
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/decode_dnn.config
@@ -0,0 +1,8 @@
+# In RM, the optimal decode LMWT is in range 2..5, which is different from usual 10..15
+# (it is caused by using simple rule-based LM, instead of n-gram LM),
+scoring_opts="--min-lmwt 2 --max-lmwt 10"
+# Still, it is better to use --acwt 0.1, both for decoding and sMBR,
+acwt=0.1
+# For this small task we can afford to have large beams,
+beam=30.0 # beam for decoding.  Was 13.0 in the scripts.
+lattice_beam=18.0 # this has most effect on size of the lattices.
diff --git a/egs/cmu_cslu_kids/s5/conf/mfcc.conf b/egs/cmu_cslu_kids/s5/conf/mfcc.conf
new file mode 100644
index 00000000000..6bbcb763153
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/mfcc.conf
@@ -0,0 +1,2 @@
+--use-energy=false   # only non-default option.
+--allow_downsample=true
diff --git a/egs/cmu_cslu_kids/s5/conf/mfcc_hires.conf b/egs/cmu_cslu_kids/s5/conf/mfcc_hires.conf
new file mode 100644
index 00000000000..40f95e97010
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/mfcc_hires.conf
@@ -0,0 +1,11 @@
+# config for high-resolution MFCC features, intended for neural network training
+# Note: we keep all cepstra, so it has the same info as filterbank features,
+# but MFCC is more easily compressible (because less correlated) which is why 
+# we prefer this method.
+--use-energy=false   # use average of log energy, not energy.
+--num-mel-bins=40     # similar to Google's setup.
+--num-ceps=40     # there is no dimensionality reduction.
+--low-freq=20     # low cutoff frequency for mel bins... this is high-bandwidth data, so
+                  # there might be some information at the low end.
+--high-freq=-400 # high cutoff frequently, relative to Nyquist of 8000 (=7600)
+--allow-downsample=true
diff --git a/egs/cmu_cslu_kids/s5/conf/online_cmvn.conf b/egs/cmu_cslu_kids/s5/conf/online_cmvn.conf
new file mode 100644
index 00000000000..7748a4a4dd3
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/online_cmvn.conf
@@ -0,0 +1 @@
+# configuration file for apply-cmvn-online, used in the script ../local/run_online_decoding.sh
diff --git a/egs/cmu_cslu_kids/s5/conf/plp.conf b/egs/cmu_cslu_kids/s5/conf/plp.conf
new file mode 100644
index 00000000000..e7e8a9e14af
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/conf/plp.conf
@@ -0,0 +1,2 @@
+# No non-default options for now.
+--allow_downsample=true 
diff --git a/egs/cmu_cslu_kids/s5/local/chain/compare_wer.sh b/egs/cmu_cslu_kids/s5/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..8ee5db2326a
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/chain/compare_wer.sh
@@ -0,0 +1,137 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer.sh exp/chain/tdnn_c_sp exp/chain/tdnn_c_sp_smbr:{1,2,3}
+
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: [--looped] [--online] <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/tdnn_{b,c}_sp"
+  echo "or (with epoch numbers for discriminative training):"
+  echo "$0 exp/chain/tdnn_b_sp_disc:{1,2,3}"
+  exit 1
+fi
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=true
+  shift
+fi
+include_online=false
+if [ "$1" == "--online" ]; then
+  include_online=true
+  shift
+fi
+
+
+used_epochs=false
+
+# this function set_names is used to separate the epoch-related parts of the name
+# [for discriminative training] and the regular parts of the name.
+# If called with a colon-free directory name, like:
+#  set_names exp/chain/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain/tdnn_d_sp_smbr and epoch_infix="_epoch3"
+
+
+set_names() {
+  if [ $# != 1 ]; then
+    echo "compare_wer_general.sh: internal error"
+    exit 1  # exit the program
+  fi
+  dirname=$(echo $1 | cut -d: -f1)
+  epoch=$(echo $1 | cut -s -d: -f2)
+  if [ -z $epoch ]; then
+    epoch_infix=""
+  else
+    used_epochs=true
+    epoch_infix=_epoch${epoch}
+  fi
+}
+
+
+
+echo -n "# System               "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+strings=(
+  "#WER dev_clean_2 (tgsmall) "
+  "#WER dev_clean_2 (tglarge) ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+    decode_names=(tgsmall_dev_clean_2 tglarge_dev_clean_2)
+
+     wer=$(cat $dirname/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+     printf "% 10s" $wer
+   done
+   echo
+   if $include_looped; then
+     echo -n "#             [looped:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat $dirname/decode_looped_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+   if $include_online; then
+     echo -n "#             [online:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat ${dirname}_online/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+done
+
+
+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
+
+echo -n "# Num-params              "
+for x in $*; do
+  printf "% 10s" $(grep num-parameters $x/log/progress.1.log | awk '{print $2}')
+done
+echo
diff --git a/egs/cmu_cslu_kids/s5/local/chain/run_tdnnf.sh b/egs/cmu_cslu_kids/s5/local/chain/run_tdnnf.sh
new file mode 120000
index 00000000000..34499362831
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/chain/run_tdnnf.sh
@@ -0,0 +1 @@
+tuning/run_tdnn_1a.sh
\ No newline at end of file
diff --git a/egs/cmu_cslu_kids/s5/local/chain/tdnnf_decode.sh b/egs/cmu_cslu_kids/s5/local/chain/tdnnf_decode.sh
new file mode 100755
index 00000000000..8d124193584
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/chain/tdnnf_decode.sh
@@ -0,0 +1,82 @@
+#! /bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Decode on new data set using trained model. 
+# The data directory should be prepared in kaldi style.
+# Usage:
+#     ./local/chain/tdnnF_decode.sh --data_src <prepared_data_dir> 
+
+set -euo pipefail
+echo "$0 $@"
+
+stage=0
+decode_nj=10
+data_src=
+affix=
+tree_affix=
+nnet3_affix=
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+data_name=$(basename $data_src)
+data_hires="data/${data_name}_hires"
+ivect_dir=exp/nnet3${nnet3_affix}/ivector_$data_name
+tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
+dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+
+mfcc=mfcc_hires_$data_name
+chunk_width=140,100,160
+reporting_email=
+
+if [ $stage -le 0 ]; then 
+    rm -rf $data_hires
+    cp -r $data_src $data_hires
+fi
+# High resolution mfcc
+if [ $stage -le 1 ]; then 
+    steps/make_mfcc.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
+        --cmd "$train_cmd" $data_hires \
+        exp/$data_name/make_feat_hires  $mfcc|| exit 1;
+    steps/compute_cmvn_stats.sh $data_hires || exit 1;
+    utils/fix_data_dir.sh $data_hires || exit 1;
+fi
+
+# Extract i-vector
+if [ $stage -le 2 ]; then 
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 20 \
+        $data_hires exp/nnet3"$affix"/extractor $ivect_dir
+fi
+
+if [ $stage -le 3 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+    (
+      nspk=$(wc -l <$data_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --frames-per-chunk $frames_per_chunk \
+          --nj $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir $ivect_dir \
+          $tree_dir/graph_tgsmall $data_hires ${dir}/decode_tgsmall_$data_name || exit 1
+      
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       $data_hires ${dir}/decode_{tgsmall,tglarge}_$data_name || exit 1
+    ) || touch $dir/.error &
+
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
diff --git a/egs/cmu_cslu_kids/s5/local/chain/tuning/run_tdnn_1a.sh b/egs/cmu_cslu_kids/s5/local/chain/tuning/run_tdnn_1a.sh
new file mode 100755
index 00000000000..51e0123d0f2
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/chain/tuning/run_tdnn_1a.sh
@@ -0,0 +1,279 @@
+#!/bin/bash
+
+# Copyright 2017-2018  Johns Hopkins University (author: Daniel Povey)
+#           2017-2018  Yiming Wang
+#           2019       Fei Wu
+
+# Based on material recipe for low-resource languages
+# Factored TDNN with skip connectiong and splicing (two bottle neck layers)
+
+#   WER results on dev
+#   Model       LM          Corpus      WER(%)
+#   tdnn_1a     tg_large    Combined    11.72
+#   tdnn_1a     tg_small    Combined    13.61
+#   tdnn_1a     tg_large    CMU_Kids    17.26
+#   tdnn_1a     tg_small    CMU_Kids    26.43
+#   tdnn_1a     tg_large    CSLU_Kids   10.80
+#   tdnn_1a     tg_small    CSLU_Kids   12.50
+
+# steps/info/chain_dir_info.pl exp/chain/tdnn1a_sp
+# exp/chain/tdnn1a_sp/: num-iters=342 nj=2..5 num-params=17.9M dim=40+100->3192 combine=-0.042->-0.041 (over 8) xent:train/valid[227,341,final]=(-0.451,-0.363,-0.346/-0.524,-0.466,-0.434) logprob:train/valid[227,341,final]=(-0.047,-0.043,-0.042/-0.058,-0.056,-0.054) 
+
+set -euo pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+nj=10
+train_set=train
+test_sets="test"
+gmm=tri3       
+nnet3_affix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=1a  
+tree_affix=
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training chunk-options
+chunk_width=140,100,160
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+common_egs_dir=
+xent_regularize=0.1
+
+# training options
+srand=0
+remove_egs=true
+reporting_email=
+
+
+# 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
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 7" if you have already
+# run those things.
+local/nnet3/run_ivector_common.sh\
+    --stage $stage \
+    --train-set $train_set \
+    --test-sets $test_sets \
+    --gmm $gmm \
+    --nnet3_affix "$nnet3_affix" || exit 1;
+
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
+lang=data/lang_chain
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
+dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+lores_train_data_dir=data/${train_set}_sp
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $lores_train_data_dir/feats.scp $ali_dir/ali.1.gz; do
+    [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 7 ]; 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/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 $lang
+    silphonelist=$(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 [ $stage -le 8 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj 75 --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+
+if [ $stage -le 10 ]; 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: $tree_dir/final.mdl already exists, refusing to overwrite it."
+       exit 1;
+    fi
+    steps/nnet3/chain/build_tree.sh \
+      --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 3500 ${lores_train_data_dir} \
+      $lang $ali_dir $tree_dir
+fi
+
+if [ $stage -le 11 ]; 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)
+    opts="l2-regularize=0.004 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
+    linear_opts="orthonormal-constraint=-1.0 l2-regularize=0.004"
+    output_opts="l2-regularize=0.002"
+    
+    mkdir -p $dir/configs
+    
+    cat <<EOF > $dir/configs/network.xconfig
+    input dim=100 name=ivector
+    input dim=40 name=input
+    
+    # please note that it is important to have input layer with the name=input
+    # as the layer immediately preceding the fixed-affine-layer to enable
+    # the use of short notation for the descriptor
+    fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+    
+    # the first splicing is moved before the lda layer, so no splicing here
+    relu-batchnorm-dropout-layer name=tdnn1 $opts dim=1024
+    linear-component name=tdnn2l0 dim=256 $linear_opts input=Append(-1,0)
+    linear-component name=tdnn2l dim=256 $linear_opts input=Append(-1,0)
+    relu-batchnorm-dropout-layer name=tdnn2 $opts input=Append(0,1) dim=1024
+    linear-component name=tdnn3l dim=256 $linear_opts input=Append(-1,0)
+    relu-batchnorm-dropout-layer name=tdnn3 $opts dim=1024 input=Append(0,1)
+    linear-component name=tdnn4l0 dim=256 $linear_opts input=Append(-1,0)
+    linear-component name=tdnn4l dim=256 $linear_opts input=Append(0,1)
+    relu-batchnorm-dropout-layer name=tdnn4 $opts input=Append(0,1) dim=1024
+    linear-component name=tdnn5l dim=256 $linear_opts
+    relu-batchnorm-dropout-layer name=tdnn5 $opts dim=1024 input=Append(0, tdnn3l)
+    linear-component name=tdnn6l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn6l dim=256 $linear_opts input=Append(-3,0)
+    relu-batchnorm-dropout-layer name=tdnn6 $opts input=Append(0,3) dim=1280
+    linear-component name=tdnn7l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn7l dim=256 $linear_opts input=Append(0,3)
+    relu-batchnorm-dropout-layer name=tdnn7 $opts input=Append(0,3,tdnn6l,tdnn4l,tdnn2l) dim=1024
+    linear-component name=tdnn8l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn8l dim=256 $linear_opts input=Append(0,3)
+    relu-batchnorm-dropout-layer name=tdnn8 $opts input=Append(0,3) dim=1280
+    linear-component name=tdnn9l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn9l dim=256 $linear_opts input=Append(-3,0)
+    relu-batchnorm-dropout-layer name=tdnn9 $opts input=Append(0,3,tdnn8l,tdnn6l,tdnn5l) dim=1024
+    linear-component name=tdnn10l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn10l dim=256 $linear_opts input=Append(0,3)
+    relu-batchnorm-dropout-layer name=tdnn10 $opts input=Append(0,3) dim=1280
+    linear-component name=tdnn11l0 dim=256 $linear_opts input=Append(-3,0)
+    linear-component name=tdnn11l dim=256 $linear_opts input=Append(-3,0)
+    relu-batchnorm-dropout-layer name=tdnn11 $opts input=Append(0,3,tdnn10l,tdnn9l,tdnn7l) dim=1024
+    linear-component name=prefinal-l dim=256 $linear_opts
+    
+    relu-batchnorm-layer name=prefinal-chain input=prefinal-l $opts dim=1280
+    linear-component name=prefinal-chain-l dim=256 $linear_opts
+    batchnorm-component name=prefinal-chain-batchnorm
+    output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+    
+    relu-batchnorm-layer name=prefinal-xent input=prefinal-l $opts dim=1280
+    linear-component name=prefinal-xent-l dim=256 $linear_opts
+    batchnorm-component name=prefinal-xent-batchnorm
+    output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+    
+EOF
+
+    steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+    
+fi
+
+
+if [ $stage -le 12 ]; 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/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+    fi
+    
+    steps/nnet3/chain/train.py --stage=$train_stage \
+      --cmd="$decode_cmd" \
+      --feat.online-ivector-dir=$train_ivector_dir \
+      --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+      --chain.xent-regularize $xent_regularize \
+      --chain.leaky-hmm-coefficient=0.1 \
+      --chain.l2-regularize=0.0 \
+      --chain.apply-deriv-weights=false \
+      --chain.lm-opts="--num-extra-lm-states=2000" \
+      --trainer.dropout-schedule $dropout_schedule \
+      --trainer.add-option="--optimization.memory-compression-level=2" \
+      --trainer.srand=$srand \
+      --trainer.max-param-change=2.0 \
+      --trainer.num-epochs=20 \
+      --trainer.frames-per-iter=3000000 \
+      --trainer.optimization.num-jobs-initial=2 \
+      --trainer.optimization.num-jobs-final=5 \
+      --trainer.optimization.initial-effective-lrate=0.002 \
+      --trainer.optimization.final-effective-lrate=0.0002 \
+      --trainer.num-chunk-per-minibatch=128,64 \
+      --egs.chunk-width=$chunk_width \
+      --egs.dir="$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 13 ]; then
+    # Note: it's not important to give mkgraph.sh the lang directory with the
+    # matched topology (since it gets the topology file from the model).
+    utils/mkgraph.sh \
+      --self-loop-scale 1.0 data/lang_test_tgsmall \
+      $tree_dir $tree_dir/graph_tgsmall || exit 1;
+fi
+
+if [ $stage -le 14 ]; then
+    frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+    rm $dir/.error 2>/dev/null || true
+
+    for data in $test_sets; do
+        (
+        nspk=$(wc -l <data/${data}_hires/spk2utt)
+        steps/nnet3/decode.sh \
+            --acwt 1.0 --post-decode-acwt 10.0 \
+            --frames-per-chunk $frames_per_chunk \
+            --nj $nspk --cmd "$decode_cmd"  --num-threads 4 \
+            --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \
+            $tree_dir/graph_tgsmall data/${data}_hires ${dir}/decode_tgsmall_${data} || exit 1
+        steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+        data/${data}_hires ${dir}/decode_{tgsmall,tglarge}_${data} || exit 1
+        ) || touch $dir/.error &
+    done
+    wait
+    [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+exit 0
diff --git a/egs/cmu_cslu_kids/s5/local/cmu_prepare_data.sh b/egs/cmu_cslu_kids/s5/local/cmu_prepare_data.sh
new file mode 100755
index 00000000000..d452b8574b6
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/cmu_prepare_data.sh
@@ -0,0 +1,89 @@
+#! /bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Prepares cmu_kids. 
+# Should be run from egs/cmu_cslu_kids
+
+set -eu
+corpus=cmu_kids/kids
+data=data/data_cmu
+test_percentage=30
+
+. ./path.sh
+. ./utils/parse_options.sh
+
+total_cnt=0
+test_cnt=0
+train_cnt=0
+
+for d in $data/train $data/test; do
+    mkdir -p $d
+    ./local/file_check.sh $d
+done
+
+echo "Preparing cmu_kids..."
+for kid in $corpus/*; do 
+	if [ -d $kid ]; then
+        # echo "Kid: $kid"
+		spkID=$(basename $kid)
+		sph="$kid/signal"
+	    if [ -d $sph ];then
+            # echo "$sph"
+            for utt in $sph/*; do
+                if [ ${utt: -4} == ".sph" ]; then
+                    total_cnt=$[$total_cnt+1]   
+                    rnd=$((1+RANDOM % 100))
+                    uttID=$(basename $utt)
+                    uttID=${uttID%".sph"}
+                    sentID=${uttID#$spkID}
+                    sentID=${sentID:0:3}
+
+                    # Find the sentence
+                    grep $sentID cmu_kids/tables/sentence.tbl > tmp
+                    cut -f 3- < tmp > out                    
+        
+                    tr '[:lower:]' '[:upper:]' < out > tmp
+                    tr -d '[:cntrl:]' < tmp > out
+                    sent=$(<out)
+
+                    # Clean transcript 
+                    cp $kid/trans/$uttID.trn tmp
+                    tr -d '\n' < tmp > out
+                    tr '[:lower:]' '[:upper:]' < tmp > out
+                    trans=$(<out)
+                     
+                    if [ $rnd -le $test_percentage ]; then
+                        target="test"
+                        test_cnt=$[$test_cnt+1]
+                    else
+                        target="train"
+                        train_cnt=$[$train_cnt+1]
+                    fi
+
+                    echo "$uttID $spkID" >> $data/$target/utt2spk
+                    echo "$uttID $KALDI_ROOT/tools/sph2pipe_v2.5/sph2pipe -f wav -p -c 1 $utt|" >> $data/$target/wav.scp
+                    echo "$spkID f" >> $data/$target/spk2gender
+                    echo "$uttID $sent" >> $data/$target/text
+                fi
+            done
+        fi
+	fi
+done
+
+for d in $data/train $data/test; do
+    utils/utt2spk_to_spk2utt.pl $d/utt2spk > $d/spk2utt
+    utils/fix_data_dir.sh $d
+done
+
+printf "\t total: %s; train: %s; test: %s.\n" "$total_cnt" "$train_cnt" "$test_cnt" 
+rm -f out tmp
+
+# Optional
+# Get data duration, just for book keeping
+# for data in $data/train $data/test; do
+#     ./local/data_duration.sh $data
+# done
+# 
+
diff --git a/egs/cmu_cslu_kids/s5/local/cslu_aud_prep.sh b/egs/cmu_cslu_kids/s5/local/cslu_aud_prep.sh
new file mode 100755
index 00000000000..735f87eca9f
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/cslu_aud_prep.sh
@@ -0,0 +1,43 @@
+#/bin/bash 
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Called by local/cslu_DataPrep.shi
+
+Assignment()
+{
+    rnd=$((1+RANDOM % 100))
+    if [ $rnd -le $test_percentage ]; then 
+        target="test"
+    else
+        target="train"
+    fi
+}
+audio=
+test_percentage=30  # Percent of data reserved as test set 
+debug=debug/cslu_dataprep_debug
+data=data/data_cslu
+. ./utils/parse_options.sh
+
+uttID=$(basename $audio)
+uttID=${uttID%'.wav'}
+sentID=${uttID: -3}
+spkID=${uttID%$sentID}
+sentID=${sentID%"0"}
+sentID=$(echo "$sentID" | tr '[:lower:]' '[:upper:]' )
+
+line=$(grep $sentID cslu/docs/all.map)
+
+if [ -z "$line" ]; then     # Can't map utterance to transcript
+    echo $audio $sentID >> $debug
+else
+    txt=$(echo $line | grep -oP '"\K.*?(?=")')
+    cap_txt=${txt^^}
+    Assignment
+    echo "$uttID $cap_txt" >> $data/$target/text
+    echo "$uttID $spkID" >> $data/$target/utt2spk
+    echo "$spkID f" >> $data/$target/spk2gender
+    echo "$uttID $audio" >> $data/$target/wav.scp
+fi
+
diff --git a/egs/cmu_cslu_kids/s5/local/cslu_prepare_data.sh b/egs/cmu_cslu_kids/s5/local/cslu_prepare_data.sh
new file mode 100755
index 00000000000..621179079b3
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/cslu_prepare_data.sh
@@ -0,0 +1,49 @@
+#! /bin/bash 
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Prepares cslu_kids
+# Should be run from egs/cmu_csli_kids
+
+set -e
+Looper()
+{
+    # echo "Looping through $1"
+    for f in $1/*; do 
+        if [ -d $f ]; then
+            Looper $f
+        else            
+            ./local/cslu_aud_prep.sh --data $data --audio $f
+        fi
+    done
+}
+
+data=data/data_cslu
+corpus=cslu
+. ./utils/parse_options.sh
+
+rm -f debug/cslu_dataprep_debug
+mkdir -p debug
+# File check, remove previous data and features files 
+for d in $data/test $data/train; do 
+    mkdir -p $d
+    ./local/file_check.sh $d
+done
+
+echo "Preparing cslu_kids..."
+Looper $corpus/speech/scripted
+
+for d in $data/test $data/train; do
+    ./utils/utt2spk_to_spk2utt.pl $d
+    ./utils/fix_data_dir.sh $d
+done
+if [ -f debug/cslu_dataprep_debug ]; then
+    echo "Missing transcripts for some utterances. See cslu_dataprep_debug"
+fi
+
+# Optional
+# Get data duration, just for book keeping
+# for data in data/data_cslu/test data/data_cslu/train; do 
+#     ./local/data_duration.sh $data
+# done
diff --git a/egs/cmu_cslu_kids/s5/local/data_duration.sh b/egs/cmu_cslu_kids/s5/local/data_duration.sh
new file mode 100755
index 00000000000..e838e365ea7
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/data_duration.sh
@@ -0,0 +1,19 @@
+#! /bin/bash 
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Get duration of the utterance given data dir
+set -eu
+echo $0 $@
+
+data_dir=$1
+mkdir -p duration
+
+./utils/data/get_utt2dur.sh $data_dir
+
+echo "$data_dir"
+python local/sum_duration.py $data_dir/utt2dur 
+echo ""
+
+
diff --git a/egs/cmu_cslu_kids/s5/local/download_cmu_dict.sh b/egs/cmu_cslu_kids/s5/local/download_cmu_dict.sh
new file mode 100755
index 00000000000..0248dd0cae1
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/download_cmu_dict.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+# Copyright 2019 Fei Wu
+set -eu 
+# Adapted from the local/prepare_dict script in 
+# the librispeech recipe. Download and prepare CMU_dict.
+# For childresn speech ASR tasks, since the vocabulary in cmu_kids and 
+# cslu_kids is relatively easy comparing to librispeech, we use only the 
+# CMU_dict, and do not handle OOV with G2P.
+# Should be run from egs/cmu_cslu_kids.
+# Usage:
+#   local/download_cmu_dict.sh --dict_dir <path_to_dict_dir>
+
+dict_dir=data/local/dict
+OOV="<UNK>"
+
+. ./utils/parse_options.sh || exit 1;
+. ./path.sh || exit 1
+
+if [ ! -d $dict_dir ]; then
+  echo "Downloading and preparing CMU dict"
+  svn co -r 12440 https://svn.code.sf.net/p/cmusphinx/code/trunk/cmudict $dict_dir/raw_dict || exit 1;
+  
+  echo "Removing the pronunciation variant markers ..."
+  grep -v ';;;' $dict_dir/raw_dict/cmudict.0.7a | \
+  perl -ane 'if(!m:^;;;:){ s:(\S+)\(\d+\) :$1 :; print; }' | \
+  sort -u > $dict_dir/lexicon.txt || exit 1;
+
+  tr -d '\r' <  $dict_dir/raw_dict/cmudict.0.7a.symbols > $dict_dir/nonsilence_phones.txt
+  
+  echo "$OOV SIL" >> $dict_dir/lexicon.txt
+  
+  echo "SIL" > $dict_dir/silence_phones.txt
+  echo "SPN" >> $dict_dir/silence_phones.txt
+  echo "SIL" > $dict_dir/optional_silence.txt
+
+  rm -rf $dict_dir/raw_dict
+fi
diff --git a/egs/cmu_cslu_kids/s5/local/download_lm.sh b/egs/cmu_cslu_kids/s5/local/download_lm.sh
new file mode 100755
index 00000000000..382f313df7c
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/download_lm.sh
@@ -0,0 +1,76 @@
+#!/bin/bash
+
+# Copyright 2014 Vassil Panayotov
+# Apache 2.0
+
+if [ $# -ne "2" ]; then
+  echo "Usage: $0 <base-url> <download_dir>"
+  echo "e.g.: $0 http://www.openslr.org/resources/11 data/local/lm"
+  exit 1
+fi
+
+base_url=$1
+dst_dir=$2
+
+# given a filename returns the corresponding file size in bytes
+# The switch cases below can be autogenerated by entering the data directory and running:
+# for f in *; do echo "\"$f\") echo \"$(du -b $f | awk '{print $1}')\";;"; done
+function filesize() {
+  case $1 in
+    "3-gram.arpa.gz") echo "759636181";;
+    "3-gram.pruned.1e-7.arpa.gz") echo "34094057";;
+    "3-gram.pruned.3e-7.arpa.gz") echo "13654242";;
+    "4-gram.arpa.gz") echo "1355172078";;
+    "g2p-model-5") echo "20098243";;
+    "librispeech-lexicon.txt") echo "5627653";;
+    "librispeech-lm-corpus.tgz") echo "1803499244";;
+    "librispeech-lm-norm.txt.gz") echo "1507274412";;
+    "librispeech-vocab.txt") echo "1737588";;
+    *) echo "";;
+  esac
+}
+
+function check_and_download () {
+  [[ $# -eq 1 ]] || { echo "check_and_download() expects exactly one argument!"; return 1; }
+  fname=$1
+  echo "Downloading file '$fname' into '$dst_dir'..."
+  expect_size="$(filesize $fname)"
+  [[ ! -z "$expect_size" ]] || { echo "Unknown file size for '$fname'"; return 1; }
+  if [[ -s $dst_dir/$fname ]]; then
+    # In the following statement, the first version works on linux, and the part
+    # after '||' works on Linux.
+    f=$dst_dir/$fname
+    fsize=$(set -o pipefail; du -b $f 2>/dev/null | awk '{print $1}' || stat '-f %z' $f)
+    if [[ "$fsize" -eq "$expect_size" ]]; then
+      echo "'$fname' already exists and appears to be complete"
+      return 0
+    else
+      echo "WARNING: '$fname' exists, but the size is wrong - re-downloading ..."
+    fi
+  fi
+  wget --no-check-certificate -O $dst_dir/$fname $base_url/$fname || {
+    echo "Error while trying to download $fname!"
+    return 1
+  }
+  f=$dst_dir/$fname
+  # In the following statement, the first version works on linux, and the part after '||'
+  # works on Linux.
+  fsize=$(set -o pipefail; du -b $f 2>/dev/null | awk '{print $1}' || stat '-f %z' $f)
+  [[ "$fsize" -eq "$expect_size" ]] || { echo "$fname: file size mismatch!"; return 1; }
+  return 0
+}
+
+mkdir -p $dst_dir
+
+for f in 3-gram.arpa.gz 3-gram.pruned.1e-7.arpa.gz 3-gram.pruned.3e-7.arpa.gz 4-gram.arpa.gz \
+         g2p-model-5 librispeech-lm-corpus.tgz librispeech-vocab.txt librispeech-lexicon.txt; do
+  check_and_download $f || exit 1
+done
+
+cd $dst_dir
+ln -sf 3-gram.pruned.1e-7.arpa.gz lm_tgmed.arpa.gz
+ln -sf 3-gram.pruned.3e-7.arpa.gz lm_tgsmall.arpa.gz
+ln -sf 3-gram.arpa.gz lm_tglarge.arpa.gz
+ln -sf 4-gram.arpa.gz lm_fglarge.arpa.gz
+
+exit 0
diff --git a/egs/cmu_cslu_kids/s5/local/file_check.sh b/egs/cmu_cslu_kids/s5/local/file_check.sh
new file mode 100755
index 00000000000..859f228058a
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/file_check.sh
@@ -0,0 +1,17 @@
+#! /bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+
+printf "\t File Check in folder: %s.\n" "$1"
+
+WavScp="$1/wav.scp"
+Text="$1/text"
+Utt2Spk="$1/utt2spk"
+Gend="$1/utt2gender"
+Spk2Utt="$1/spk2utt"
+rm -f $WavScp $Text $Utt2Spk $Gend $Spk2Utt
+
+
+
diff --git a/egs/cmu_cslu_kids/s5/local/format_lms.sh b/egs/cmu_cslu_kids/s5/local/format_lms.sh
new file mode 100755
index 00000000000..b530f61d2d9
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/format_lms.sh
@@ -0,0 +1,60 @@
+#!/bin/bash
+
+# Copyright 2014 Vassil Panayotov
+# Apache 2.0
+
+# Prepares the test time language model(G) transducers
+# (adapted from wsj/s5/local/wsj_format_data.sh)
+
+. ./path.sh || exit 1;
+
+# begin configuration section
+src_dir=data/lang
+# end configuration section
+
+. utils/parse_options.sh || exit 1;
+
+set -e
+
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 <lm-dir>"
+  echo "e.g.: $0 /export/a15/vpanayotov/data/lm"
+  echo ", where:"
+  echo "    <lm-dir> is the directory in which the language model is stored/downloaded"
+  echo "Options:"
+  echo "   --src-dir  <dir>           # source lang directory, default data/lang"
+  exit 1
+fi
+
+lm_dir=$1
+
+if [ ! -d $lm_dir ]; then
+  echo "$0: expected source LM directory $lm_dir to exist"
+  exit 1;
+fi
+if [ ! -f $src_dir/words.txt ]; then
+  echo "$0: expected $src_dir/words.txt to exist."
+  exit 1;
+fi
+
+
+tmpdir=data/local/lm_tmp.$$
+trap "rm -r $tmpdir" EXIT
+
+mkdir -p $tmpdir
+
+for lm_suffix in tgsmall tgmed; do
+  # tglarge is prepared by a separate command, called from run.sh; we don't
+  # want to compile G.fst for tglarge, as it takes a while.
+  test=${src_dir}_test_${lm_suffix}
+  mkdir -p $test
+  cp -r ${src_dir}/* $test
+  gunzip -c $lm_dir/lm_${lm_suffix}.arpa.gz | \
+    arpa2fst --disambig-symbol=#0 \
+             --read-symbol-table=$test/words.txt - $test/G.fst
+  utils/validate_lang.pl --skip-determinization-check $test || exit 1;
+done
+
+echo "Succeeded in formatting data."
+
+exit 0
diff --git a/egs/cmu_cslu_kids/s5/local/make_lm.pl b/egs/cmu_cslu_kids/s5/local/make_lm.pl
new file mode 100755
index 00000000000..80eea5a6198
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/make_lm.pl
@@ -0,0 +1,119 @@
+#!/usr/bin/env perl
+
+# Copyright 2010-2011 Yanmin Qian  Microsoft Corporation
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+# This file takes as input the file wp_gram.txt that comes with the RM
+# distribution, and creates the language model as an acceptor in FST form.
+
+# make_rm_lm.pl   wp_gram.txt > G.txt
+
+if (@ARGV != 1) {
+    print "usage: make_rm_lm.pl  wp_gram.txt > G.txt\n";
+    exit(0);
+}
+unless (open(IN_FILE, "@ARGV[0]")) {
+    die ("can't open @ARGV[0]");
+}
+
+
+$flag = 0;
+$count_wrd = 0;
+$cnt_ends = 0;
+$init = "";
+
+while ($line = <IN_FILE>)
+{	
+	chop($line);        # Return the last char
+
+    $line =~ s/ //g;    # Selete all spaces 
+    
+	if(($line =~ /^>/)) # If line has ">"
+	{
+		if($flag == 0)          # Flip flag
+		{
+			$flag = 1;
+		}
+		$line =~ s/>//g;        # Delete ">" 
+		$hashcnt{$init} = $i;
+		$init = $line;
+		$i = 0;
+		$count_wrd++;
+		@LineArray[$count_wrd - 1] = $init;
+ 		$hashwrd{$init} = 0;
+	}
+	elsif($flag != 0)
+	{
+		
+		$hash{$init}[$i] = $line;
+		$i++; 			
+		if($line =~ /SENTENCE-END/)
+		{
+			$cnt_ends++;
+		}
+ 	} 
+	else
+	{}
+}
+
+$hashcnt{$init} = $i;
+
+$num = 0;
+$weight = 0;
+$init_wrd = "SENTENCE-END";
+$hashwrd{$init_wrd} = @LineArray;
+for($i = 0; $i < $hashcnt{$init_wrd}; $i++)
+{
+	$weight = -log(1/$hashcnt{$init_wrd});
+	$hashwrd{$hash{$init_wrd}[$i]} = $i + 1;
+	print "0    $hashwrd{$hash{$init_wrd}[$i]}    $hash{$init_wrd}[$i]    $hash{$init_wrd}[$i]    $weight\n";
+}
+$num = $i;
+
+for($i = 0; $i < @LineArray; $i++)
+{
+	if(@LineArray[$i] eq 'SENTENCE-END')
+	{}
+	else
+	{
+		if($hashwrd{@LineArray[$i]} == 0)
+		{
+			$num++;
+			$hashwrd{@LineArray[$i]} = $num;
+		}
+		for($j = 0; $j < $hashcnt{@LineArray[$i]}; $j++)
+		{
+			$weight = -log(1/$hashcnt{@LineArray[$i]});
+			if($hashwrd{$hash{@LineArray[$i]}[$j]} == 0)
+			{
+				$num++;
+				$hashwrd{$hash{@LineArray[$i]}[$j]} = $num;
+			}
+			if($hash{@LineArray[$i]}[$j] eq 'SENTENCE-END')
+			{
+				print "$hashwrd{@LineArray[$i]}    $hashwrd{$hash{@LineArray[$i]}[$j]}    <eps>    <eps>    $weight\n"
+                }
+			else
+			{
+				print "$hashwrd{@LineArray[$i]}    $hashwrd{$hash{@LineArray[$i]}[$j]}    $hash{@LineArray[$i]}[$j]    $hash{@LineArray[$i]}[$j]    $weight\n";
+			}
+		}
+	}
+}
+
+print "$hashwrd{$init_wrd}    0\n";
+close(IN_FILE);
+
+
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/compare_wer.sh b/egs/cmu_cslu_kids/s5/local/nnet3/compare_wer.sh
new file mode 100755
index 00000000000..095e85cc338
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/compare_wer.sh
@@ -0,0 +1,132 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer.sh exp/chain/tdnn_c_sp exp/chain/tdnn_c_sp_smbr:{1,2,3}
+
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: [--looped] [--online] <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/tdnn_{b,c}_sp"
+  echo "or (with epoch numbers for discriminative training):"
+  echo "$0 exp/chain/tdnn_b_sp_disc:{1,2,3}"
+  exit 1
+fi
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=true
+  shift
+fi
+include_online=false
+if [ "$1" == "--online" ]; then
+  include_online=true
+  shift
+fi
+
+
+used_epochs=false
+
+# this function set_names is used to separate the epoch-related parts of the name
+# [for discriminative training] and the regular parts of the name.
+# If called with a colon-free directory name, like:
+#  set_names exp/chain/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain/tdnn_d_sp_smbr and epoch_infix="_epoch3"
+
+
+set_names() {
+  if [ $# != 1 ]; then
+    echo "compare_wer_general.sh: internal error"
+    exit 1  # exit the program
+  fi
+  dirname=$(echo $1 | cut -d: -f1)
+  epoch=$(echo $1 | cut -s -d: -f2)
+  if [ -z $epoch ]; then
+    epoch_infix=""
+  else
+    used_epochs=true
+    epoch_infix=_epoch${epoch}
+  fi
+}
+
+
+
+echo -n "# System               "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+strings=(
+  "#WER dev_clean_2 (tgsmall) "
+  "#WER dev_clean_2 (tglarge) ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+    decode_names=(tgsmall_dev_clean_2 tglarge_dev_clean_2)
+
+     wer=$(cat $dirname/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+     printf "% 10s" $wer
+   done
+   echo
+   if $include_looped; then
+     echo -n "#             [looped:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat $dirname/decode_looped_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+   if $include_online; then
+     echo -n "#             [online:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat ${dirname}_online/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+done
+
+
+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,combined}.log 2>/dev/null | grep log-like | 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,combined}.log 2>/dev/null | grep log-like | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train acc      "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.{final,combined}.log 2>/dev/null | grep accuracy | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid acc      "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.{final,combined}.log 2>/dev/null | grep accuracy | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/run_ivector_common.sh b/egs/cmu_cslu_kids/s5/local/nnet3/run_ivector_common.sh
new file mode 100755
index 00000000000..c695f2c9f1c
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/run_ivector_common.sh
@@ -0,0 +1,148 @@
+#!/bin/bash
+
+set -euo pipefail
+
+# This script is called from local/nnet3/run_tdnn.sh and
+# local/chain/run_tdnn.sh (and may eventually be called by more
+# scripts).  It contains the common feature preparation and
+# iVector-related parts of the script.  See those scripts for examples
+# of usage.
+
+stage=0
+train_set=train
+test_sets="test"
+gmm=tri3b
+
+nnet3_affix=
+
+. ./cmd.sh
+. ./path.sh
+. utils/parse_options.sh
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+
+for f in data/${train_set}/feats.scp ${gmm_dir}/final.mdl; do
+  if [ ! -f $f ]; then
+    echo "$0: expected file $f to exist"
+    exit 1
+  fi
+done
+
+if [ $stage -le 1 ]; then
+  # Although the nnet will be trained by high resolution data, we still have to
+  # perturb the normal data to get the alignment _sp stands for speed-perturbed
+  echo "$0: preparing directory for low-resolution speed-perturbed data (for alignment)"
+  utils/data/perturb_data_dir_speed_3way.sh data/${train_set} data/${train_set}_sp
+  echo "$0: making MFCC features for low-resolution speed-perturbed data"
+  steps/make_mfcc.sh --cmd "$train_cmd" --nj 10 data/${train_set}_sp || exit 1;
+  steps/compute_cmvn_stats.sh data/${train_set}_sp || exit 1;
+  utils/fix_data_dir.sh data/${train_set}_sp
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: aligning with the perturbed low-resolution data"
+  steps/align_fmllr.sh --nj 20 --cmd "$train_cmd" \
+    data/${train_set}_sp data/lang $gmm_dir $ali_dir || exit 1
+fi
+
+if [ $stage -le 3 ]; then
+  # Create high-resolution MFCC features (with 40 cepstra instead of 13).
+  # this shows how you can split across multiple file-systems.
+  echo "$0: creating high-resolution MFCC features"
+  mfccdir=data/${train_set}_sp_hires/data
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
+    utils/create_split_dir.pl /export/fs0{1,2}/$USER/kaldi-data/mfcc/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$mfccdir/storage $mfccdir/storage
+  fi
+
+  for datadir in ${train_set}_sp ${test_sets}; do
+    utils/copy_data_dir.sh data/$datadir data/${datadir}_hires
+  done
+
+  # do volume-perturbation on the training data prior to extracting hires
+  # features; this helps make trained nnets more invariant to test data volume.
+  utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires || exit 1;
+
+  for datadir in ${train_set}_sp ${test_sets}; do
+    steps/make_mfcc.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
+      --cmd "$train_cmd" data/${datadir}_hires || exit 1;
+    steps/compute_cmvn_stats.sh data/${datadir}_hires || exit 1;
+    utils/fix_data_dir.sh data/${datadir}_hires || exit 1;
+  done
+fi
+
+if [ $stage -le 4 ]; then
+  echo "$0: computing a subset of data to train the diagonal UBM."
+  # We'll use about a quarter of the data.
+  mkdir -p exp/nnet3${nnet3_affix}/diag_ubm
+  temp_data_root=exp/nnet3${nnet3_affix}/diag_ubm
+
+  num_utts_total=$(wc -l <data/${train_set}_sp_hires/utt2spk)
+  num_utts=$[$num_utts_total/4]
+  utils/data/subset_data_dir.sh data/${train_set}_sp_hires \
+     $num_utts ${temp_data_root}/${train_set}_sp_hires_subset
+
+  echo "$0: computing a PCA transform from the hires data."
+  steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
+      --splice-opts "--left-context=3 --right-context=3" \
+      --max-utts 10000 --subsample 2 \
+       ${temp_data_root}/${train_set}_sp_hires_subset \
+       exp/nnet3${nnet3_affix}/pca_transform
+
+  echo "$0: training the diagonal UBM."
+  # Use 512 Gaussians in the UBM.
+  steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj 30 \
+    --num-frames 700000 \
+    --num-threads 8 \
+    ${temp_data_root}/${train_set}_sp_hires_subset 512 \
+    exp/nnet3${nnet3_affix}/pca_transform exp/nnet3${nnet3_affix}/diag_ubm
+fi
+
+if [ $stage -le 5 ]; then
+  # Train the iVector extractor.  Use all of the speed-perturbed data since iVector extractors
+  # can be sensitive to the amount of data.  The script defaults to an iVector dimension of
+  # 100.
+  echo "$0: training the iVector extractor"
+  steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj 10 \
+     data/${train_set}_sp_hires exp/nnet3${nnet3_affix}/diag_ubm \
+     exp/nnet3${nnet3_affix}/extractor || exit 1;
+fi
+
+
+if [ $stage -le 6 ]; then
+  # We extract iVectors on the speed-perturbed training data after combining
+  # short segments, which will be what we train the system on.  With
+  # --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
+  # each of these pairs as one speaker; this gives more diversity in iVectors..
+  # Note that these are extracted 'online'.
+
+  # note, we don't encode the 'max2' in the name of the ivectordir even though
+  # that's the data we extract the ivectors from, as it's still going to be
+  # valid for the non-'max2' data, the utterance list is the same.
+
+  ivectordir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $ivectordir/storage ]; then
+    utils/create_split_dir.pl /export/fs0{1,2}/$USER/kaldi-data/ivectors/cmu_cslu_kids-$(date +'%m_%d_%H_%M')/s5/$ivectordir/storage $ivectordir/storage
+  fi
+
+
+  # having a larger number of speakers is helpful for generalization, and to
+  # handle per-utterance decoding well (iVector starts at zero).
+  temp_data_root=${ivectordir}
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${train_set}_sp_hires ${temp_data_root}/${train_set}_sp_hires_max2
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 20 \
+    ${temp_data_root}/${train_set}_sp_hires_max2 \
+    exp/nnet3${nnet3_affix}/extractor $ivectordir
+
+  # Also extract iVectors for the test data, but in this case we don't need the speed
+  # perturbation (sp).
+  for data in $test_sets; do
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 20 \
+      data/${data}_hires exp/nnet3${nnet3_affix}/extractor \
+      exp/nnet3${nnet3_affix}/ivectors_${data}_hires
+  done
+fi
+
+exit 0
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/run_tdnn_lstm.sh b/egs/cmu_cslu_kids/s5/local/nnet3/run_tdnn_lstm.sh
new file mode 120000
index 00000000000..9669251c14a
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/run_tdnn_lstm.sh
@@ -0,0 +1 @@
+tuning/run_tdnn_lstm_1c.sh
\ No newline at end of file
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh
new file mode 100755
index 00000000000..de858973c98
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh
@@ -0,0 +1,223 @@
+#!/bin/bash
+
+# This is a basic TDNN+LSTM nnet3 experiment.
+
+
+# steps/info/nnet3_dir_info.pl exp/nnet3/tdnn_lstm1a_sp
+# exp/nnet3/tdnn_lstm1a_sp: num-iters=32 nj=2..2 num-params=8.4M dim=40+100->2041 combine=-0.47->-0.38 loglike:train/valid[20,31,combined]=(-0.62,-0.38,-0.37/-1.03,-1.03,-1.02) accuracy:train/valid[20,31,combined]=(0.79,0.87,0.87/0.70,0.72,0.72)
+
+# Below, comparing with the chain TDNN system.  It's a little better with the
+# small-vocab decoding.  Both systems are probably super-badly tuned, and the
+# chain system probably used too many jobs.
+#
+# local/nnet3/compare_wer.sh exp/chain/tdnn1a_sp exp/nnet3/tdnn_lstm1a_sp
+# System                tdnn1a_sp tdnn_lstm1a_sp
+#WER dev_clean_2 (tgsmall)      18.43     17.37
+#WER dev_clean_2 (tglarge)      13.15     13.43
+# Final train prob                  -0.3933
+# Final valid prob                  -0.9662
+# Final train acc                    0.8652
+# Final valid acc                    0.7206
+
+# Set -e here so that we catch if any executable fails immediately
+set -euo pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+decode_nj=10
+train_set=train_clean_5
+test_sets=dev_clean_2
+gmm=tri3b
+nnet3_affix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=1a   # affix for the TDNN directory name
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+# training chunk-options
+chunk_width=40,30,20
+chunk_left_context=40
+chunk_right_context=0
+common_egs_dir=
+xent_regularize=0.1
+
+# training options
+srand=0
+remove_egs=true
+reporting_email=
+
+#decode options
+test_online_decoding=true  # if true, it will run the last decoding stage.
+
+
+# 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
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
+# run those things.
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --nnet3-affix "$nnet3_affix" || exit 1;
+
+# Problem: We have removed the "train_" prefix of our training set in
+# the alignment directory names! Bad!
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+lang=data/lang
+label_delay=5
+
+dir=exp/nnet3${nnet3_affix}/tdnn_lstm${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $gmm_dir/graph_tgsmall/HCLG.fst $ali_dir/ali.1.gz; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 10 ]; then
+  mkdir -p $dir
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $ali_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda delay=$label_delay input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  relu-renorm-layer name=tdnn1 dim=520
+  relu-renorm-layer name=tdnn2 dim=520 input=Append(-1,0,1)
+  fast-lstmp-layer name=lstm1 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 decay-time=20 delay=-3
+  relu-renorm-layer name=tdnn3 dim=520 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn4 dim=520 input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm2 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 decay-time=20 delay=-3
+  relu-renorm-layer name=tdnn5 dim=520 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn6 dim=520 input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm3 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 decay-time=20 delay=-3
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets 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 11 ]; 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/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/train_rnn.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir=$train_ivector_dir \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=6 \
+    --trainer.deriv-truncate-margin=10 \
+    --trainer.samples-per-iter=20000 \
+    --trainer.optimization.num-jobs-initial=1 \
+    --trainer.optimization.num-jobs-final=2 \
+    --trainer.optimization.initial-effective-lrate=0.0003 \
+    --trainer.optimization.final-effective-lrate=0.00003 \
+    --trainer.optimization.shrink-value=0.99 \
+    --trainer.rnn.num-chunk-per-minibatch=128,64 \
+    --trainer.optimization.momentum=0.5 \
+    --egs.chunk-width=$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="$common_egs_dir" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --ali-dir=$ali_dir \
+    --lang=$lang \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --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 $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \
+          $gmm_dir/graph_tgsmall data/${data}_hires ${dir}/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+# Not testing the 'looped' decoding separately, because for
+# TDNN systems it would give exactly the same results as the
+# normal decoding.
+
+if $test_online_decoding && [ $stage -le 13 ]; then
+  # note: if the features change (e.g. you add pitch features), you will have to
+  # change the options of the following command line.
+  steps/online/nnet3/prepare_online_decoding.sh \
+    --mfcc-config conf/mfcc_hires.conf \
+    $lang exp/nnet3${nnet3_affix}/extractor ${dir} ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      # note: we just give it "data/${data}" as it only uses the wav.scp, the
+      # feature type does not matter.
+      steps/online/nnet3/decode.sh \
+        --nj $nspk --cmd "$decode_cmd" \
+        $gmm_dir/graph_tgsmall data/${data} ${dir}_online/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}_online/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+
+exit 0;
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh
new file mode 100755
index 00000000000..ba4ecc268df
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh
@@ -0,0 +1,228 @@
+#!/bin/bash
+
+# This is like 1a, but adding dropout.   It's definitely helpful,
+# and you can see in the objf values that the train-test difference
+# is less.
+
+
+# steps/info/nnet3_dir_info.pl exp/nnet3/tdnn_lstm1b_sp
+# exp/nnet3/tdnn_lstm1b_sp: num-iters=32 nj=2..2 num-params=8.4M dim=40+100->2041 combine=-0.71->-0.58 loglike:train/valid[20,31,combined]=(-2.78,-0.95,-0.57/-2.94,-1.31,-0.98) accuracy:train/valid[20,31,combined]=(0.48,0.75,0.81/0.45,0.67,0.71)
+
+# local/nnet3/compare_wer.sh --online exp/nnet3/tdnn_lstm1a_sp exp/nnet3/tdnn_lstm1b_sp
+# System                tdnn_lstm1a_sp tdnn_lstm1b_sp
+#WER dev_clean_2 (tgsmall)      17.67     17.01
+#             [online:]         18.06     17.26
+#WER dev_clean_2 (tglarge)      13.43     12.63
+#             [online:]         13.73     12.94
+# Final train prob        -0.3660   -0.5680
+# Final valid prob        -1.0236   -0.9771
+# Final train acc          0.8737    0.8067
+# Final valid acc          0.7222    0.7144
+
+
+
+# Set -e here so that we catch if any executable fails immediately
+set -euo pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+decode_nj=10
+train_set=train_clean_5
+test_sets=dev_clean_2
+gmm=tri3b
+nnet3_affix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=1b   # affix for the TDNN+LSTM directory name
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+# training chunk-options
+chunk_width=40,30,20
+chunk_left_context=40
+chunk_right_context=0
+common_egs_dir=
+xent_regularize=0.1
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+
+# training options
+srand=0
+remove_egs=true
+reporting_email=
+
+#decode options
+test_online_decoding=true  # if true, it will run the last decoding stage.
+
+
+# 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
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
+# run those things.
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --nnet3-affix "$nnet3_affix" || exit 1;
+
+# Problem: We have removed the "train_" prefix of our training set in
+# the alignment directory names! Bad!
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+lang=data/lang
+label_delay=5
+
+dir=exp/nnet3${nnet3_affix}/tdnn_lstm${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $gmm_dir/graph_tgsmall/HCLG.fst $ali_dir/ali.1.gz; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 10 ]; then
+  mkdir -p $dir
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $ali_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  lstm_opts="decay-time=20 delay=-3 dropout-proportion=0.0"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda delay=$label_delay input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  relu-renorm-layer name=tdnn1 dim=520
+  relu-renorm-layer name=tdnn2 dim=520 input=Append(-1,0,1)
+  fast-lstmp-layer name=lstm1 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+  relu-renorm-layer name=tdnn3 dim=520 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn4 dim=520 input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm2 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+  relu-renorm-layer name=tdnn5 dim=520 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn6 dim=520 input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm3 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets 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 11 ]; 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/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/train_rnn.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir=$train_ivector_dir \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=6 \
+    --trainer.deriv-truncate-margin=10 \
+    --trainer.samples-per-iter=20000 \
+    --trainer.optimization.num-jobs-initial=1 \
+    --trainer.optimization.num-jobs-final=2 \
+    --trainer.optimization.initial-effective-lrate=0.0003 \
+    --trainer.optimization.final-effective-lrate=0.00003 \
+    --trainer.optimization.shrink-value=0.99 \
+    --trainer.dropout-schedule="$dropout_schedule" \
+    --trainer.rnn.num-chunk-per-minibatch=128,64 \
+    --trainer.optimization.momentum=0.5 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --ali-dir=$ali_dir \
+    --lang=$lang \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --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 $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \
+          $gmm_dir/graph_tgsmall data/${data}_hires ${dir}/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+# Not testing the 'looped' decoding separately, because for
+# TDNN systems it would give exactly the same results as the
+# normal decoding.
+
+if $test_online_decoding && [ $stage -le 13 ]; then
+  # note: if the features change (e.g. you add pitch features), you will have to
+  # change the options of the following command line.
+  steps/online/nnet3/prepare_online_decoding.sh \
+    --mfcc-config conf/mfcc_hires.conf \
+    $lang exp/nnet3${nnet3_affix}/extractor ${dir} ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      # note: we just give it "data/${data}" as it only uses the wav.scp, the
+      # feature type does not matter.
+      steps/online/nnet3/decode.sh \
+        --nj $nspk --cmd "$decode_cmd" \
+        $gmm_dir/graph_tgsmall data/${data} ${dir}_online/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}_online/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+
+exit 0;
diff --git a/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1c.sh b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1c.sh
new file mode 100755
index 00000000000..74df56b0537
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/nnet3/tuning/run_tdnn_lstm_1c.sh
@@ -0,0 +1,227 @@
+#!/bin/bash
+
+# 1c is like 1b, but changing renorm to batchnorm and adding l2 regularization.
+
+# local/nnet3/compare_wer.sh --online exp/nnet3/tdnn_lstm1b_sp exp/nnet3/tdnn_lstm1c_sp
+# System                tdnn_lstm1b_sp tdnn_lstm1c_sp
+#WER dev_clean_2 (tgsmall)      17.20     16.03
+#             [online:]         17.33     16.16
+#WER dev_clean_2 (tglarge)      12.69     11.66
+#             [online:]         12.90     11.70
+# Final train prob        -0.5626   -0.6092
+# Final valid prob        -0.9453   -0.9147
+# Final train acc          0.8068    0.7999
+# Final valid acc          0.7202    0.7235
+
+# steps/info/nnet3_dir_info.pl exp/nnet3/tdnn_lstm1c_sp
+# exp/nnet3/tdnn_lstm1c_sp: num-iters=32 nj=1..2 num-params=8.4M dim=40+100->2041 combine=-0.99->-0.81 loglike:train/valid[20,31,combined]=(-1.22,-0.69,-0.61/-1.34,-1.02,-0.91) accuracy:train/valid[20,31,combined]=(0.68,0.779,0.800/0.64,0.70,0.724)
+
+
+
+
+# Set -e here so that we catch if any executable fails immediately
+set -euo pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+decode_nj=10
+train_set=train_clean_5
+test_sets=dev_clean_2
+gmm=tri3b
+nnet3_affix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=1c   # affix for the TDNN+LSTM directory name
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+# training chunk-options
+chunk_width=40,30,20
+chunk_left_context=40
+chunk_right_context=0
+common_egs_dir=
+xent_regularize=0.1
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+
+# training options
+srand=0
+remove_egs=true
+reporting_email=
+
+#decode options
+test_online_decoding=true  # if true, it will run the last decoding stage.
+
+
+# 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
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
+# run those things.
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --nnet3-affix "$nnet3_affix" || exit 1;
+
+# Problem: We have removed the "train_" prefix of our training set in
+# the alignment directory names! Bad!
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+lang=data/lang
+label_delay=5
+
+dir=exp/nnet3${nnet3_affix}/tdnn_lstm${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $gmm_dir/graph_tgsmall/HCLG.fst $ali_dir/ali.1.gz; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 10 ]; then
+  mkdir -p $dir
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $ali_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  tdnn_opts="l2-regularize=0.05"
+  lstm_opts="l2-regularize=0.01 decay-time=20 delay=-3 dropout-proportion=0.0"
+  output_opts="l2-regularize=0.01"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda delay=$label_delay input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  relu-batchnorm-layer name=tdnn1 dim=520 $tdnn_opts
+  relu-batchnorm-layer name=tdnn2 dim=520 $tdnn_opts input=Append(-1,0,1)
+  fast-lstmp-layer name=lstm1 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+  relu-batchnorm-layer name=tdnn3 dim=520 $tdnn_opts input=Append(-3,0,3)
+  relu-batchnorm-layer name=tdnn4 dim=520 $tdnn_opts input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm2 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+  relu-batchnorm-layer name=tdnn5 dim=520 $tdnn_opts input=Append(-3,0,3)
+  relu-batchnorm-layer name=tdnn6 dim=520 $tdnn_opts input=Append(-3,0,3)
+  fast-lstmp-layer name=lstm3 cell-dim=520 recurrent-projection-dim=130 non-recurrent-projection-dim=130 $lstm_opts
+
+  output-layer name=output input=lstm3 $output_opts output-delay=$label_delay dim=$num_targets 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 11 ]; 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/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/train_rnn.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir=$train_ivector_dir \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=6 \
+    --trainer.deriv-truncate-margin=10 \
+    --trainer.samples-per-iter=20000 \
+    --trainer.optimization.num-jobs-initial=1 \
+    --trainer.optimization.num-jobs-final=2 \
+    --trainer.optimization.initial-effective-lrate=0.0003 \
+    --trainer.optimization.final-effective-lrate=0.00003 \
+    --trainer.dropout-schedule="$dropout_schedule" \
+    --trainer.rnn.num-chunk-per-minibatch=128,64 \
+    --trainer.optimization.momentum=0.5 \
+    --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" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --ali-dir=$ali_dir \
+    --lang=$lang \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --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 $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \
+          $gmm_dir/graph_tgsmall data/${data}_hires ${dir}/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+# Not testing the 'looped' decoding separately, because for
+# TDNN systems it would give exactly the same results as the
+# normal decoding.
+
+if $test_online_decoding && [ $stage -le 13 ]; then
+  # note: if the features change (e.g. you add pitch features), you will have to
+  # change the options of the following command line.
+  steps/online/nnet3/prepare_online_decoding.sh \
+    --mfcc-config conf/mfcc_hires.conf \
+    $lang exp/nnet3${nnet3_affix}/extractor ${dir} ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      # note: we just give it "data/${data}" as it only uses the wav.scp, the
+      # feature type does not matter.
+      steps/online/nnet3/decode.sh \
+        --nj $nspk --cmd "$decode_cmd" \
+        $gmm_dir/graph_tgsmall data/${data} ${dir}_online/decode_tgsmall_${data} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       data/${data}_hires ${dir}_online/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+
+exit 0;
diff --git a/egs/cmu_cslu_kids/s5/local/prepare_grammar.sh b/egs/cmu_cslu_kids/s5/local/prepare_grammar.sh
new file mode 100755
index 00000000000..5b26615eaf7
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/prepare_grammar.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+#
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey).  Apache 2.0.
+
+# Takes no arguments. 
+
+
+
+tmpdir=data/local/tmp
+[ ! -f $tmpdir/G.txt ] && echo "No such file $tmpdir/G.txt" && exit 1;
+
+. ./path.sh || exit 1; # for KALDI_ROOT
+
+fstcompile --isymbols=data/lang/words.txt --osymbols=data/lang/words.txt --keep_isymbols=false \
+    --keep_osymbols=false $tmpdir/G.txt | fstarcsort --sort_type=ilabel > data/lang/G.fst || exit 1;
+
+# Checking that G is stochastic [note, it wouldn't be for an Arpa]
+fstisstochastic data/lang/G.fst || echo Error: G is not stochastic
+
+# Checking that G.fst is determinizable.
+fstdeterminize data/lang/G.fst /dev/null || echo Error determinizing G.
+
+# Checking that L_disambig.fst is determinizable.
+fstdeterminize data/lang/L_disambig.fst /dev/null || echo Error determinizing L.
+
+# Checking that disambiguated lexicon times G is determinizable
+fsttablecompose data/lang/L_disambig.fst data/lang/G.fst | \
+   fstdeterminize >/dev/null || echo Error
+
+# Checking that LG is stochastic:
+fsttablecompose data/lang/L.fst data/lang/G.fst | \
+   fstisstochastic || echo Error: LG is not stochastic.
+
+# Checking that L_disambig.G is stochastic:
+fsttablecompose data/lang/L_disambig.fst data/lang/G.fst | \
+   fstisstochastic || echo Error: LG is not stochastic.
+
+echo "Succeeded preparing grammar for CMU_kids."
diff --git a/egs/cmu_cslu_kids/s5/local/score.sh b/egs/cmu_cslu_kids/s5/local/score.sh
new file mode 100755
index 00000000000..c812199fc98
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/score.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey)
+#           2014  Guoguo Chen
+# Apache 2.0
+
+[ -f ./path.sh ] && . ./path.sh
+
+# begin configuration section.
+cmd=run.pl
+stage=0
+decode_mbr=true
+word_ins_penalty=0.0,0.5,1.0
+min_lmwt=7
+max_lmwt=17
+iter=final
+#end configuration section.
+
+[ -f ./path.sh ] && . ./path.sh
+. parse_options.sh || exit 1;
+
+if [ $# -ne 3 ]; then
+  echo "Usage: local/score.sh [--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=$3
+
+symtab=$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
+
+mkdir -p $dir/scoring/log
+
+cat $data/text | sed 's:<NOISE>::g' | sed 's:<SPOKEN_NOISE>::g' > $dir/scoring/test_filt.txt
+
+for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+  $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/best_path.LMWT.$wip.log \
+    lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \
+    lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \
+    lattice-best-path --word-symbol-table=$symtab \
+      ark:- ark,t:$dir/scoring/LMWT.$wip.tra || exit 1;
+done
+
+# Note: the double level of quoting for the sed command
+for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+  $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/score.LMWT.$wip.log \
+    cat $dir/scoring/LMWT.$wip.tra \| \
+    utils/int2sym.pl -f 2- $symtab \| sed 's:\<UNK\>::g' \| \
+    compute-wer --text --mode=present \
+    ark:$dir/scoring/test_filt.txt  ark,p:- ">&" $dir/wer_LMWT_$wip || exit 1;
+done
+
+exit 0;
diff --git a/egs/cmu_cslu_kids/s5/local/sort_result.sh b/egs/cmu_cslu_kids/s5/local/sort_result.sh
new file mode 100755
index 00000000000..aedec9dc344
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/sort_result.sh
@@ -0,0 +1,46 @@
+#! /bin/bash 
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Sorts and reports results in results/results.txt
+# for all models in exp. Expects decode directories 
+# to be named as exp/<model>/decode* or exp/chain/tdnn*/decode*
+# Should be run from egs/cmu_cslu_kids.
+
+res=${1:-"results/results.txt"}
+exp=exp
+mkdir -p results
+rm -f $res
+
+echo "Sorting results in: "
+echo "# ---------- GMM-HMM Models ----------" >> $res
+for mdl in $exp/mono* $exp/tri*; do
+    echo "  $mdl"
+    if [ -d $mdl ];then
+        for dec in $mdl/decode*;do
+            echo "    $dec"
+            if [ -d $dec ];then
+                grep WER $dec/wer* | \
+                    sort -k2 -n > $dec/WERs
+                head -n 1 $dec/WERs >> $res
+            fi
+        done
+    fi
+done
+
+echo "# ---------- DNN-HMM Models ----------" >> $res
+# DNN results
+for mdl in $exp/chain/tdnn*; do
+    echo "  $mdl"
+    for dec in $mdl/decode*; do
+        if [ -d $dec ]; then
+            echo "    $dec"
+            grep WER $dec/wer* | \
+                sort -k2 -n > $dec/WERs
+            head -n 1 $dec/WERs >> $res
+        fi
+    done
+done
+
+sed -i "s/:/    /g" $res 
diff --git a/egs/cmu_cslu_kids/s5/local/subset_dataset.sh b/egs/cmu_cslu_kids/s5/local/subset_dataset.sh
new file mode 100755
index 00000000000..050128247a4
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/subset_dataset.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+
+# Copyright 2017  Luminar Technologies, Inc. (author: Daniel Galvez)
+# Apache 2.0
+
+# The following commands were used to generate the mini_librispeech dataset:
+#
+# Note that data generation is random. This could be fixed by
+# providing a seed argument to the shuf program.
+
+if [ "$#" -ne 3 ]; then
+  echo "Usage: $0 <src-dir> <dst-dir> <num-hours>"
+  echo "e.g.: $0 /export/a05/dgalvez/LibriSpeech/train-clean-100 \\
+                 /export/a05/dgalvez/LibriSpeech/train-clean-5 5"
+  exit 1
+fi
+
+src_dir=$1
+dest_dir=$2
+dest_num_hours=$3
+
+src=$(basename $src_dir)
+dest=$(basename $dest_dir)
+librispeech_dir=$(dirname $src_dir)
+
+# TODO: Possibly improve this to ensure gender balance and speaker
+# balance.
+# TODO: Use actual time values instead of assuming that to make sure we get $dest_num_hours of data
+src_num_hours=$(grep "$src" $librispeech_dir/CHAPTERS.TXT | awk -F'|' '{ print $3 }' | \
+python -c '
+from __future__ import print_function
+from sys import stdin
+minutes_str = stdin.read().split()
+print(int(round(sum([float(minutes) for minutes in minutes_str]) / 60.0)))')
+src_num_chapters=$(grep "$src" $librispeech_dir/CHAPTERS.TXT | \
+                      awk -F'|' '{ print $1 }' | sort -u | wc -l)
+mkdir -p data/subset_tmp
+grep "$src" $librispeech_dir/CHAPTERS.TXT | \
+  awk -F'|' '{ print $1 }' | \
+  shuf -n $(((dest_num_hours * src_num_chapters) / src_num_hours)) > \
+       data/subset_tmp/${dest}_chapter_id_list.txt
+
+while read -r chapter_id || [[ -n "$chapter_id" ]]; do
+  chapter_dir=$(find $src_dir/ -mindepth 2 -name "$chapter_id" -type d)
+  speaker_id=$(basename $(dirname $chapter_dir))
+  mkdir -p $dest_dir/$speaker_id/
+  cp -r $chapter_dir $dest_dir/$speaker_id/
+done  < data/subset_tmp/${dest}_chapter_id_list.txt
diff --git a/egs/cmu_cslu_kids/s5/local/sum_duration.py b/egs/cmu_cslu_kids/s5/local/sum_duration.py
new file mode 100644
index 00000000000..0af7ba62151
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/sum_duration.py
@@ -0,0 +1,15 @@
+# Sum duration obtained by using 
+# utils/data/get_utt2dur.sh
+
+import sys
+file = sys.argv[1]
+sum = 0
+with open(file, 'r') as fp:
+    line = fp.readline()
+    while(line):
+        toks = line.strip().split()
+        sum += float(toks[1])
+        line = fp.readline()
+fp.close()
+h=sum/3600
+sys.stdout.write("%f hour data.\n"%h)
diff --git a/egs/cmu_cslu_kids/s5/local/train_lms.sh b/egs/cmu_cslu_kids/s5/local/train_lms.sh
new file mode 100755
index 00000000000..0807210be18
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/train_lms.sh
@@ -0,0 +1,217 @@
+#!/bin/bash
+
+# This script trains LMs on the WSJ LM-training data.
+# It requires that you have already run wsj_extend_dict.sh,
+# to get the larger-size dictionary including all of CMUdict
+# plus any OOVs and possible acronyms that we could easily 
+# derive pronunciations for.
+
+dict_suffix=
+
+echo "$0 $@"  # Print the command line for logging
+. utils/parse_options.sh || exit 1;
+
+dir=data/local/local_lm
+srcdir=data/local/dict${dict_suffix}_larger
+mkdir -p $dir
+. ./path.sh || exit 1; # for KALDI_ROOT
+export PATH=$KALDI_ROOT/tools/kaldi_lm:$PATH
+( # First make sure the kaldi_lm toolkit is installed.
+ cd $KALDI_ROOT/tools || exit 1;
+ if [ -d kaldi_lm ]; then
+   echo Not installing the kaldi_lm toolkit since it is already there.
+ else
+   echo Downloading and installing the kaldi_lm tools
+   if [ ! -f kaldi_lm.tar.gz ]; then
+     wget http://www.danielpovey.com/files/kaldi/kaldi_lm.tar.gz || exit 1;
+   fi
+   tar -xvzf kaldi_lm.tar.gz || exit 1;
+   cd kaldi_lm
+   make || exit 1;
+   echo Done making the kaldi_lm tools
+ fi
+) || exit 1;
+
+
+
+if [ ! -f $srcdir/cleaned.gz -o ! -f $srcdir/lexicon.txt ]; then
+  echo "Expecting files $srcdir/cleaned.gz and $srcdir/lexicon.txt to exist";
+  echo "You need to run local/wsj_extend_dict.sh before running this script."
+  exit 1;
+fi
+
+# Get a wordlist-- keep everything but silence, which should not appear in
+# the LM.
+awk '{print $1}' $srcdir/lexicon.txt | grep -v -w '!SIL' > $dir/wordlist.txt
+
+# Get training data with OOV words (w.r.t. our current vocab) replaced with <UNK>.
+echo "Getting training data with OOV words replaced with <UNK> (train_nounk.gz)" 
+gunzip -c $srcdir/cleaned.gz | awk -v w=$dir/wordlist.txt \
+  'BEGIN{while((getline<w)>0) v[$1]=1;}
+  {for (i=1;i<=NF;i++) if ($i in v) printf $i" ";else printf "<UNK> ";print ""}'|sed 's/ $//g' \
+  | gzip -c > $dir/train_nounk.gz
+
+# Get unigram counts (without bos/eos, but this doens't matter here, it's
+# only to get the word-map, which treats them specially & doesn't need their
+# counts).
+# Add a 1-count for each word in word-list by including that in the data,
+# so all words appear.
+gunzip -c $dir/train_nounk.gz | cat - $dir/wordlist.txt | \
+  awk '{ for(x=1;x<=NF;x++) count[$x]++; } END{for(w in count){print count[w], w;}}' | \
+ sort -nr > $dir/unigram.counts
+
+# Get "mapped" words-- a character encoding of the words that makes the common words very short.
+cat $dir/unigram.counts  | awk '{print $2}' | get_word_map.pl "<s>" "</s>" "<UNK>" > $dir/word_map
+
+gunzip -c $dir/train_nounk.gz | awk -v wmap=$dir/word_map 'BEGIN{while((getline<wmap)>0)map[$1]=$2;}
+  { for(n=1;n<=NF;n++) { printf map[$n]; if(n<NF){ printf " "; } else { print ""; }}}' | gzip -c >$dir/train.gz
+
+# To save disk space, remove the un-mapped training data.  We could
+# easily generate it again if needed.
+rm $dir/train_nounk.gz 
+
+train_lm.sh --arpa --lmtype 3gram-mincount $dir
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 141.444826
+# 7.8 million N-grams.
+
+prune_lm.sh --arpa 6.0 $dir/3gram-mincount/
+# 1.45 million N-grams.
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 165.394139
+
+train_lm.sh --arpa --lmtype 4gram-mincount $dir
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 126.734180
+# 10.3 million N-grams.
+
+prune_lm.sh --arpa 7.0 $dir/4gram-mincount
+# 1.50 million N-grams
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 155.663757
+
+
+exit 0
+
+### Below here, this script is showing various commands that 
+## were run during LM tuning.
+
+train_lm.sh --arpa --lmtype 3gram-mincount $dir
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 141.444826
+# 7.8 million N-grams.
+
+prune_lm.sh --arpa 3.0 $dir/3gram-mincount/
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 156.408740
+# 2.5 million N-grams.
+
+prune_lm.sh --arpa 6.0 $dir/3gram-mincount/
+# 1.45 million N-grams.
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 165.394139
+
+train_lm.sh --arpa --lmtype 4gram-mincount $dir
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 126.734180
+# 10.3 million N-grams.
+
+prune_lm.sh --arpa 3.0 $dir/4gram-mincount
+#Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 143.206294
+# 2.6 million N-grams.
+
+prune_lm.sh --arpa 4.0 $dir/4gram-mincount
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 146.927717
+# 2.15 million N-grams.
+
+prune_lm.sh --arpa 5.0 $dir/4gram-mincount
+# 1.86 million N-grams
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 150.162023
+
+prune_lm.sh --arpa 7.0 $dir/4gram-mincount
+# 1.50 million N-grams
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 155.663757
+
+train_lm.sh --arpa --lmtype 3gram $dir
+# Perplexity over 228518.000000 words (excluding 478.000000 OOVs) is 135.692866
+# 20.0 million N-grams
+
+! which ngram-count  \
+  && echo "SRILM tools not installed so not doing the comparison" && exit 1;
+
+#################
+# You could finish the script here if you wanted.
+# Below is to show how to do baselines with SRILM.
+#  You'd have to install the SRILM toolkit first.
+
+heldout_sent=10000 # Don't change this if you want result to be comparable with
+    # kaldi_lm results
+sdir=$dir/srilm # in case we want to use SRILM to double-check perplexities.
+mkdir -p $sdir
+gunzip -c $srcdir/cleaned.gz | head -$heldout_sent > $sdir/cleaned.heldout
+gunzip -c $srcdir/cleaned.gz | tail -n +$heldout_sent > $sdir/cleaned.train
+(echo "<s>"; echo "</s>" ) | cat - $dir/wordlist.txt > $sdir/wordlist.final.s
+
+# 3-gram:
+ngram-count -text $sdir/cleaned.train -order 3 -limit-vocab -vocab $sdir/wordlist.final.s -unk \
+  -map-unk "<UNK>" -kndiscount -interpolate -lm $sdir/srilm.o3g.kn.gz
+ngram -lm $sdir/srilm.o3g.kn.gz -ppl $sdir/cleaned.heldout # consider -debug 2
+#file data/local/local_lm/srilm/cleaned.heldout: 10000 sentences, 218996 words, 478 OOVs
+#0 zeroprobs, logprob= -491456 ppl= 141.457 ppl1= 177.437
+
+# Trying 4-gram:
+ngram-count -text $sdir/cleaned.train -order 4 -limit-vocab -vocab $sdir/wordlist.final.s -unk \
+  -map-unk "<UNK>" -kndiscount -interpolate -lm $sdir/srilm.o4g.kn.gz
+ngram -order 4 -lm $sdir/srilm.o4g.kn.gz -ppl $sdir/cleaned.heldout 
+#file data/local/local_lm/srilm/cleaned.heldout: 10000 sentences, 218996 words, 478 OOVs
+#0 zeroprobs, logprob= -480939 ppl= 127.233 ppl1= 158.822
+
+#3-gram with pruning:
+ngram-count -text $sdir/cleaned.train -order 3 -limit-vocab -vocab $sdir/wordlist.final.s -unk \
+  -prune 0.0000001 -map-unk "<UNK>" -kndiscount -interpolate -lm $sdir/srilm.o3g.pr7.kn.gz
+ngram -lm $sdir/srilm.o3g.pr7.kn.gz -ppl $sdir/cleaned.heldout 
+#file data/local/local_lm/srilm/cleaned.heldout: 10000 sentences, 218996 words, 478 OOVs
+#0 zeroprobs, logprob= -510828 ppl= 171.947 ppl1= 217.616
+# Around 2.25M N-grams.
+# Note: this is closest to the experiment done with "prune_lm.sh --arpa 3.0 $dir/3gram-mincount/"
+# above, which gave 2.5 million N-grams and a perplexity of 156.
+
+# Note: all SRILM experiments above fully discount all singleton 3 and 4-grams.
+# You can use -gt3min=0 and -gt4min=0 to stop this (this will be comparable to
+# the kaldi_lm experiments above without "-mincount".
+
+##  From here is how to train with
+# IRSTLM.  This is not really working at the moment.
+
+if [ -z $IRSTLM ] ; then
+  export IRSTLM=$KALDI_ROOT/tools/irstlm/
+fi
+export PATH=${PATH}:$IRSTLM/bin
+if ! command -v prune-lm >/dev/null 2>&1 ; then
+  echo "$0: Error: the IRSTLM is not available or compiled" >&2
+  echo "$0: Error: We used to install it by default, but." >&2
+  echo "$0: Error: this is no longer the case." >&2
+  echo "$0: Error: To install it, go to $KALDI_ROOT/tools" >&2
+  echo "$0: Error: and run extras/install_irstlm.sh" >&2
+  exit 1
+fi
+
+idir=$dir/irstlm
+mkdir $idir
+gunzip -c $srcdir/cleaned.gz | tail -n +$heldout_sent | add-start-end.sh | \
+  gzip -c > $idir/train.gz
+
+dict -i=WSJ.cleaned.irstlm.txt -o=dico -f=y -sort=no
+ cat dico | gawk 'BEGIN{while (getline<"vocab.20k.nooov") v[$1]=1; print "DICTIONARY 0 "length(v);}FNR>1{if ($1 in v)\
+{print $0;}}' > vocab.irstlm.20k
+
+
+build-lm.sh -i "gunzip -c $idir/train.gz" -o $idir/lm_3gram.gz  -p yes \
+  -n 3 -s improved-kneser-ney -b yes
+# Testing perplexity with SRILM tools:
+ngram -lm $idir/lm_3gram.gz  -ppl $sdir/cleaned.heldout 
+#data/local/local_lm/irstlm/lm_3gram.gz: line 162049: warning: non-zero probability for <unk> in closed-vocabulary LM
+#file data/local/local_lm/srilm/cleaned.heldout: 10000 sentences, 218996 words, 0 OOVs
+#0 zeroprobs, logprob= -513670 ppl= 175.041 ppl1= 221.599
+
+# Perplexity is very bad (should be ~141, since we used -p option,
+# not 175),
+# but adding -debug 3 to the command line shows that
+# the IRSTLM LM does not seem to sum to one properly, so it seems that
+# it produces an LM that isn't interpretable in the normal way as an ARPA
+# LM.
+
+
+
diff --git a/egs/cmu_cslu_kids/s5/local/vtln.sh b/egs/cmu_cslu_kids/s5/local/vtln.sh
new file mode 100755
index 00000000000..0ca179ce89f
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/local/vtln.sh
@@ -0,0 +1,61 @@
+#!/bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Run VTLN. This will be run if the vtln option 
+# is set to be true in run.sh.
+
+set -eu
+stage=0
+featdir=mfcc/vtln
+data=data
+mdl=exp/tri3
+mdl_vtln=${mdl}_vtln
+vtln_lda=exp/tri4
+vtln_sat=exp/tri5
+
+. ./cmd.sh
+. ./utils/parse_options.sh
+
+mkdir -p $featdir
+
+steps/train_lvtln.sh --cmd "$train_cmd" 1800 9000 $data/train $data/lang $mdl $mdl_vtln
+
+if [ $stage -le 0 ]; then
+    mkdir -p $data/train_vtln
+    cp $data/train/* $data/train_vtln || true
+    cp $mdl_vtln/final.warp $data/train_vtln/spk2warp
+    steps/make_mfcc.sh --nj 8 --cmd "$train_cmd" $data/train_vtln exp/make_mfcc/train_vtln $featdir  
+    steps/compute_cmvn_stats.sh $data/train_vtln exp/make_mfcc/train_vtln $featdir  
+fi
+
+if [ $stage -le 1 ]; then 
+    utils/mkgraph.sh $data/lang_test_tgmed $mdl_vtln $mdl_vtln/graph
+    steps/decode_lvtln.sh --config conf/decode.config --nj 20 --cmd "$decode_cmd" \
+        $mdl_vtln/graph $data/test $mdl_vtln/decode
+fi 
+
+if [ $stage -le 2 ]; then
+    mkdir -p $data/test_vtln
+    cp $data/test/* $data/test_vtln || true
+    cp $mdl_vtln/decode/final.warp $data/test_vtln/spk2warp
+    steps/make_mfcc.sh --nj 8 --cmd "$train_cmd" $data/test_vtln exp/make_mfcc/test_vtln $featdir  
+    steps/compute_cmvn_stats.sh $data/test_vtln exp/make_mfcc/test_vtln $featdir  
+fi 
+
+if [ $stage -le 3 ]; then
+    steps/train_lda_mllt.sh --cmd "$train_cmd" --splice-opts "--left-context=3 --right-context=3" 1800 9000 \
+        $data/train_vtln $data/lang $mdl_vtln $vtln_lda
+    utils/mkgraph.sh $data/lang_test_tgmed $vtln_lda $vtln_lda/graph
+    echo "$mdl_vtln + lda + mllt" > $vtln_lda/mcodel_discription
+    steps/decode.sh --config conf/decode.config --nj 20 --cmd "$decode_cmd" \
+        $vtln_lda/graph $data/test_vtln $vtln_lda/decode
+fi
+
+if [ $stage -le 4 ]; then
+    steps/train_sat.sh 1800 9000 $data/train_vtln $data/lang $vtln_lda $vtln_sat
+    utils/mkgraph.sh $data/lang_test_tgmed $vtln_sat $vtln_sat/graph 
+    steps/decode_fmllr.sh --config conf/decode.config --nj 20 --cmd "$decode_cmd" $vtln_sat/graph $data/test_vtln $vtln_sat/decode 
+    echo  "$mdl_vtln + lda + mllt + SAT" > $vtln_sat/model_discription
+fi
diff --git a/egs/cmu_cslu_kids/s5/path.sh b/egs/cmu_cslu_kids/s5/path.sh
new file mode 100755
index 00000000000..2d17b17a84a
--- /dev/null
+++ b/egs/cmu_cslu_kids/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/cmu_cslu_kids/s5/run.sh b/egs/cmu_cslu_kids/s5/run.sh
new file mode 100755
index 00000000000..43ae1ea9426
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/run.sh
@@ -0,0 +1,177 @@
+#! /bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+set -eo
+
+stage=0
+cmu_kids=               # path to cmu_kids corpus
+cslu_kids=              # path to cslu_kids corpus
+lm_src=                 # path of existing librispeech lm 
+extra_features=false    # Extra features for GMM model (MMI, boosting and MPE)
+vtln=false              # Optional, run VLTN on gmm and tdnnf models if set true 
+email=                  # Reporting email for tdnn-f training
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+lm_url=www.openslr.org/resources/11
+mkdir -p data
+mkdir -p data/local
+
+# Prepare data
+if [ $stage -le 0 ]; then
+  # Make soft link to the corpora
+  if [ ! -e cmu_kids ]; then
+    ln -sf $cmu_kids cmu_kids
+  fi
+  if [ ! -e cslu ]; then
+    ln -sf $cslu_kids cslu
+  fi
+  
+  # Make softlink to lm, if lm_src provided
+  if [ ! -z "$lm_src" ] && [ ! -e data/local/lm ] ; then
+    ln -sf $lm_src data/local/lm
+  fi
+  
+  # Remove old data dirs
+  rm -rf data/data_cmu
+  rm -rf data/data_cslu
+
+  # Data Prep
+  ./local/cmu_prepare_data.sh --corpus cmu_kids/kids --data data/data_cmu
+  ./local/cslu_prepare_data.sh --corpus cslu --data data/data_cslu 
+fi
+
+# Combine data
+if [ $stage -le 1 ]; then
+   mkdir -p data/train
+   mkdir -p data/test
+   rm -rf data/train/*
+   rm -rf data/test/*
+   ./utils/combine_data.sh data/train data/data_cmu/train data/data_cslu/train
+   ./utils/combine_data.sh data/test data/data_cmu/test data/data_cslu/test
+fi
+
+# LM, WFST Preparation
+if [ $stage -le 2 ]; then
+  if [ ! -d data/local/dict ]; then
+      ./local/download_cmu_dict.sh
+  fi
+
+  if [ ! -e data/local/lm ]; then
+    echo "lm_src not provided. Downloading lm from openslr."
+    ./local/download_lm.sh $lm_url data/local/lm
+  fi
+
+  utils/prepare_lang.sh data/local/dict "<UNK>"  data/local/lang data/lang
+  local/format_lms.sh --src_dir data/lang  data/local/lm 
+   
+  # Create ConstArpaLm format language model for full 3-gram and 4-gram LMs
+  utils/build_const_arpa_lm.sh data/local/lm/lm_tglarge.arpa.gz data/lang data/lang_test_tglarge
+  utils/build_const_arpa_lm.sh data/local/lm/lm_fglarge.arpa.gz data/lang data/lang_test_fglarge 
+fi
+
+# Make MFCC features
+if [ $stage -le 3 ]; then
+  mkdir -p mfcc
+  mkdir -p exp
+  steps/make_mfcc.sh --nj 40 --cmd "$train_cmd" data/test exp/make_feat/test mfcc
+  steps/compute_cmvn_stats.sh data/test exp/make_feat/test mfcc
+  steps/make_mfcc.sh --nj 40 --cmd "$train_cmd" data/train exp/make_feat/train mfcc 
+  steps/compute_cmvn_stats.sh data/train exp/make_feat/train mfcc
+fi
+
+# Mono-phone 
+if [ $stage -le 4 ]; then
+  # Train
+  steps/train_mono.sh --nj 40 --cmd "$train_cmd" data/train data/lang exp/mono 
+  #Decode
+  utils/mkgraph.sh data/lang_test_tgsmall exp/mono exp/mono/graph
+  steps/decode.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/mono/graph data/test exp/mono/decode
+  #Align
+  steps/align_si.sh --nj 20 --cmd "$train_cmd" data/train data/lang exp/mono exp/mono_ali
+fi
+
+# Tri1 [Vanilla tri phone model]
+if [ $stage -le 5 ]; then
+  # Train
+  steps/train_deltas.sh --cmd "$train_cmd" 1800 9000 data/train data/lang exp/mono_ali exp/tri1
+  # Decode 
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri1 exp/tri1/graph 
+  steps/decode.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri1/graph data/test exp/tri1/decode
+  # Align - make graph - decode again   
+  steps/align_si.sh --nj 20 --cmd "queue.pl" --use-graphs true data/train data/lang_test_tgmed exp/tri1 exp/tri1_ali
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri1_ali exp/tri1_ali/graph
+  steps/decode.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri1_ali/graph data/test exp/tri1_ali/decode
+fi
+
+# Add LDA and MLLT
+if [ $stage -le 6 ]; then
+  # Train
+  steps/train_lda_mllt.sh --cmd "$train_cmd" --splice-opts "--left-context=3 --right-context=3" 1800 9000 data/train data/lang exp/tri1_ali exp/tri2
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri2 exp/tri2/graph
+  # Decode
+  steps/decode.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2/decode
+  # Align - make graph - dcode again 
+  steps/align_si.sh --nj 20 --cmd "$train_cmd" --use-graphs true data/train data/lang_test_tgmed exp/tri2 exp/tri2_ali
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri2_ali exp/tri2_ali/graph
+  steps/decode_fmllr.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri2_ali/graph data/test exp/tri2_ali/decode
+fi 
+
+# Add other features
+if [ $stage -le 7 ]; then
+  if [ $extra_features = true ]; then
+    # Add MMI
+    steps/make_denlats.sh --nj 20 --cmd "$train_cmd" data/train data/lang exp/tri2 exp/tri2_denlats
+    steps/train_mmi.sh data/train data/lang exp/tri2_ali exp/tri2_denlats exp/tri2_mmi
+    steps/decode.sh --config conf/decode.config --iter 4 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mmi/decode_it4
+    steps/decode.sh --config conf/decode.config --iter 3 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mmi/decode_it3
+    
+    # Add Boosting 
+    steps/train_mmi.sh --boost 0.05 data/train data/lang exp/tri2_ali exp/tri2_denlats exp/tri2_mmi_b0.05
+    steps/decode.sh --config conf/decode.config --iter 4 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mmi_b0.05/decode_it4
+    steps/decode.sh --config conf/decode.config --iter 3 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mmi_b0.05/decode_it3
+    
+    # Add MPE 
+    steps/train_mpe.sh data/train data/lang exp/tri2_ali exp/tri2_denlats exp/tri2_mpe
+    steps/decode.sh --config conf/decode.config --iter 4 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mpe/decode_it4
+    steps/decode.sh --config conf/decode.config --iter 3 --nj 20 --cmd "$decode_cmd" exp/tri2/graph data/test exp/tri2_mpe/decode_it3
+  fi
+fi
+
+# Add SAT
+if [ $stage -le 8 ]; then 
+  # Do LDA+MLLT+SAT, and decode.
+  steps/train_sat.sh 1800 9000 data/train data/lang exp/tri2_ali exp/tri3
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri3 exp/tri3/graph
+  steps/decode_fmllr.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri3/graph data/test exp/tri3/decode
+fi
+
+if [ $stage -le 9 ]; then
+  # Align all data with LDA+MLLT+SAT system (tri3)
+  steps/align_fmllr.sh --nj 20 --cmd "$train_cmd" --use-graphs true data/train data/lang_test_tgmed exp/tri3 exp/tri3_ali
+  utils/mkgraph.sh data/lang_test_tgmed exp/tri3_ali exp/tri3_ali/graph   
+  steps/decode_fmllr.sh --config conf/decode.config --nj 40 --cmd "$decode_cmd" exp/tri3_ali/graph data/test exp/tri3_ali/decode
+fi
+
+if [ $stage -le 10 ]; then 
+    # Uncomment reporting email option to get training progress updates by email
+  ./local/chain/run_tdnnf.sh --train_set train \
+      --test_sets test --gmm tri3  # --reporting_email $email 
+fi
+
+
+# Optional VTLN. Run if vtln is set to true
+if [ $stage -le 11 ]; then
+  if [ $vtln = true ]; then
+    ./local/vtln.sh
+    ./local/chain/run_tdnnf.sh --nnet3_affix vtln --train_set train_vtln \
+        --test_sets test_vtln --gmm tri5 # --reporting_email $email
+  fi
+fi
+
+# Collect and resport WER results for all models
+./local/sort_result.sh
diff --git a/egs/cmu_cslu_kids/s5/steps b/egs/cmu_cslu_kids/s5/steps
new file mode 120000
index 00000000000..1b186770dd1
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps/
\ No newline at end of file
diff --git a/egs/cmu_cslu_kids/s5/utils b/egs/cmu_cslu_kids/s5/utils
new file mode 120000
index 00000000000..a3279dc8679
--- /dev/null
+++ b/egs/cmu_cslu_kids/s5/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils/
\ No newline at end of file