diff --git a/egs/librispeech/s5/local/data_prep.sh b/egs/librispeech/s5/local/data_prep.sh
index 5a264a07464..dea93525e28 100755
--- a/egs/librispeech/s5/local/data_prep.sh
+++ b/egs/librispeech/s5/local/data_prep.sh
@@ -33,7 +33,7 @@ utt2spk=$dst/utt2spk; [[ -f "$utt2spk" ]] && rm $utt2spk
 spk2gender=$dst/spk2gender; [[ -f $spk2gender ]] && rm $spk2gender
 utt2dur=$dst/utt2dur; [[ -f "$utt2dur" ]] && rm $utt2dur
 
-for reader_dir in $(find $src -mindepth 1 -maxdepth 1 -type d | sort); do
+for reader_dir in $(find -L $src -mindepth 1 -maxdepth 1 -type d | sort); do
   reader=$(basename $reader_dir)
   if ! [ $reader -eq $reader ]; then  # not integer.
     echo "$0: unexpected subdirectory name $reader"
@@ -53,7 +53,7 @@ for reader_dir in $(find $src -mindepth 1 -maxdepth 1 -type d | sort); do
       exit 1;
     fi
 
-    find $chapter_dir/ -iname "*.flac" | sort | xargs -I% basename % .flac | \
+    find -L $chapter_dir/ -iname "*.flac" | sort | xargs -I% basename % .flac | \
       awk -v "dir=$chapter_dir" '{printf "%s flac -c -d -s %s/%s.flac |\n", $0, dir, $0}' >>$wav_scp|| exit 1
 
     chapter_trans=$chapter_dir/${reader}-${chapter}.trans.txt
diff --git a/egs/mini_librispeech/s5/RESULTS b/egs/mini_librispeech/s5/RESULTS
new file mode 100755
index 00000000000..463c059bdbb
--- /dev/null
+++ b/egs/mini_librispeech/s5/RESULTS
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+for x in exp/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
+
+for x in exp/chain/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
+exit 0
+
+# Results on on dev_clean_2
+%WER 49.18 [ 9903 / 20138, 439 ins, 2332 del, 7132 sub ] exp/mono/decode_nosp_tgsmall_dev_clean_2/wer_8_0.0
+%WER 20.42 [ 4113 / 20138, 469 ins, 545 del, 3099 sub ] exp/tri1/decode_nosp_tglarge_dev_clean_2/wer_14_0.0
+%WER 24.56 [ 4945 / 20138, 447 ins, 792 del, 3706 sub ] exp/tri1/decode_nosp_tgmed_dev_clean_2/wer_14_0.0
+%WER 27.37 [ 5512 / 20138, 425 ins, 969 del, 4118 sub ] exp/tri1/decode_nosp_tgsmall_dev_clean_2/wer_14_0.0
+%WER 18.59 [ 3743 / 20138, 435 ins, 517 del, 2791 sub ] exp/tri2b/decode_nosp_tglarge_dev_clean_2/wer_15_0.0
+%WER 22.06 [ 4443 / 20138, 400 ins, 748 del, 3295 sub ] exp/tri2b/decode_nosp_tgmed_dev_clean_2/wer_15_0.0
+%WER 24.32 [ 4898 / 20138, 413 ins, 899 del, 3586 sub ] exp/tri2b/decode_nosp_tgsmall_dev_clean_2/wer_15_0.0
+%WER 13.45 [ 2708 / 20138, 358 ins, 330 del, 2020 sub ] exp/tri3b/decode_nosp_tglarge_dev_clean_2/wer_17_0.0
+%WER 16.25 [ 3273 / 20138, 332 ins, 485 del, 2456 sub ] exp/tri3b/decode_nosp_tgmed_dev_clean_2/wer_16_0.0
+%WER 18.10 [ 3645 / 20138, 332 ins, 603 del, 2710 sub ] exp/tri3b/decode_nosp_tgsmall_dev_clean_2/wer_16_0.0
+
+
+%WER 18.58 [ 3742 / 20138, 366 ins, 763 del, 2613 sub ] exp/chain/tdnn1a_sp/decode_tgsmall_dev_clean_2/wer_10_0.0
+%WER 13.35 [ 2689 / 20138, 318 ins, 491 del, 1880 sub ] exp/chain/tdnn1a_sp/decode_tglarge_dev_clean_2/wer_9_0.5
diff --git a/egs/mini_librispeech/s5/cmd.sh b/egs/mini_librispeech/s5/cmd.sh
new file mode 100644
index 00000000000..71dd849a93b
--- /dev/null
+++ b/egs/mini_librispeech/s5/cmd.sh
@@ -0,0 +1,15 @@
+# you can change cmd.sh depending on what type of queue you are using.
+# If you have no queueing system and want to run on a local machine, you
+# can change all instances 'queue.pl' to run.pl (but be careful and run
+# commands one by one: most recipes will exhaust the memory on your
+# machine).  queue.pl works with GridEngine (qsub).  slurm.pl works
+# with slurm.  Different queues are configured differently, with different
+# queue names and different ways of specifying things like memory;
+# to account for these differences you can create and edit the file
+# conf/queue.conf to match your queue's configuration.  Search for
+# conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
+# or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
+
+export train_cmd="queue.pl --mem 2G"
+export decode_cmd="queue.pl --mem 4G"
+export mkgraph_cmd="queue.pl --mem 8G"
diff --git a/egs/mini_librispeech/s5/conf/mfcc.conf b/egs/mini_librispeech/s5/conf/mfcc.conf
new file mode 100644
index 00000000000..7361509099f
--- /dev/null
+++ b/egs/mini_librispeech/s5/conf/mfcc.conf
@@ -0,0 +1 @@
+--use-energy=false   # only non-default option.
diff --git a/egs/mini_librispeech/s5/conf/mfcc_hires.conf b/egs/mini_librispeech/s5/conf/mfcc_hires.conf
new file mode 100644
index 00000000000..434834a6725
--- /dev/null
+++ b/egs/mini_librispeech/s5/conf/mfcc_hires.conf
@@ -0,0 +1,10 @@
+# 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) 
diff --git a/egs/mini_librispeech/s5/conf/online_cmvn.conf b/egs/mini_librispeech/s5/conf/online_cmvn.conf
new file mode 100644
index 00000000000..7748a4a4dd3
--- /dev/null
+++ b/egs/mini_librispeech/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/mini_librispeech/s5/local/chain/compare_wer.sh b/egs/mini_librispeech/s5/local/chain/compare_wer.sh
new file mode 100755
index 00000000000..cd6be14ed88
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/chain/compare_wer.sh
@@ -0,0 +1,131 @@
+#!/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
diff --git a/egs/mini_librispeech/s5/local/chain/run_tdnn.sh b/egs/mini_librispeech/s5/local/chain/run_tdnn.sh
new file mode 120000
index 00000000000..34499362831
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/chain/run_tdnn.sh
@@ -0,0 +1 @@
+tuning/run_tdnn_1a.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1a.sh b/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1a.sh
new file mode 100755
index 00000000000..642c20ec191
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1a.sh
@@ -0,0 +1,298 @@
+#!/bin/bash
+
+# This is a basic TDNN experiment.
+
+# steps/info/chain_dir_info.pl exp/chain/tdnn1a_sp
+# exp/chain/tdnn1a_sp: num-iters=6 nj=2..5 num-params=7.0M dim=40+100->2309 combine=-0.072->-0.069 xent:train/valid[3,5,final]=(-2.10,-1.62,-1.48/-2.26,-1.85,-1.77) logprob:train/valid[3,5,final]=(-0.096,-0.069,-0.060/-0.124,-0.107,-0.104)
+
+# local/chain/compare_wer.sh --online exp/chain/tdnn1a_sp
+# System                tdnn1a_sp
+#WER dev_clean_2 (tgsmall)      18.58
+#             [online:]         18.49
+#WER dev_clean_2 (tglarge)      13.35
+#             [online:]         13.47
+# Final train prob        -0.0596
+# Final valid prob        -0.1036
+# Final train prob (xent)   -1.4843
+# Final valid prob (xent)   -1.7723
+
+# 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
+tree_affix=
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+# training chunk-options
+chunk_width=140,100,160
+# we don't need extra left/right context for TDNN systems.
+chunk_left_context=0
+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_${train_set}_sp
+tree_dir=exp/chain${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 10 ]; 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 [ $stage -le 11 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj 75 --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 12 ]; 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 13 ]; 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)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-renorm-layer name=tdnn1 dim=512
+  relu-renorm-layer name=tdnn2 dim=512 input=Append(-1,0,1)
+  relu-renorm-layer name=tdnn3 dim=512 input=Append(-1,0,1)
+  relu-renorm-layer name=tdnn4 dim=512 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn5 dim=512 input=Append(-3,0,3)
+  relu-renorm-layer name=tdnn6 dim=512 input=Append(-6,-3,0)
+
+  ## adding the layers for chain branch
+  relu-renorm-layer name=prefinal-chain dim=512 target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-renorm-layer name=prefinal-xent input=tdnn6 dim=512 target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+
+if [ $stage -le 14 ]; 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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=3000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=5 \
+    --trainer.optimization.initial-effective-lrate=0.001 \
+    --trainer.optimization.final-effective-lrate=0.0001 \
+    --trainer.optimization.shrink-value=1.0 \
+    --trainer.num-chunk-per-minibatch=256,128,64 \
+    --trainer.optimization.momentum=0.0 \
+    --egs.chunk-width=$chunk_width \
+    --egs.chunk-left-context=$chunk_left_context \
+    --egs.chunk-right-context=$chunk_right_context \
+    --egs.chunk-left-context-initial=0 \
+    --egs.chunk-right-context-final=0 \
+    --egs.dir="$common_egs_dir" \
+    --egs.opts="--frames-overlap-per-eg 0" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --tree-dir=$tree_dir \
+    --lat-dir=$lat_dir \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 15 ]; 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 [ $stage -le 16 ]; then
+  frames_per_chunk=$(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 \
+          --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 \
+          $tree_dir/graph_tgsmall data/${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
+
+# 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 17 ]; 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 \
+        --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $nspk --cmd "$decode_cmd" \
+        $tree_dir/graph_tgsmall data/${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/mini_librispeech/s5/local/data_prep.sh b/egs/mini_librispeech/s5/local/data_prep.sh
new file mode 120000
index 00000000000..9d3777b7497
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/data_prep.sh
@@ -0,0 +1 @@
+../../../librispeech/s5/local/data_prep.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/download_lm.sh b/egs/mini_librispeech/s5/local/download_lm.sh
new file mode 120000
index 00000000000..d8bae7f4d32
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/download_lm.sh
@@ -0,0 +1 @@
+../../../librispeech/s5/local/download_lm.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/format_lms.sh b/egs/mini_librispeech/s5/local/format_lms.sh
new file mode 120000
index 00000000000..cd7ba62c0f3
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/format_lms.sh
@@ -0,0 +1 @@
+../../../librispeech/s5/local/format_lms.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/nnet3/compare_wer.sh b/egs/mini_librispeech/s5/local/nnet3/compare_wer.sh
new file mode 100755
index 00000000000..095e85cc338
--- /dev/null
+++ b/egs/mini_librispeech/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" $prob
+done
+echo
+
+echo
diff --git a/egs/mini_librispeech/s5/local/nnet3/run_ivector_common.sh b/egs/mini_librispeech/s5/local/nnet3/run_ivector_common.sh
new file mode 100755
index 00000000000..82bb46d64a9
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/nnet3/run_ivector_common.sh
@@ -0,0 +1,148 @@
+#!/bin/bash
+
+set -euo pipefail
+
+# This script is called from local/nnet3/run_tdnn.sh and
+# local/chain/run_tdnn.sh (and may eventually be called by more
+# scripts).  It contains the common feature preparation and
+# iVector-related parts of the script.  See those scripts for examples
+# of usage.
+
+stage=0
+train_set=train_clean_5
+test_sets="dev_clean_2"
+gmm=tri3b
+
+nnet3_affix=
+
+. ./cmd.sh
+. ./path.sh
+. utils/parse_options.sh
+
+gmm_dir=exp/${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 $mfccdir/storage ]; then
+    utils/create_split_dir.pl /export/b1{5,6,7,8}/$USER/kaldi-data/egs/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 $ivectordir/storage ]; then
+    utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/egs/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$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/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh b/egs/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh
new file mode 100755
index 00000000000..48f7f6cdee5
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1a.sh
@@ -0,0 +1,220 @@
+#!/bin/bash
+
+# This is a basic TDNN+LSTM nnet3 experiment.
+
+
+# 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)      18.58     17.67
+#WER dev_clean_2 (tglarge)      13.35     13.43
+# Final train prob                  -0.3660
+# Final valid prob                  -1.0236
+# Final train acc                    0.8737
+# Final valid acc                    0.7222
+
+# 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/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh b/egs/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh
new file mode 100755
index 00000000000..af0779c0109
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/nnet3/tuning/run_tdnn_lstm_1b.sh
@@ -0,0 +1,210 @@
+#!/bin/bash
+
+# This is like 1a, but adding dropout.   It seems g
+
+
+# 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/mini_librispeech/s5/local/prepare_dict.sh b/egs/mini_librispeech/s5/local/prepare_dict.sh
new file mode 120000
index 00000000000..aa044af81d3
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/prepare_dict.sh
@@ -0,0 +1 @@
+../../../librispeech/s5/local/prepare_dict.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/score.sh b/egs/mini_librispeech/s5/local/score.sh
new file mode 100755
index 00000000000..c812199fc98
--- /dev/null
+++ b/egs/mini_librispeech/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/mini_librispeech/s5/local/subset_dataset.sh b/egs/mini_librispeech/s5/local/subset_dataset.sh
new file mode 100755
index 00000000000..050128247a4
--- /dev/null
+++ b/egs/mini_librispeech/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/${dest}_chapter_id_list.txt
diff --git a/egs/mini_librispeech/s5/path.sh b/egs/mini_librispeech/s5/path.sh
new file mode 100644
index 00000000000..705600ad47a
--- /dev/null
+++ b/egs/mini_librispeech/s5/path.sh
@@ -0,0 +1,8 @@
+export KALDI_ROOT=`pwd`/../../..
+export PATH=$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
+
+# For now, don't include any of the optional dependenices of the main
+# librispeech recipe
diff --git a/egs/mini_librispeech/s5/run.sh b/egs/mini_librispeech/s5/run.sh
new file mode 100755
index 00000000000..bf1ded337f6
--- /dev/null
+++ b/egs/mini_librispeech/s5/run.sh
@@ -0,0 +1,205 @@
+#!/bin/bash
+
+# Note: this works only on pre-downloaded data on the CLSP servers
+data=/export/a05/dgalvez/
+
+data_url=www.openslr.org/resources/TODO # TODO
+lm_url=www.openslr.org/resources/11
+
+. ./cmd.sh
+. ./path.sh
+
+stage=0
+. utils/parse_options.sh
+
+# TODO(galv): Reconsider this
+set -euxo pipefail
+
+# TODO(galv): Modify openslr.org to contain the minified training dataset.
+# for part in dev-clean-2 train-clean-5; do
+#   local/download_and_untar.sh $data $data_url $part
+# done
+
+if [ $stage -le 0 ]; then
+  local/download_lm.sh $lm_url data/local/lm
+fi
+
+if [ $stage -le 1 ]; then
+  # format the data as Kaldi data directories
+  for part in dev-clean-2 train-clean-5; do
+    # use underscore-separated names in data directories.
+    local/data_prep.sh $data/LibriSpeech/$part data/$(echo $part | sed s/-/_/g)
+  done
+
+  local/prepare_dict.sh --stage 3 --nj 30 --cmd "$train_cmd" \
+    data/local/lm data/local/lm data/local/dict_nosp
+
+  utils/prepare_lang.sh data/local/dict_nosp \
+    "<UNK>" data/local/lang_tmp_nosp data/lang_nosp
+
+  local/format_lms.sh --src-dir data/lang_nosp 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_nosp data/lang_nosp_test_tglarge
+fi
+
+if [ $stage -le 2 ]; then
+  mfccdir=mfcc
+  # spread the mfccs over various machines, as this data-set is quite large.
+  if [[  $(hostname -f) ==  *.clsp.jhu.edu ]]; then
+    mfcc=$(basename mfccdir) # in case was absolute pathname (unlikely), get basename.
+    utils/create_split_dir.pl /export/b{07,14,16,17}/$USER/kaldi-data/egs/librispeech/s5/$mfcc/storage \
+      $mfccdir/storage
+  fi
+
+  for part in dev_clean_2 train_clean_5; do
+    steps/make_mfcc.sh --cmd "$train_cmd" --nj 10 data/$part exp/make_mfcc/$part $mfccdir
+    steps/compute_cmvn_stats.sh data/$part exp/make_mfcc/$part $mfccdir
+  done
+
+  # Get the shortest 500 utterances first because those are more likely
+  # to have accurate alignments.
+  utils/subset_data_dir.sh --shortest data/train_clean_5 500 data/train_500short
+fi
+
+# train a monophone system
+if [ $stage -le 3 ]; then
+  # TODO(galv): Is this too many jobs for a smaller dataset?
+  steps/train_mono.sh --boost-silence 1.25 --nj 5 --cmd "$train_cmd" \
+    data/train_500short data/lang_nosp exp/mono
+  # TODO: Understand why we use lang_nosp here...
+  (
+    utils/mkgraph.sh data/lang_nosp_test_tgsmall \
+      exp/mono exp/mono/graph_nosp_tgsmall
+    for test in dev_clean_2; do
+      steps/decode.sh --nj 10 --cmd "$decode_cmd" exp/mono/graph_nosp_tgsmall \
+        data/$test exp/mono/decode_nosp_tgsmall_$test
+    done
+  )&
+
+  steps/align_si.sh --boost-silence 1.25 --nj 5 --cmd "$train_cmd" \
+    data/train_clean_5 data/lang_nosp exp/mono exp/mono_ali_train_clean_5
+fi
+
+# train a first delta + delta-delta triphone system on all utterances
+if [ $stage -le 4 ]; then
+  steps/train_deltas.sh --boost-silence 1.25 --cmd "$train_cmd" \
+    2000 10000 data/train_clean_5 data/lang_nosp exp/mono_ali_train_clean_5 exp/tri1
+
+  # decode using the tri1 model
+  (
+    utils/mkgraph.sh data/lang_nosp_test_tgsmall \
+      exp/tri1 exp/tri1/graph_nosp_tgsmall
+    for test in dev_clean_2; do
+      steps/decode.sh --nj 5 --cmd "$decode_cmd" exp/tri1/graph_nosp_tgsmall \
+      data/$test exp/tri1/decode_nosp_tgsmall_$test
+      steps/lmrescore.sh --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tgmed} \
+        data/$test exp/tri1/decode_nosp_{tgsmall,tgmed}_$test
+      steps/lmrescore_const_arpa.sh \
+        --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tglarge} \
+        data/$test exp/tri1/decode_nosp_{tgsmall,tglarge}_$test
+    done
+  )&
+
+  steps/align_si.sh --nj 5 --cmd "$train_cmd" \
+    data/train_clean_5 data/lang_nosp exp/tri1 exp/tri1_ali_train_clean_5
+fi
+
+# train an LDA+MLLT system.
+if [ $stage -le 5 ]; then
+  steps/train_lda_mllt.sh --cmd "$train_cmd" \
+    --splice-opts "--left-context=3 --right-context=3" 2500 15000 \
+    data/train_clean_5 data/lang_nosp exp/tri1_ali_train_clean_5 exp/tri2b
+
+  # decode using the LDA+MLLT model
+  (
+    utils/mkgraph.sh data/lang_nosp_test_tgsmall \
+      exp/tri2b exp/tri2b/graph_nosp_tgsmall
+    for test in dev_clean_2; do
+      steps/decode.sh --nj 10 --cmd "$decode_cmd" exp/tri2b/graph_nosp_tgsmall \
+        data/$test exp/tri2b/decode_nosp_tgsmall_$test
+      steps/lmrescore.sh --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tgmed} \
+        data/$test exp/tri2b/decode_nosp_{tgsmall,tgmed}_$test
+      steps/lmrescore_const_arpa.sh \
+        --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tglarge} \
+        data/$test exp/tri2b/decode_nosp_{tgsmall,tglarge}_$test
+    done
+  )&
+
+  # Align utts using the tri2b model
+  steps/align_si.sh  --nj 5 --cmd "$train_cmd" --use-graphs true \
+    data/train_clean_5 data/lang_nosp exp/tri2b exp/tri2b_ali_train_clean_5
+fi
+
+# Train tri3b, which is LDA+MLLT+SAT
+if [ $stage -le 6 ]; then
+  steps/train_sat.sh --cmd "$train_cmd" 2500 15000 \
+    data/train_clean_5 data/lang_nosp exp/tri2b_ali_train_clean_5 exp/tri3b
+
+  # decode using the tri3b model
+  (
+    utils/mkgraph.sh data/lang_nosp_test_tgsmall \
+      exp/tri3b exp/tri3b/graph_nosp_tgsmall
+    for test in dev_clean_2; do
+      steps/decode_fmllr.sh --nj 10 --cmd "$decode_cmd" \
+        exp/tri3b/graph_nosp_tgsmall data/$test \
+        exp/tri3b/decode_nosp_tgsmall_$test
+      steps/lmrescore.sh --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tgmed} \
+        data/$test exp/tri3b/decode_nosp_{tgsmall,tgmed}_$test
+      steps/lmrescore_const_arpa.sh \
+        --cmd "$decode_cmd" data/lang_nosp_test_{tgsmall,tglarge} \
+        data/$test exp/tri3b/decode_nosp_{tgsmall,tglarge}_$test
+    done
+  )&
+fi
+
+# Now we compute the pronunciation and silence probabilities from training data,
+# and re-create the lang directory.
+if [ $stage -le 7 ]; then
+  steps/get_prons.sh --cmd "$train_cmd" \
+    data/train_clean_5 data/lang_nosp exp/tri3b
+  utils/dict_dir_add_pronprobs.sh --max-normalize true \
+    data/local/dict_nosp \
+    exp/tri3b/pron_counts_nowb.txt exp/tri3b/sil_counts_nowb.txt \
+    exp/tri3b/pron_bigram_counts_nowb.txt data/local/dict
+
+  utils/prepare_lang.sh data/local/dict \
+    "<UNK>" data/local/lang_tmp data/lang
+
+  local/format_lms.sh --src-dir data/lang data/local/lm
+
+  utils/build_const_arpa_lm.sh \
+    data/local/lm/lm_tglarge.arpa.gz data/lang data/lang_test_tglarge
+
+  steps/align_fmllr.sh --nj 5 --cmd "$train_cmd" \
+    data/train_clean_5 data/lang exp/tri3b exp/tri3b_ali_train_clean_5
+fi
+
+
+if [ $stage -le 8 ]; then
+  # Test the tri3b system with the silprobs and pron-probs.
+
+  # decode using the tri3b model
+  utils/mkgraph.sh data/lang_test_tgsmall \
+                   exp/tri3b exp/tri3b/graph_tgsmall
+  for test in dev_clean_2; do
+    steps/decode_fmllr.sh --nj 10 --cmd "$decode_cmd" \
+                          exp/tri3b/graph_tgsmall data/$test \
+                          exp/tri3b/decode_tgsmall_$test
+    steps/lmrescore.sh --cmd "$decode_cmd" data/lang_test_{tgsmall,tgmed} \
+                       data/$test exp/tri3b/decode_{tgsmall,tgmed}_$test
+    steps/lmrescore_const_arpa.sh \
+      --cmd "$decode_cmd" data/lang_test_{tgsmall,tglarge} \
+      data/$test exp/tri3b/decode_{tgsmall,tglarge}_$test
+  done
+fi
+
+exit 0 # temp
+
+# Train a chain model
+if [ $stage -le 9 ]; then
+  local/chain/run_tdnn.sh --stage 0
+fi
+
+# Don't finish until all background decoding jobs are finished.
+wait
diff --git a/egs/mini_librispeech/s5/steps b/egs/mini_librispeech/s5/steps
new file mode 120000
index 00000000000..1b186770dd1
--- /dev/null
+++ b/egs/mini_librispeech/s5/steps
@@ -0,0 +1 @@
+../../wsj/s5/steps/
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/utils b/egs/mini_librispeech/s5/utils
new file mode 120000
index 00000000000..a3279dc8679
--- /dev/null
+++ b/egs/mini_librispeech/s5/utils
@@ -0,0 +1 @@
+../../wsj/s5/utils/
\ No newline at end of file
diff --git a/egs/wsj/s5/local/chain/tuning/run_tdnn_1a.sh b/egs/wsj/s5/local/chain/tuning/run_tdnn_1a.sh
index d874eb0986a..9d48ec7a898 100755
--- a/egs/wsj/s5/local/chain/tuning/run_tdnn_1a.sh
+++ b/egs/wsj/s5/local/chain/tuning/run_tdnn_1a.sh
@@ -226,7 +226,7 @@ fi
 if [ $stage -le 16 ]; 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/tedlium-$(date +'%m_%d_%H_%M')/s5_r2/$dir/egs/storage $dir/egs/storage
+     /export/b0{3,4,5,6}/$USER/kaldi-data/egs/wsj-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
   fi
 
   steps/nnet3/chain/train.py --stage=$train_stage \