diff --git a/egs/fisher_english/s5/local/chain/compare_wer_general.sh b/egs/fisher_english/s5/local/chain/compare_wer_general.sh
new file mode 100755
index 00000000000..2f724c8ff81
--- /dev/null
+++ b/egs/fisher_english/s5/local/chain/compare_wer_general.sh
@@ -0,0 +1,106 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer_general.sh exp/chain_cleaned/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer_general.sh exp/chain_cleaned/tdnn_c_sp exp/chain_cleaned/tdnn_c_sp_smbr:{1,2,3}
+
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=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_cleaned/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain_cleaned/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain_cleaned/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain_cleaned/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 $x)";   done
+echo
+
+strings=("# WER on dev    "  "# WER on test    ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+     decode_names=(dev${epoch_infix} test${epoch_infix})
+     wer=$(grep WER $dirname/decode_${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 $epoch_infix
+       decode_names=(dev${epoch_infix} test${epoch_infix})
+       wer=$(grep WER $dirname/decode_looped_${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/fisher_english/s5/local/chain/run_semisupervised.sh b/egs/fisher_english/s5/local/chain/run_semisupervised.sh
new file mode 100755
index 00000000000..7b9c56e063e
--- /dev/null
+++ b/egs/fisher_english/s5/local/chain/run_semisupervised.sh
@@ -0,0 +1,148 @@
+#!/bin/bash
+
+set -e -o pipefail
+
+stage=-2
+nj=30
+decode_nj=30
+base_train_set=train_comb350k # for reference
+
+unsupervised_set=train_unsup250k  # set this to your choice of unsupervised data
+supervised_set=train_sup
+semi_affix=350k  # affix relating train-set splitting proportion
+
+tdnn_affix=_sup1a  # affix for the supervised chain-model directory
+train_supervised_opts="--stage -10 --train-stage -10"
+
+# combination options
+decode_affix=
+egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
+comb_affix=_comb1a  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
+unsup_frames_per_eg=  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
+unsup_egs_weight=1.0
+lattice_lm_scale=0.0  # lm-scale for using the weights from unsupervised lattices
+lattice_prune_beam=2.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
+left_tolerance=2
+right_tolerance=2
+train_combined_opts="--num-epochs 4.5"
+graph_affix=   # can be used to decode the unsup data with another lm/graph
+phone_insertion_penalty=
+# to tune:
+# frames_per_eg for unsupervised
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+nnet3_affix=_semi${semi_affix}  # affix for nnet3 and chain dirs
+decode_affix=${decode_affix}${graph_affix}
+
+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
+
+if [ $stage -le -2 ]; then
+  ali_dir=exp/tri4a_ali_$supervised_set
+  if [ ! -f $ali_dir/ali.1.gz ]; then
+    steps/align_fmllr.sh --nj 30 --cmd "queue.pl" data/$supervised_set data/lang exp/tri4a $ali_dir
+  fi
+  echo "$0: chain training on the supervised subset data/${supervised_set}"
+  local/chain/run_tdnn.sh $train_supervised_opts --remove-egs false \
+                          --train-set $supervised_set \
+                          --build-tree-ali-dir $ali_dir \
+                          --nnet3-affix $nnet3_affix --tdnn-affix $tdnn_affix
+fi
+
+if [ $stage -le -1 ] && [ ! -f exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires/ivector_online.scp ]; then
+  echo "$0: getting ivectors for the hires unsupervised data data/${unsupervised_set}_hires"
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj "$nj" \
+            data/${unsupervised_set}_hires exp/nnet3${nnet3_affix}/extractor \
+            exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires
+fi
+
+chaindir=exp/chain${nnet3_affix}/tdnn${tdnn_affix}_sp
+
+left_context=`cat $chaindir/egs/info/left_context`
+right_context=`cat $chaindir/egs/info/right_context`
+left_context_initial=`cat $chaindir/egs/info/left_context_initial`
+right_context_final=`cat $chaindir/egs/info/right_context_final`
+[ -z $unsup_frames_per_eg ] && unsup_frames_per_eg=`cat $chaindir/egs/info/frames_per_eg`
+frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
+cmvn_opts=`cat $chaindir/cmvn_opts`
+
+if [ $stage -le 0 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  graphdir=$chaindir/graph${graph_affix}
+  if [ ! -f $graphdir/HCLG.fst ]; then
+    utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test${graph_affix} $chaindir $graphdir
+  fi
+  steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 \
+            --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
+            --scoring-opts "--min-lmwt 5 " \
+            $chaindir/graph${graph_affix} data/${unsupervised_set}_hires $chaindir/decode_${unsupervised_set}${decode_affix}
+  ln -s ../final.mdl $chaindir/decode_${unsupervised_set}${decode_affix}/final.mdl || true
+fi
+
+if [ $stage -le 1 ]; then
+  echo "$0: generating egs from the unsupervised data"
+  steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+             --left-tolerance $left_tolerance --right-tolerance $right_tolerance \
+             --left-context $left_context --right-context $right_context \
+             --left-context-initial $left_context_initial --right-context-final $right_context_final \
+             --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+             --frame-subsampling-factor $frame_subsampling_factor \
+             --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+             --lattice-prune-beam "$lattice_prune_beam" \
+             --egs-weight $unsup_egs_weight \
+             --phone-insertion-penalty "$phone_insertion_penalty" \
+             --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
+             data/${unsupervised_set}_hires $chaindir \
+             ${chaindir}/decode_${unsupervised_set}${decode_affix} $chaindir/unsup_egs${decode_affix}${egs_affix}
+fi
+
+sup_egs_dir=$chaindir/egs
+unsup_egs_dir=$chaindir/unsup_egs${decode_affix}${egs_affix}
+comb_egs_dir=$chaindir/comb_egs${decode_affix}${egs_affix}
+if [ $stage -le 2 ]; then
+  echo "$0: combining supervised/unsupervised egs"
+  n1=`cat $sup_egs_dir/info/num_archives`
+  n2=`cat $unsup_egs_dir/info/num_archives`
+  num_archives=$(($n2>$n1?$n2:$n1))
+  num_archives=$[num_archives*3/2]
+  mkdir -p $comb_egs_dir/log
+  cp {$sup_egs_dir,$comb_egs_dir}/train_diagnostic.cegs
+  cp {$sup_egs_dir,$comb_egs_dir}/valid_diagnostic.cegs
+  nnet3-chain-copy-egs "ark:cat $sup_egs_dir/combine.cegs $unsup_egs_dir/combine.cegs |" ark:$comb_egs_dir/combine.cegs
+  cp {$sup_egs_dir,$comb_egs_dir}/cmvn_opts
+  cp -r $sup_egs_dir/info $comb_egs_dir
+  echo $num_archives > $comb_egs_dir/info/num_archives
+  cat {$sup_egs_dir,$unsup_egs_dir}/info/num_frames | awk '{s+=$1} END{print s}' > $comb_egs_dir/info/num_frames
+  cat {$sup_egs_dir,$unsup_egs_dir}/info/egs_per_archive | awk '{s+=$1} END{print s}' > $comb_egs_dir/info/egs_per_archive
+  out_egs_list=
+  egs_list=
+  for n in $(seq $num_archives); do
+      [ -f $sup_egs_dir/cegs.$n.ark ] && egs_list="$egs_list $sup_egs_dir/cegs.$n.ark"
+      [ -f $unsup_egs_dir/cegs.$n.ark ] && egs_list="$egs_list $unsup_egs_dir/cegs.$n.ark"
+      out_egs_list="$out_egs_list ark:$comb_egs_dir/cegs.$n.ark"
+  done
+  srand=0
+  $decode_cmd $comb_egs_dir/log/combine.log \
+              nnet3-chain-copy-egs "ark:cat $egs_list|" $out_egs_list
+fi
+
+if [ $stage -le 3 ]; then
+  echo "$0: training on the supervised+unsupervised subset"
+  # the train-set and gmm do not matter as we are providing the egs
+  local/chain/run_tdnn.sh --stage 12 --remove-egs false --train-set $supervised_set \
+                          --nnet3-affix $nnet3_affix \
+                          --tdnn-affix ${tdnn_affix}${decode_affix}${egs_affix}${comb_affix} \
+                          --common-egs-dir $comb_egs_dir $train_combined_opts
+fi
diff --git a/egs/fisher_english/s5/local/chain/run_tdnn.sh b/egs/fisher_english/s5/local/chain/run_tdnn.sh
new file mode 100755
index 00000000000..7352152a3bc
--- /dev/null
+++ b/egs/fisher_english/s5/local/chain/run_tdnn.sh
@@ -0,0 +1,202 @@
+#!/bin/bash
+set -e
+
+# Based on run_tdnn_7b.sh in the fisher swbd recipe
+
+# configs for 'chain'
+stage=0
+tdnn_affix=7b
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+train_set=train
+tree_affix=
+nnet3_affix=
+xent_regularize=0.1
+hidden_dim=725
+num_leaves=11000  # number of PdfIds in chain modeling
+
+# training options
+num_epochs=4
+remove_egs=false
+common_egs_dir=
+minibatch_size=128
+num_jobs_initial=3
+num_jobs_final=16
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+frames_per_iter=1500000
+
+gmm=tri5a
+build_tree_ali_dir=exp/tri4a_ali  # used to make a new tree for chain topology, should match train data
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+
+gmm_dir=exp/$gmm   # used to get training lattices (for chain supervision)
+treedir=exp/chain${nnet3_affix}/tree_${tree_affix}
+lat_dir=exp/chain${nnet3_affix}/tri5a_${train_set}_sp_lats  # training lattices directory
+dir=exp/chain${nnet3_affix}/tdnn${tdnn_affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+lores_train_data_dir=data/${train_set}_sp
+build_tree_train_data_dir=data/${train_set}
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+lang=data/lang_chain
+
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --speed-perturb true \
+                                  --train-set $train_set \
+                                  --nnet3-affix $nnet3_affix \
+                                  --generate-alignments false || exit 1;
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat $build_tree_ali_dir/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" $lores_train_data_dir \
+    data/lang $gmm_dir $lat_dir || exit 1;
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 10 ]; then
+  # Create a version of the lang/ directory that has one state per phone in the
+  # topo file. [note, it really has two states.. the first one is only repeated
+  # once, the second one has zero or more repeats.]
+  rm -rf $lang
+  cp -r data/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate -1 \
+      --cmd "$train_cmd" $num_leaves $build_tree_train_data_dir $lang $build_tree_ali_dir $treedir || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1,2) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn6 input=Append(-6,-3,0) dim=$hidden_dim
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain input=tdnn6 dim=$hidden_dim target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn6 dim=$hidden_dim target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+if [ $stage -le 13 ]; then
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
+  fi
+
+  touch $dir/egs/.nodelete # keep egs around when that run dies.
+
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $train_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize 0.1 \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width 150 \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter $frames_per_iter \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate $initial_effective_lrate \
+    --trainer.optimization.final-effective-lrate $final_effective_lrate \
+    --trainer.max-param-change 2.0 \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir $train_data_dir \
+    --tree-dir $treedir \
+    --lat-dir $lat_dir \
+    --dir $dir  || exit 1;
+fi
+
+graph_dir=$dir/graph
+if [ $stage -le 14 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test $dir $graph_dir
+fi
+
+decode_suff=
+if [ $stage -le 15 ]; then
+  iter_opts=
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in dev test; do
+      (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+      ) &
+  done
+fi
+wait;
+exit 0;
diff --git a/egs/fisher_english/s5/local/nnet3/run_ivector_common.sh b/egs/fisher_english/s5/local/nnet3/run_ivector_common.sh
new file mode 100755
index 00000000000..6505381b03f
--- /dev/null
+++ b/egs/fisher_english/s5/local/nnet3/run_ivector_common.sh
@@ -0,0 +1,147 @@
+#!/bin/bash
+
+. ./cmd.sh
+set -e
+stage=1
+generate_alignments=true  # false if doing chain training
+speed_perturb=true
+train_set=train
+
+lda_train_set=train_100k
+nnet3_affix=
+gmm=tri2_ali   # should also contain alignments for $lda_train_set
+
+. ./path.sh
+. ./utils/parse_options.sh
+
+gmm_dir=exp/$gmm
+
+# perturbed data preparation
+if [ "$speed_perturb" == "true" ]; then
+  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 alignments.
+    # _sp stands for speed-perturbed
+
+    for datadir in ${train_set}; do
+      utils/perturb_data_dir_speed.sh 0.9 data/${datadir} data/temp1
+      utils/perturb_data_dir_speed.sh 1.1 data/${datadir} data/temp2
+      utils/combine_data.sh data/${datadir}_tmp data/temp1 data/temp2
+      utils/validate_data_dir.sh --no-feats data/${datadir}_tmp
+      rm -r data/temp1 data/temp2
+
+      mfccdir=mfcc_perturbed
+      steps/make_mfcc.sh --cmd "$train_cmd" --nj 50 \
+        data/${datadir}_tmp exp/make_mfcc/${datadir}_tmp $mfccdir || exit 1;
+      steps/compute_cmvn_stats.sh data/${datadir}_tmp exp/make_mfcc/${datadir}_tmp $mfccdir || exit 1;
+      utils/fix_data_dir.sh data/${datadir}_tmp
+
+      utils/copy_data_dir.sh --spk-prefix sp1.0- --utt-prefix sp1.0- data/${datadir} data/temp0
+      utils/combine_data.sh data/${datadir}_sp data/${datadir}_tmp data/temp0
+      utils/fix_data_dir.sh data/${datadir}_sp
+      rm -r data/temp0 data/${datadir}_tmp
+    done
+  fi
+
+  if [ $stage -le 2 ] && [ "$generate_alignments" == "true" ]; then
+    #obtain the alignment of the perturbed data
+    steps/align_fmllr.sh --nj 100 --cmd "$train_cmd" \
+      data/${train_set}_sp data/lang exp/tri5a exp/tri5a_ali_${train_set}_sp || exit 1
+  fi
+  train_set=${train_set}_sp
+fi
+
+if [ $stage -le 3 ]; then
+  mfccdir=mfcc_hires
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
+    date=$(date +'%m_%d_%H_%M')
+    utils/create_split_dir.pl /export/b0{1,2,3,4}/$USER/kaldi-data/egs/fisher_english-$date/s5b/$mfccdir/storage $mfccdir/storage
+  fi
+
+  # the 100k directory is copied seperately, as
+  # we want to use exp/tri2_ali for lda_mllt training
+  # the main train directory might be speed_perturbed
+  for dataset in $train_set $lda_train_set; do
+    utils/copy_data_dir.sh data/$dataset data/${dataset}_hires
+
+    # scale the waveforms, this is useful as we don't use CMVN
+    data_dir=data/${dataset}_hires
+    cat $data_dir/wav.scp | python -c "
+import sys, os, subprocess, re, random
+scale_low = 1.0/8
+scale_high = 2.0
+for line in sys.stdin.readlines():
+  if len(line.strip()) == 0:
+    continue
+  print '{0} sox --vol {1} -t wav - -t wav - |'.format(line.strip(), random.uniform(scale_low, scale_high))
+"| sort -k1,1 -u  > $data_dir/wav.scp_scaled || exit 1;
+    mv $data_dir/wav.scp_scaled $data_dir/wav.scp
+
+    steps/make_mfcc.sh --nj 70 --mfcc-config conf/mfcc_hires.conf \
+        --cmd "$train_cmd" data/${dataset}_hires exp/make_hires/$dataset $mfccdir;
+    steps/compute_cmvn_stats.sh data/${dataset}_hires exp/make_hires/${dataset} $mfccdir;
+
+    # Remove the small number of utterances that couldn't be extracted for some
+    # reason (e.g. too short; no such file).
+    utils/fix_data_dir.sh data/${dataset}_hires;
+  done
+
+  for dataset in test dev; do
+    # Create MFCCs for the eval set
+    utils/copy_data_dir.sh data/$dataset data/${dataset}_hires
+    steps/make_mfcc.sh --cmd "$train_cmd" --nj 10 --mfcc-config conf/mfcc_hires.conf \
+        data/${dataset}_hires exp/make_hires/$dataset $mfccdir;
+    steps/compute_cmvn_stats.sh data/${dataset}_hires exp/make_hires/$dataset $mfccdir;
+    utils/fix_data_dir.sh data/${dataset}_hires  # remove segments with problems
+  done
+
+  # Take the first 30k utterances (about 1/8th of the data) this will be used
+  # for the diagubm training
+  utils/subset_data_dir.sh --first data/${train_set}_hires 30000 data/${train_set}_30k_hires
+  utils/data/remove_dup_utts.sh 200 data/${train_set}_30k_hires data/${train_set}_30k_nodup_hires  # 33hr
+fi
+
+# ivector extractor training
+if [ $stage -le 4 ]; then
+  # We need to build a small system just because we need the LDA+MLLT transform
+  # to train the diag-UBM on top of.  We use --num-iters 13 because after we get
+  # the transform (12th iter is the last), any further training is pointless.
+  # this decision is based on fisher_english
+  steps/train_lda_mllt.sh --cmd "$train_cmd" --num-iters 13 \
+    --splice-opts "--left-context=3 --right-context=3" \
+    5500 90000 data/${lda_train_set}_hires \
+    data/lang $gmm_dir exp/nnet3${nnet3_affix}/tri3a
+fi
+
+if [ $stage -le 5 ]; then
+  # To train a diagonal UBM we don't need very much data, so use the smallest subset.
+  steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj 30 --num-frames 200000 \
+    data/${train_set}_30k_nodup_hires 512 exp/nnet3${nnet3_affix}/tri3a exp/nnet3${nnet3_affix}/diag_ubm
+fi
+
+if [ $stage -le 6 ]; then
+  # iVector extractors can be sensitive to the amount of data, but this one has a
+  # fairly small dim (defaults to 100) so we don't use all of it, we use just the
+  # 100k subset (just under half the data).
+  steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj 10 \
+    data/${lda_train_set}_hires exp/nnet3${nnet3_affix}/diag_ubm exp/nnet3${nnet3_affix}/extractor || exit 1;
+fi
+
+if [ $stage -le 7 ]; then
+  # We extract iVectors on all the ${train_set} data, which will be what we
+  # train the system on.
+
+  # having a larger number of speakers is helpful for generalization, and to
+  # handle per-utterance decoding well (iVector starts at zero).
+  steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 data/${train_set}_hires data/${train_set}_max2_hires
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
+    data/${train_set}_max2_hires exp/nnet3${nnet3_affix}/extractor exp/nnet3${nnet3_affix}/ivectors_${train_set}_hires || exit 1;
+
+  for dataset in test dev; do
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
+      data/${dataset}_hires exp/nnet3${nnet3_affix}/extractor exp/nnet3${nnet3_affix}/ivectors_${dataset}_hires || exit 1;
+  done
+fi
+
+exit 0;
diff --git a/egs/fisher_english/s5/local/nnet3/run_tdnn.sh b/egs/fisher_english/s5/local/nnet3/run_tdnn.sh
new file mode 100644
index 00000000000..f055b853b61
--- /dev/null
+++ b/egs/fisher_english/s5/local/nnet3/run_tdnn.sh
@@ -0,0 +1,98 @@
+#!/bin/bash
+
+# This script is not tested.
+
+# this is the standard "tdnn" system, built in nnet3; it's what we used to
+# call multi-splice.
+
+. ./cmd.sh
+
+
+# At this script level we don't support not running on GPU, as it would be painfully slow.
+# If you want to run without GPU you'd have to call train_tdnn.sh with --gpu false,
+# --num-threads 16 and --minibatch-size 128.
+
+stage=0
+affix=
+train_stage=-10
+common_egs_dir=
+reporting_email=
+remove_egs=true
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+dir=exp/nnet3/tdnn
+dir=$dir${affix:+_$affix}
+train_set=train_sp
+ali_dir=exp/tri5a_ali_train_sp
+
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --speed-perturb true || exit 1;
+
+if [ $stage -le 9 ]; then
+  echo "$0: creating neural net configs";
+
+  # create the config files for nnet initialization
+  python steps/nnet3/tdnn/make_configs.py  \
+    --feat-dir data/${train_set}_hires \
+    --ivector-dir exp/nnet3/ivectors_${train_set} \
+    --ali-dir $ali_dir \
+    --relu-dim 1024 \
+    --splice-indexes "-2,-1,0,1,2 -1,2 -3,3 -3,3 -7,2 0"  \
+    --use-presoftmax-prior-scale true \
+   $dir/configs || exit 1;
+fi
+
+
+if [ $stage -le 10 ]; 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/fisher_swbd-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/train_dnn.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.0017 \
+    --trainer.optimization.final-effective-lrate 0.00017 \
+    --egs.dir "$common_egs_dir" \
+    --cleanup.remove-egs $remove_egs \
+    --cleanup.preserve-model-interval 500 \
+    --use-gpu true \
+    --feat-dir=data/${train_set}_hires \
+    --ali-dir $ali_dir \
+    --lang data/lang \
+    --reporting.email="$reporting_email" \
+    --dir=$dir  || exit 1;
+
+fi
+
+graph_dir=exp/tri5a/graph
+if [ $stage -le 11 ]; then
+  for decode_set in dev test; do
+    (
+    num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+    steps/nnet3/decode.sh --nj $num_jobs --cmd "$decode_cmd" \
+        --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+       $graph_dir data/${decode_set}_hires $dir/decode_${decode_set} || exit 1;
+    ) &
+  done
+fi
+wait;
+exit 0;
+
diff --git a/egs/fisher_english/s5/local/score.sh b/egs/fisher_english/s5/local/score.sh
index c381abf7277..702a57c94c6 100755
--- a/egs/fisher_english/s5/local/score.sh
+++ b/egs/fisher_english/s5/local/score.sh
@@ -5,6 +5,7 @@
 cmd=run.pl
 min_lmwt=5
 max_lmwt=17
+iter=
 #end configuration section.
 
 [ -f ./path.sh ] && . ./path.sh
diff --git a/egs/tedlium/s5_r2/local/chain/run_semisupervised.sh b/egs/tedlium/s5_r2/local/chain/run_semisupervised.sh
new file mode 100755
index 00000000000..674b8745c42
--- /dev/null
+++ b/egs/tedlium/s5_r2/local/chain/run_semisupervised.sh
@@ -0,0 +1,181 @@
+#!/bin/bash
+
+set -e -o pipefail
+
+# e.g. try lm-scale:
+# local/chain/run_semisupervised.sh --stage 1 --tdnn-affix _sup1a --egs-affix _lmwt1.0 --lattice-lm-scale 1.0
+
+
+# frames_per_eg 300
+# local/chain/run_semisupervised.sh --stage 1 --tdnn-affix _sup1d --unsup-frames-per-eg 300 --egs-affix _fpe300
+
+stage=0
+nj=30
+decode_nj=30
+base_train_set=train_cleaned # the starting point train-set
+base_gmm=tri3_cleaned  # the starting point of training on the supervised data (no flat start for now)
+semi_affix=  # affix relating train-set splitting proportion
+             # (currently supervised 25%) and the base train set (currently _cleaned), etc.
+tdnn_affix=_sup1a  # affix for the supervised chain-model directory
+train_supervised_opts="--stage -10 --train-stage -10"
+
+# combination options
+decode_affix=
+egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
+comb_affix=_comb1a  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
+unsup_frames_per_eg=  # if empty will be equal to the supervised model's config
+unsup_egs_weight=1.0
+lattice_lm_scale=0.1  # lm-scale for using the weights from unsupervised lattices
+lattice_prune_beam=  # If supplied will prune the lattices prior to getting egs for unsupervised data
+left_tolerance=2
+right_tolerance=2
+train_combined_opts="--num-epochs 5"
+
+# to tune:
+# frames_per_eg for unsupervised
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+supervised_set=${base_train_set}_sup${semi_affix}
+unsupervised_set=${base_train_set}_unsup${semi_affix}
+gmm=${base_gmm}_semi${semi_affix}  # the gmm to be supplied to chain/run_tdnn.sh
+nnet3_affix=_cleaned_semi${semi_affix}  # affix for nnet3 and chain dirs
+
+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
+
+if [ $stage -le -4 ]; then
+  echo "$0: preparing the supervised and unsupervised subsets of data"
+  if [ -f data/$supervised_set/feats.scp ]; then
+    echo "$0: supervised subset of data already exists; either delete it or use a later stage"
+    exit 1;
+  fi
+  mkdir -p data/$supervised_set
+  # get the list of supervised utts
+  num_utts=`wc -l data/$base_train_set/feats.scp | cut -d' ' -f1`
+  num_supervised_utts=$[num_utts/4]
+  num_unsupervised_utts=$[num_utts-num_supervised_utts]
+  echo "$0: spliting data/$base_train_set to supervised subset with"
+  echo "    $num_supervised_utts utts and unsupervised subset with $num_unsupervised_utts utts."
+  utils/shuffle_list.pl data/$base_train_set/feats.scp | cut -d' ' -f1 | \
+                        head -$num_supervised_utts > data/$supervised_set/supervised_uttlist || true
+  utils/shuffle_list.pl data/$base_train_set/feats.scp | cut -d' ' -f1 | \
+                        tail -$num_unsupervised_utts > data/$supervised_set/unsupervised_uttlist || true
+  utils/subset_data_dir.sh --utt-list data/$supervised_set/supervised_uttlist \
+                           data/$base_train_set data/$supervised_set || exit 1
+  utils/subset_data_dir.sh --utt-list data/$supervised_set/unsupervised_uttlist \
+                           data/$base_train_set data/$unsupervised_set || exit 1
+  utils/data/subset_data_dir.sh --utt-list data/$unsupervised_set/feats.scp \
+                                data/${base_train_set}_sp_hires data/${unsupervised_set}_hires
+fi
+
+if [ $stage -le -3 ]; then
+  # align the supervised subset with the current cleaned gmm
+  if [ -f $gmm/ali.1.gz ]; then
+    echo "$0: alignments in $gmm appear to already exist.  Please either remove them "
+    echo " ... or use a later --stage option."
+    exit 1
+  fi
+  echo "$0: aligning the supervised data data/${supervised_set}"
+  steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
+                       data/${supervised_set} data/lang exp/$base_gmm exp/$gmm
+fi
+
+if [ $stage -le -2 ]; then
+  echo "$0: chain training on the supervised subset data/${supervised_set}"
+  local/chain/run_tdnn.sh $train_supervised_opts --remove-egs false \
+                          --train-set $supervised_set --gmm $gmm \
+                          --nnet3-affix $nnet3_affix --tdnn-affix $tdnn_affix
+fi
+
+if [ $stage -le -1 ]; then
+  echo "$0: getting ivectors for the hires unsupervised data data/${unsupervised_set}_hires"
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj "$nj" \
+            data/${unsupervised_set}_hires exp/nnet3${nnet3_affix}/extractor \
+            exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires
+fi
+
+chaindir=exp/chain${nnet3_affix}/tdnn${tdnn_affix}_sp_bi
+
+left_context=`cat $chaindir/egs/info/left_context`
+right_context=`cat $chaindir/egs/info/right_context`
+left_context_initial=`cat $chaindir/egs/info/left_context_initial`
+right_context_final=`cat $chaindir/egs/info/right_context_final`
+[ -z $unsup_frames_per_eg ] && unsup_frames_per_eg=`cat $chaindir/egs/info/frames_per_eg`
+frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
+cmvn_opts=`cat $chaindir/cmvn_opts`
+
+if [ $stage -le 0 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 \
+            --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
+            --scoring-opts "--min-lmwt 5 " \
+            $chaindir/graph data/${unsupervised_set}_hires $chaindir/decode_${unsupervised_set}${decode_affix}
+  steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" data/lang data/lang_rescore \
+              data/${unsupervised_set}_hires \
+              ${chaindir}/decode_${unsupervised_set}${decode_affix} ${chaindir}/decode_${unsupervised_set}${decode_affix}_rescore
+  ln -s ../final.mdl $chaindir/decode_${unsupervised_set}${decode_affix}_rescore/final.mdl || true
+fi
+
+if [ $stage -le 1 ]; then
+  echo "$0: generating egs from the unsupervised data"
+  steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+             --left-tolerance $left_tolerance --right-tolerance $right_tolerance \
+             --left-context $left_context --right-context $right_context \
+             --left-context-initial $left_context_initial --right-context-final $right_context_final \
+             --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+             --frame-subsampling-factor $frame_subsampling_factor \
+             --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+             --lattice-prune-beam "$lattice_prune_beam" \
+             --egs-weight $unsup_egs_weight \
+             --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
+             data/${unsupervised_set}_hires $chaindir \
+             ${chaindir}/decode_${unsupervised_set}${decode_affix}_rescore $chaindir/unsup_egs${decode_affix}${egs_affix}
+fi
+
+sup_egs_dir=$chaindir/egs
+unsup_egs_dir=$chaindir/unsup_egs${decode_affix}${egs_affix}
+comb_egs_dir=$chaindir/comb_egs${decode_affix}${egs_affix}
+if [ $stage -le 2 ]; then
+  echo "$0: combining supervised/unsupervised egs"
+  num_archives=`cat $chaindir/egs/info/num_archives`
+  mkdir -p $comb_egs_dir/log
+  cp {$sup_egs_dir,$comb_egs_dir}/train_diagnostic.cegs
+  cp {$sup_egs_dir,$comb_egs_dir}/valid_diagnostic.cegs
+  cp {$sup_egs_dir,$comb_egs_dir}/combine.cegs
+  cp {$sup_egs_dir,$comb_egs_dir}/cmvn_opts
+  cp -r $sup_egs_dir/info $comb_egs_dir
+  cat {$sup_egs_dir,$unsup_egs_dir}/info/num_frames | awk '{s+=$1} END{print s}' > $comb_egs_dir/info/num_frames
+  cat {$sup_egs_dir,$unsup_egs_dir}/info/egs_per_archive | awk '{s+=$1} END{print s}' > $comb_egs_dir/info/egs_per_archive
+  out_egs_list=
+  egs_list=
+  for n in $(seq $num_archives); do
+      egs_list="$egs_list $sup_egs_dir/cegs.$n.ark"
+      egs_list="$egs_list $unsup_egs_dir/cegs.$n.ark"
+      out_egs_list="$out_egs_list ark:$comb_egs_dir/cegs.$n.ark"
+  done
+  srand=0
+  $decode_cmd $comb_egs_dir/log/combine.log \
+              nnet3-chain-copy-egs "ark:cat $egs_list|" $out_egs_list
+fi
+
+if [ $stage -le 3 ]; then
+  echo "$0: training on the supervised+unsupervised subset"
+  # the train-set and gmm do not matter as we are providing the egs
+  local/chain/run_tdnn.sh --stage 17 --remove-egs false --train-set $supervised_set --gmm $gmm \
+                          --nnet3-affix $nnet3_affix \
+                          --tdnn-affix ${tdnn_affix}${decode_affix}${egs_affix}${comb_affix} \
+                          --common-egs-dir $comb_egs_dir $train_combined_opts
+fi
+
diff --git a/egs/tedlium/s5_r2/local/chain/tuning/run_tdnn_1d.sh b/egs/tedlium/s5_r2/local/chain/tuning/run_tdnn_1d.sh
index 99921a9bf61..1c4a032fc57 100755
--- a/egs/tedlium/s5_r2/local/chain/tuning/run_tdnn_1d.sh
+++ b/egs/tedlium/s5_r2/local/chain/tuning/run_tdnn_1d.sh
@@ -52,6 +52,7 @@ train_set=train_cleaned
 gmm=tri3_cleaned  # the gmm for the target data
 num_threads_ubm=32
 nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
+num_epochs=4
 
 # The rest are configs specific to this script.  Most of the parameters
 # are just hardcoded at this level, in the commands below.
@@ -59,6 +60,7 @@ train_stage=-10
 tree_affix=  # affix for tree directory, e.g. "a" or "b", in case we change the configuration.
 tdnn_affix=1d  #affix for TDNN directory, e.g. "a" or "b", in case we change the configuration.
 common_egs_dir=  # you can set this to use previously dumped egs.
+remove_egs=true
 
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
@@ -212,13 +214,13 @@ if [ $stage -le 18 ]; then
     --egs.chunk-width 150 \
     --trainer.num-chunk-per-minibatch 128 \
     --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs 4 \
+    --trainer.num-epochs $num_epochs \
     --trainer.optimization.num-jobs-initial 2 \
     --trainer.optimization.num-jobs-final 12 \
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.max-param-change 2.0 \
-    --cleanup.remove-egs true \
+    --cleanup.remove-egs $remove_egs \
     --feat-dir $train_data_dir \
     --tree-dir $tree_dir \
     --lat-dir $lat_dir \
diff --git a/egs/wsj/s5/steps/nnet3/chain/get_egs.sh b/egs/wsj/s5/steps/nnet3/chain/get_egs.sh
index c1d45e41de7..076dc95b2d7 100755
--- a/egs/wsj/s5/steps/nnet3/chain/get_egs.sh
+++ b/egs/wsj/s5/steps/nnet3/chain/get_egs.sh
@@ -64,6 +64,14 @@ online_ivector_dir=  # can be used if we are including speaker information as iV
 cmvn_opts=  # can be used for specifying CMVN options, if feature type is not lda (if lda,
             # it doesn't make sense to use different options than were used as input to the
             # LDA transform).  This is used to turn off CMVN in the online-nnet experiments.
+lattice_lm_scale=     # If supplied, the graph/lm weight of the lattices will be
+                      # used (with this scale) in generating supervisions
+egs_weight=1.0    # The weight which determines how much each training example
+                           # contributes to gradients while training (can be used
+                           # to down/up-weight a dataset)
+lattice_prune_beam=         # If supplied, the lattices will be pruned to this beam,
+                            # before being used to get supervisions.
+phone_insertion_penalty=
 
 echo "$0 $@"  # Print the command line for logging
 
@@ -288,8 +296,10 @@ fi
 if [ $stage -le 2 ]; then
   echo "$0: copying training lattices"
 
+  [ ! -z $lattice_prune_beam ] && \
+    prune_cmd="ark:- | lattice-prune --acoustic-scale=0.1 --beam=$lattice_prune_beam ark:-"
   $cmd --max-jobs-run 6 JOB=1:$num_lat_jobs $dir/log/lattice_copy.JOB.log \
-    lattice-copy "ark:gunzip -c $latdir/lat.JOB.gz|" ark,scp:$dir/lat.JOB.ark,$dir/lat.JOB.scp || exit 1;
+    lattice-copy "ark:gunzip -c $latdir/lat.JOB.gz|" $prune_cmd ark,scp:$dir/lat.JOB.ark,$dir/lat.JOB.scp || exit 1;
 
   for id in $(seq $num_lat_jobs); do cat $dir/lat.$id.scp; done > $dir/lat.scp
 fi
@@ -307,6 +317,12 @@ chain_supervision_all_opts="--lattice-input=true --frame-subsampling-factor=$ali
 [ ! -z $left_tolerance ] && \
   chain_supervision_all_opts="$chain_supervision_all_opts --left-tolerance=$left_tolerance"
 
+[ ! -z $lattice_lm_scale ] && \
+  chain_supervision_all_opts="$chain_supervision_all_opts --lm-scale=$lattice_lm_scale"
+
+[ ! -z $phone_insertion_penalty ] && \
+  chain_supervision_all_opts="$chain_supervision_all_opts --phone-ins-penalty=$phone_insertion_penalty"
+
 echo $left_context > $dir/info/left_context
 echo $right_context > $dir/info/right_context
 echo $left_context_initial > $dir/info/left_context_initial
@@ -385,6 +401,7 @@ if [ $stage -le 4 ]; then
     utils/filter_scp.pl $sdata/JOB/utt2spk $dir/lat.scp \| \
     lattice-align-phones --replace-output-symbols=true $latdir/final.mdl scp:- ark:- \| \
     chain-get-supervision $chain_supervision_all_opts \
+      --weight=$egs_weight \
       $chaindir/tree $chaindir/0.trans_mdl ark:- ark:- \| \
     nnet3-chain-get-egs $ivector_opts --srand=\$[JOB+$srand] $egs_opts \
       --num-frames-overlap=$frames_overlap_per_eg \
diff --git a/src/chain/chain-supervision.cc b/src/chain/chain-supervision.cc
index b5597b15667..35489ca5e22 100644
--- a/src/chain/chain-supervision.cc
+++ b/src/chain/chain-supervision.cc
@@ -19,6 +19,7 @@
 
 #include "chain/chain-supervision.h"
 #include "lat/lattice-functions.h"
+#include "lat/push-lattice.h"
 #include "util/text-utils.h"
 #include "hmm/hmm-utils.h"
 #include <numeric>
@@ -142,9 +143,9 @@ bool ProtoSupervision::operator == (const ProtoSupervision &other) const {
           fst::Equal(fst, other.fst));
 }
 
-bool PhoneLatticeToProtoSupervision(const SupervisionOptions &opts,
-                                    const CompactLattice &lat,
-                                    ProtoSupervision *proto_supervision) {
+bool PhoneLatticeToProtoSupervisionInternal(const SupervisionOptions &opts,
+                                            const CompactLattice &lat,
+                                          ProtoSupervision *proto_supervision) {
   opts.Check();
   if (lat.NumStates() == 0) {
     KALDI_WARN << "Empty lattice provided";
@@ -176,9 +177,10 @@ bool PhoneLatticeToProtoSupervision(const SupervisionOptions &opts,
         return false;
       }
       proto_supervision->fst.AddArc(state,
-                                    fst::StdArc(phone, phone,
-                                                fst::TropicalWeight::One(),
-                                                lat_arc.nextstate));
+        fst::StdArc(phone, phone,
+                    fst::TropicalWeight(lat_arc.weight.Weight().Value1()
+                                      * opts.lm_scale + opts.phone_ins_penalty),
+                    lat_arc.nextstate));
       int32 t_begin = std::max<int32>(0, (state_time - opts.left_tolerance)),
               t_end = std::min<int32>(num_frames,
                                       (next_state_time + opts.right_tolerance)),
@@ -189,7 +191,8 @@ bool PhoneLatticeToProtoSupervision(const SupervisionOptions &opts,
       proto_supervision->allowed_phones[t_subsampled].push_back(phone);
     }
     if (lat.Final(state) != CompactLatticeWeight::Zero()) {
-      proto_supervision->fst.SetFinal(state, fst::TropicalWeight::One());
+      proto_supervision->fst.SetFinal(state, fst::TropicalWeight(
+                           lat.Final(state).Weight().Value1() * opts.lm_scale));
       if (state_times[state] != num_frames) {
         KALDI_WARN << "Time of final state " << state << " in lattice is "
                    << "not equal to number of frames " << num_frames
@@ -207,6 +210,16 @@ bool PhoneLatticeToProtoSupervision(const SupervisionOptions &opts,
   return true;
 }
 
+bool PhoneLatticeToProtoSupervision(const SupervisionOptions &opts,
+                                    const CompactLattice &lat,
+                                    ProtoSupervision *proto_supervision) {
+  if (!PhoneLatticeToProtoSupervisionInternal(opts, lat, proto_supervision))
+    return false;
+  if (opts.lm_scale != 0.0)
+    fst::Push(&(proto_supervision->fst),
+              fst::REWEIGHT_TO_INITIAL, fst::kDelta, true);
+  return true;
+}
 
 bool TimeEnforcerFst::GetArc(StateId s, Label ilabel, fst::StdArc* oarc) {
   // the following call will do the range-check on 'ilabel'.
diff --git a/src/chain/chain-supervision.h b/src/chain/chain-supervision.h
index a94f68ade90..ce755f0cb63 100644
--- a/src/chain/chain-supervision.h
+++ b/src/chain/chain-supervision.h
@@ -50,10 +50,16 @@ struct SupervisionOptions {
   int32 left_tolerance;
   int32 right_tolerance;
   int32 frame_subsampling_factor;
+  BaseFloat weight;
+  BaseFloat lm_scale;
+  BaseFloat phone_ins_penalty;
 
   SupervisionOptions(): left_tolerance(5),
                         right_tolerance(5),
-                        frame_subsampling_factor(1) { }
+                        frame_subsampling_factor(1),
+                        weight(1.0),
+                        lm_scale(0.0),
+                        phone_ins_penalty(0.0) { }
 
   void Register(OptionsItf *opts) {
     opts->Register("left-tolerance", &left_tolerance, "Left tolerance for "
@@ -65,6 +71,13 @@ struct SupervisionOptions {
                    "frame-rate of the original alignment.  Applied after "
                    "left-tolerance and right-tolerance are applied (so they are "
                    "in terms of the original num-frames.");
+    opts->Register("weight", &weight,
+                   "Use this to set the supervision weight for training");
+    opts->Register("lm-scale", &lm_scale, "The scale with which the graph/lm "
+                    "weights from the phone lattice are included in the "
+                    "supervision fst.");
+    opts->Register("phone-ins-penalty", &phone_ins_penalty,
+                   "The penalty to penalize longer paths");
   }
   void Check() const;
 };