From ffda51ea5c46fe8a92a5cd23a9b1eca0d36017f5 Mon Sep 17 00:00:00 2001
From: Desh Raj <r.desh26@gmail.com>
Date: Fri, 2 Aug 2019 13:57:22 -0400
Subject: [PATCH] [scripts,egs] Replace LDA layer with delta and delta-delta
 features (#3490)

WER is about the same but this is simpler to implement and more standard.
---
 .../s5/local/chain/run_tdnn.sh                |   2 +-
 .../s5/local/chain/tuning/run_tdnn_1j.sh      | 301 +++++++++++++++
 egs/swbd/s5c/local/chain/run_tdnn.sh          |   2 +-
 .../s5c/local/chain/tuning/run_tdnn_7r.sh     | 266 +++++++++++++
 egs/wsj/s5/local/chain/run_tdnn.sh            |   2 +-
 egs/wsj/s5/local/chain/tuning/run_tdnn_1h.sh  | 349 ++++++++++++++++++
 egs/wsj/s5/steps/libs/nnet3/xconfig/parser.py |   3 +-
 .../libs/nnet3/xconfig/trivial_layers.py      |  71 ++++
 8 files changed, 992 insertions(+), 4 deletions(-)
 create mode 100755 egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1j.sh
 create mode 100755 egs/swbd/s5c/local/chain/tuning/run_tdnn_7r.sh
 create mode 100755 egs/wsj/s5/local/chain/tuning/run_tdnn_1h.sh

diff --git a/egs/mini_librispeech/s5/local/chain/run_tdnn.sh b/egs/mini_librispeech/s5/local/chain/run_tdnn.sh
index deb68d515d2..05a7c2d345b 120000
--- a/egs/mini_librispeech/s5/local/chain/run_tdnn.sh
+++ b/egs/mini_librispeech/s5/local/chain/run_tdnn.sh
@@ -1 +1 @@
-tuning/run_tdnn_1i.sh
\ No newline at end of file
+tuning/run_tdnn_1j.sh
\ No newline at end of file
diff --git a/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1j.sh b/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1j.sh
new file mode 100755
index 00000000000..7a6604f9773
--- /dev/null
+++ b/egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1j.sh
@@ -0,0 +1,301 @@
+#!/bin/bash
+
+# 1j is as 1i but replaces the LDA layer at the input of the
+# network with delta and delta-delta features.
+# Below, 1j2 is just a different rerun of 1j with different
+# --affix option, to give some idea of the run-to-run variation.
+
+# local/chain/compare_wer.sh --online exp/chain/tdnn1i_sp exp/chain/tdnn1j_sp exp/chain/tdnn1j2_sp
+# System                tdnn1i_sp tdnn1j_sp tdnn1j2_sp
+#WER dev_clean_2 (tgsmall)      10.73     10.78     10.79
+#             [online:]         10.81     10.80     10.80
+#WER dev_clean_2 (tglarge)       7.54      7.43      7.54
+#             [online:]          7.59      7.37      7.55
+# Final train prob        -0.0621   -0.0620   -0.0615
+# Final valid prob        -0.0787   -0.0787   -0.0785
+# Final train prob (xent)   -1.4603   -1.4363   -1.4438
+# Final valid prob (xent)   -1.5847   -1.5629   -1.5712
+# Num-params                 5210944   5210944   5210944
+
+# steps/info/chain_dir_info.pl exp/chain/tdnn1j_sp
+# exp/chain/tdnn1j_sp: num-iters=34 nj=2..5 num-params=5.2M dim=40+100->2336 combine=-0.049->-0.046 (over 4) xent:train/valid[21,33,final]=(-1.40,-1.18,-1.17/-1.56,-1.38,-1.37) logprob:train/valid[21,33,final]=(-0.064,-0.052,-0.047/-0.089,-0.085,-0.078)
+
+# 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=1j   # 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
+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)
+
+  tdnn_opts="l2-regularize=0.03"
+  tdnnf_opts="l2-regularize=0.03 bypass-scale=0.66"
+  linear_opts="l2-regularize=0.03 orthonormal-constraint=-1.0"
+  prefinal_opts="l2-regularize=0.03"
+  output_opts="l2-regularize=0.015"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # this takes the MFCCs and generates filterbank coefficients.  The MFCCs
+  # are more compressible so we prefer to dump the MFCCs to disk rather
+  # than filterbanks.
+  idct-layer name=idct input=input dim=40 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat
+  batchnorm-component name=batchnorm0 input=idct
+  spec-augment-layer name=spec-augment freq-max-proportion=0.5 time-zeroed-proportion=0.2 time-mask-max-frames=20
+  
+  delta-layer name=delta input=spec-augment
+  no-op-component name=input2 input=Append(delta, Scale(0.4, ReplaceIndex(ivector, t, 0)))
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 $tdnn_opts dim=768 input=input2
+  tdnnf-layer name=tdnnf2 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf3 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf4 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf5 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=0
+  tdnnf-layer name=tdnnf6 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf7 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf8 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf9 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf10 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf11 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf12 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf13 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+  linear-component name=prefinal-l dim=192 $linear_opts
+
+  ## adding the layers for chain branch
+  prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts small-dim=192 big-dim=768
+  output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+  # adding the layers for xent branch
+  prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts small-dim=192 big-dim=768
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+
+if [ $stage -le 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 $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.0 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=20 \
+    --trainer.frames-per-iter=3000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=5 \
+    --trainer.optimization.initial-effective-lrate=0.002 \
+    --trainer.optimization.final-effective-lrate=0.0002 \
+    --trainer.num-chunk-per-minibatch=128,64 \
+    --egs.chunk-width=$chunk_width \
+    --egs.dir="$common_egs_dir" \
+    --egs.opts="--frames-overlap-per-eg 0" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --tree-dir=$tree_dir \
+    --lat-dir=$lat_dir \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 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 \
+          --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/swbd/s5c/local/chain/run_tdnn.sh b/egs/swbd/s5c/local/chain/run_tdnn.sh
index 2f256c1a5aa..af535373188 120000
--- a/egs/swbd/s5c/local/chain/run_tdnn.sh
+++ b/egs/swbd/s5c/local/chain/run_tdnn.sh
@@ -1 +1 @@
-tuning/run_tdnn_7q.sh
\ No newline at end of file
+tuning/run_tdnn_7r.sh
\ No newline at end of file
diff --git a/egs/swbd/s5c/local/chain/tuning/run_tdnn_7r.sh b/egs/swbd/s5c/local/chain/tuning/run_tdnn_7r.sh
new file mode 100755
index 00000000000..10e4ae1ddca
--- /dev/null
+++ b/egs/swbd/s5c/local/chain/tuning/run_tdnn_7r.sh
@@ -0,0 +1,266 @@
+#!/bin/bash
+
+# 7r is as 7q but replaces the LDA layer at the input of the
+# network with traditional delta and delta-delta features.
+
+# local/chain/compare_wer_general.sh --rt03 tdnn7q_sp tdnn7r_sp
+# System                tdnn7q_sp tdnn7r_sp
+# WER on train_dev(tg)      11.87     11.94
+# WER on train_dev(fg)      11.00     11.05
+# WER on eval2000(tg)        14.1      14.1
+# WER on eval2000(fg)        12.7      12.8
+# WER on rt03(tg)            17.3      17.2
+# WER on rt03(fg)            15.0      15.0
+# Final train prob         -0.057    -0.056
+# Final valid prob         -0.072    -0.072
+# Final train prob (xent)        -0.880    -0.894
+# Final valid prob (xent)       -0.8997   -0.9036
+# Num-parameters               18705712  18705712
+
+# steps/info/chain_dir_info.pl exp/chain/tdnn7r_sp
+# exp/chain/tdnn7r_sp: num-iters=394 nj=3..16 num-params=18.7M dim=40+100->6040 combine=-0.054->-0.054 (over 5) xent:train/valid[261,393,final]=(-1.12,-0.897,-0.894/-1.10,-0.905,-0.904) logprob:train/valid[261,393,final]=(-0.084,-0.057,-0.056/-0.089,-0.072,-0.072)
+
+set -e
+
+# configs for 'chain'
+stage=0
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+affix=7r
+if [ -e data/rt03 ]; then maybe_rt03=rt03; else maybe_rt03= ; fi
+
+decode_iter=
+decode_nj=50
+
+# training options
+frames_per_eg=150,110,100
+remove_egs=false
+common_egs_dir=
+xent_regularize=0.1
+dropout_schedule='0,0@0.20,0.5@0.50,0'
+
+test_online_decoding=false  # if true, it will run the last decoding stage.
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+suffix=
+$speed_perturb && suffix=_sp
+dir=exp/chain/tdnn${affix}${suffix}
+
+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 8" if you have already
+# run those things.
+
+train_set=train_nodup$suffix
+ali_dir=exp/tri4_ali_nodup$suffix
+treedir=exp/chain/tri5_7d_tree$suffix
+lang=data/lang_chain_2y
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage \
+  --speed-perturb $speed_perturb \
+  --generate-alignments $speed_perturb || exit 1;
+
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the LF-MMI training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri4 exp/tri4_lats_nodup$suffix
+  rm exp/tri4_lats_nodup$suffix/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 [ $stage -le 11 ]; then
+  # Build a tree using our new topology. This is the critically different
+  # step compared with other recipes.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
+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)
+  affine_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
+  tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
+  linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
+  prefinal_opts="l2-regularize=0.01"
+  output_opts="l2-regularize=0.002"
+
+  mkdir -p $dir/configs
+
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  delta-layer name=delta
+  no-op-component name=input2 input=Append(delta, Scale(1.0, ReplaceIndex(ivector, t, 0)))
+
+  relu-batchnorm-dropout-layer name=tdnn1 $affine_opts dim=1536 input=input2
+  tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+  tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+  tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+  tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=0
+  tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+  linear-component name=prefinal-l dim=256 $linear_opts
+
+  prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
+  output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+  prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+if [ $stage -le 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/swbd-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
+  fi
+
+#    --cmd "queue.pl --config /home/dpovey/queue_conly.conf" \
+
+
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$train_cmd" \
+    --feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.0 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0 --constrained false" \
+    --egs.chunk-width $frames_per_eg \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 6 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.00025 \
+    --trainer.optimization.final-effective-lrate 0.000025 \
+    --trainer.max-param-change 2.0 \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri4_lats_nodup$suffix \
+    --dir $dir  || exit 1;
+
+fi
+
+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_sw1_tg $dir $dir/graph_sw1_tg
+fi
+
+
+graph_dir=$dir/graph_sw1_tg
+iter_opts=
+if [ ! -z $decode_iter ]; then
+  iter_opts=" --iter $decode_iter "
+fi
+if [ $stage -le 15 ]; then
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000 $maybe_rt03; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj $decode_nj --cmd "$decode_cmd" $iter_opts \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+          $graph_dir data/${decode_set}_hires \
+          $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) || touch $dir/.error &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in decoding"
+    exit 1
+  fi
+fi
+
+if $test_online_decoding && [ $stage -le 16 ]; 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/extractor $dir ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000 $maybe_rt03; do
+    (
+      # note: we just give it "$decode_set" as it only uses the wav.scp, the
+      # feature type does not matter.
+
+      steps/online/nnet3/decode.sh --nj $decode_nj --cmd "$decode_cmd" \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+         $graph_dir data/${decode_set}_hires \
+         ${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            ${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+    ) || touch $dir/.error &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in decoding"
+    exit 1
+  fi
+fi
+
+
+exit 0;
diff --git a/egs/wsj/s5/local/chain/run_tdnn.sh b/egs/wsj/s5/local/chain/run_tdnn.sh
index cebb2b84f16..3922170ac12 120000
--- a/egs/wsj/s5/local/chain/run_tdnn.sh
+++ b/egs/wsj/s5/local/chain/run_tdnn.sh
@@ -1 +1 @@
-tuning/run_tdnn_1g.sh
\ No newline at end of file
+tuning/run_tdnn_1h.sh
\ No newline at end of file
diff --git a/egs/wsj/s5/local/chain/tuning/run_tdnn_1h.sh b/egs/wsj/s5/local/chain/tuning/run_tdnn_1h.sh
new file mode 100755
index 00000000000..3e00048c9ae
--- /dev/null
+++ b/egs/wsj/s5/local/chain/tuning/run_tdnn_1h.sh
@@ -0,0 +1,349 @@
+#!/bin/bash
+
+# 1h is as 1g but replaces the LDA layer at the input of the
+# network with traditional delta and delta-delta features.
+
+# local/chain/compare_wer.sh exp/chain/tdnn1g_sp exp/chain/tdnn1h2_sp/ exp/chain/tdnn1h_sp/
+# System                tdnn1g_sp tdnn1h_sp
+#WER dev93 (tgpr)                6.87      6.75
+#WER dev93 (tg)                  6.66      6.70
+#WER dev93 (big-dict,tgpr)       4.77      4.60
+#WER dev93 (big-dict,fg)         4.29      4.40
+#WER eval92 (tgpr)               4.63      4.71
+#WER eval92 (tg)                 4.36      4.41
+#WER eval92 (big-dict,tgpr)      2.71      2.96
+#WER eval92 (big-dict,fg)        2.39      2.39
+# Final train prob        -0.0419   -0.0424
+# Final valid prob        -0.0540   -0.0534
+# Final train prob (xent)   -0.6582   -0.6583
+# Final valid prob (xent)   -0.7220   -0.7281
+# Num-params                 8364672   8364672
+
+# steps/info/chain_dir_info.pl exp/chain/tdnn1h_sp
+# exp/chain/tdnn1h_sp: num-iters=108 nj=2..8 num-params=8.4M dim=40+100->2880 combine=-0.044->-0.044 (over 2) xent:train/valid[71,107,final]=(-0.863,-0.648,-0.658/-0.932,-0.719,-0.728) logprob:train/valid[71,107,final]=(-0.065,-0.043,-0.042/-0.073,-0.055,-0.053)
+
+set -e -o pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+nj=30
+train_set=train_si284
+test_sets="test_dev93 test_eval92"
+gmm=tri4b        # this is the source gmm-dir that we'll use for alignments; it
+                 # should have alignments for the specified training data.
+
+num_threads_ubm=32
+
+nj_extractor=10
+# It runs a JOB with '-pe smp N', where N=$[threads*processes]
+num_threads_extractor=4
+num_processes_extractor=4
+
+nnet3_affix=       # affix for exp dirs, e.g. it was _cleaned in tedlium.
+
+# Options which are not passed through to run_ivector_common.sh
+affix=1h   #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+common_egs_dir=
+reporting_email=
+
+# LSTM/chain options
+train_stage=-10
+xent_regularize=0.1
+dropout_schedule='0,0@0.20,0.5@0.50,0'
+
+# 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
+
+# training options
+srand=0
+remove_egs=false
+
+#decode options
+test_online_decoding=false  # 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
+
+local/nnet3/run_ivector_common.sh \
+  --stage $stage --nj $nj \
+  --train-set $train_set --gmm $gmm \
+  --num-threads-ubm $num_threads_ubm \
+  --nj-extractor $nj_extractor \
+  --num-processes-extractor $num_processes_extractor \
+  --num-threads-extractor $num_threads_extractor \
+  --nnet3-affix "$nnet3_affix"
+
+
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
+dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+lores_train_data_dir=data/${train_set}_sp
+
+# note: you don't necessarily have to change the treedir name
+# each time you do a new experiment-- only if you change the
+# configuration in a way that affects the tree.
+tree_dir=exp/chain${nnet3_affix}/tree_a_sp
+# the 'lang' directory is created by this script.
+# If you create such a directory with a non-standard topology
+# you should probably name it differently.
+lang=data/lang_chain
+
+for f in $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $lores_train_data_dir/feats.scp $gmm_dir/final.mdl \
+    $ali_dir/ali.1.gz $gmm_dir/final.mdl; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+
+if [ $stage -le 12 ]; 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 [ $stage -le 13 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 14 ]; 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 15 ]; 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)
+  tdnn_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim-continuous=true"
+  tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
+  linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
+  prefinal_opts="l2-regularize=0.01"
+  output_opts="l2-regularize=0.005"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  idct-layer name=idct input=input dim=40 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat
+  delta-layer name=delta input=idct
+  no-op-component name=input2 input=Append(delta, Scale(1.0, ReplaceIndex(ivector, t, 0)))
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 $tdnn_opts dim=1024 input=input2
+  tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
+  tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
+  tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
+  tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=0
+  tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
+  linear-component name=prefinal-l dim=192 $linear_opts
+
+
+  prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1024 small-dim=192
+  output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+  prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1024 small-dim=192
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+
+if [ $stage -le 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/wsj-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/chain/train.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir=$train_ivector_dir \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.0 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=10 \
+    --trainer.frames-per-iter=5000000 \
+    --trainer.optimization.num-jobs-initial=2 \
+    --trainer.optimization.num-jobs-final=8 \
+    --trainer.optimization.initial-effective-lrate=0.0005 \
+    --trainer.optimization.final-effective-lrate=0.00005 \
+    --trainer.num-chunk-per-minibatch=128,64 \
+    --trainer.optimization.momentum=0.0 \
+    --egs.chunk-width=$chunk_width \
+    --egs.chunk-left-context=0 \
+    --egs.chunk-right-context=0 \
+    --egs.dir="$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 17 ]; then
+  # The reason we are using data/lang here, instead of $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/lang/check_phones_compatible.sh \
+    data/lang_test_tgpr/phones.txt $lang/phones.txt
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/lang_test_tgpr \
+    $tree_dir $tree_dir/graph_tgpr || exit 1;
+
+  utils/lang/check_phones_compatible.sh \
+    data/lang_test_bd_tgpr/phones.txt $lang/phones.txt
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/lang_test_bd_tgpr \
+    $tree_dir $tree_dir/graph_bd_tgpr || exit 1;
+fi
+
+if [ $stage -le 18 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      data_affix=$(echo $data | sed s/test_//)
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      for lmtype in tgpr bd_tgpr; do
+        steps/nnet3/decode.sh \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --extra-left-context 0 --extra-right-context 0 \
+          --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 \
+          $tree_dir/graph_${lmtype} data/${data}_hires ${dir}/decode_${lmtype}_${data_affix} || exit 1
+      done
+      steps/lmrescore.sh \
+        --self-loop-scale 1.0 \
+        --cmd "$decode_cmd" data/lang_test_{tgpr,tg} \
+        data/${data}_hires ${dir}/decode_{tgpr,tg}_${data_affix} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_bd_{tgpr,fgconst} \
+       data/${data}_hires ${dir}/decode_${lmtype}_${data_affix}{,_fg} || 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 19 ]; 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 $test_sets; do
+    (
+      data_affix=$(echo $data | sed s/test_//)
+      nspk=$(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.
+      for lmtype in tgpr bd_tgpr; do
+        steps/online/nnet3/decode.sh \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj $nspk --cmd "$decode_cmd" \
+          $tree_dir/graph_${lmtype} data/${data} ${dir}_online/decode_${lmtype}_${data_affix} || exit 1
+      done
+      steps/lmrescore.sh \
+        --self-loop-scale 1.0 \
+        --cmd "$decode_cmd" data/lang_test_{tgpr,tg} \
+        data/${data}_hires ${dir}_online/decode_{tgpr,tg}_${data_affix} || exit 1
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_bd_{tgpr,fgconst} \
+       data/${data}_hires ${dir}_online/decode_${lmtype}_${data_affix}{,_fg} || 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/wsj/s5/steps/libs/nnet3/xconfig/parser.py b/egs/wsj/s5/steps/libs/nnet3/xconfig/parser.py
index ee39b9312bc..5e21c4c0274 100644
--- a/egs/wsj/s5/steps/libs/nnet3/xconfig/parser.py
+++ b/egs/wsj/s5/steps/libs/nnet3/xconfig/parser.py
@@ -85,7 +85,8 @@
         'scale-component':  xlayers.XconfigPerElementScaleComponent,
         'dim-range-component': xlayers.XconfigDimRangeComponent,
         'offset-component':  xlayers.XconfigPerElementOffsetComponent,
-        'combine-feature-maps-layer': xlayers.XconfigCombineFeatureMapsLayer
+        'combine-feature-maps-layer': xlayers.XconfigCombineFeatureMapsLayer,
+        'delta-layer': xlayers.XconfigDeltaLayer
 }
 
 # Turn a config line and a list of previous layers into
diff --git a/egs/wsj/s5/steps/libs/nnet3/xconfig/trivial_layers.py b/egs/wsj/s5/steps/libs/nnet3/xconfig/trivial_layers.py
index 00d6028ab09..4afea78ad3f 100644
--- a/egs/wsj/s5/steps/libs/nnet3/xconfig/trivial_layers.py
+++ b/egs/wsj/s5/steps/libs/nnet3/xconfig/trivial_layers.py
@@ -186,6 +186,77 @@ def _generate_config(self):
         return configs
 
 
+class XconfigDeltaLayer(XconfigLayerBase):
+    """This class is for parsing lines like
+     'delta-layer name=delta input=idct'
+    which appends the central frame with the delta features
+    (i.e. -1,0,1 since scale equals 1) and delta-delta features 
+    (i.e. 1,0,-2,0,1), and then applies batchnorm to it.
+
+    Parameters of the class, and their defaults:
+      input='[-1]'             [Descriptor giving the input of the layer]
+    """
+    def __init__(self, first_token, key_to_value, prev_names=None):
+        XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
+
+    def set_default_configs(self):
+        self.config = {'input': '[-1]'}
+
+    def check_configs(self):
+        pass
+
+    def output_name(self, auxiliary_output=None):
+        assert auxiliary_output is None
+        return self.name
+
+    def output_dim(self, auxiliary_output=None):
+        assert auxiliary_output is None
+        input_dim = self.descriptors['input']['dim']
+        return (3*input_dim)
+
+    def get_full_config(self):
+        ans = []
+        config_lines = self._generate_config()
+
+        for line in config_lines:
+            for config_name in ['ref', 'final']:
+                # we do not support user specified matrices in this layer
+                # so 'ref' and 'final' configs are the same.
+                ans.append((config_name, line))
+        return ans
+
+    def _generate_config(self):
+        # by 'descriptor_final_string' we mean a string that can appear in
+        # config-files, i.e. it contains the 'final' names of nodes.
+        input_desc = self.descriptors['input']['final-string']
+        input_dim = self.descriptors['input']['dim']
+        output_dim = self.output_dim()
+
+        configs = []
+        line = ('dim-range-node name={0}_copy1 input-node={0} dim={1} dim-offset=0'.format(
+            input_desc, input_dim))
+        configs.append(line)
+        line = ('dim-range-node name={0}_copy2 input-node={0} dim={1} dim-offset=0'.format(
+            input_desc, input_dim))
+        configs.append(line)
+
+        line = ('component name={0}_2 type=NoOpComponent dim={1}'.format(
+            input_desc, output_dim))
+        configs.append(line)
+        line = ('component-node name={0}_2 component={0}_2 input=Append(Offset({0},0),'
+            ' Sum(Offset(Scale(-1.0,{0}_copy1),-1), Offset({0},1)), Sum(Offset({0},-2), Offset({0},2),' 
+            ' Offset(Scale(-2.0,{0}_copy2),0)))'.format(input_desc))
+        configs.append(line)
+        
+        line = ('component name={0} type=BatchNormComponent dim={1}'.format(
+            self.name, output_dim))
+        configs.append(line)
+        line = ('component-node name={0} component={0} input={1}_2'.format(
+            self.name, input_desc))
+        configs.append(line)
+        return configs
+
+
 class XconfigLinearComponent(XconfigLayerBase):
     """This class is for parsing lines like
      'linear-component name=linear1 dim=1024 input=Append(-3,0,3)'