Merge branch 'master' of github.com:kaldi-asr/kaldi

Author: vimalmanohar

.gitignore

@@ -83,6 +83,7 @@ GSYMS
 /tools/ATLAS/
 /tools/atlas3.8.3.tar.gz
 /tools/irstlm/
+/tools/mitlm/
 /tools/openfst
 /tools/openfst-1.3.2.tar.gz
 /tools/openfst-1.3.2/

egs/babel/s5d/conf/lang/404-georgian.FLP.official.conf

@@ -75,8 +75,8 @@ unsup_data_list=./conf/lists/404-georgian/untranscribed-training.list
 unsup_nj=32
 
 
-lexicon_file=
-lexiconFlags="--romanized --oov <unk>"
+lexicon_file=/export/corpora/LDC/LDC2016S12/IARPA_BABEL_OP3_404/conversational/reference_materials/lexicon.txt
+lexiconFlags=" --romanized --oov <unk>"
 
 
 

egs/babel/s5d/local/make_L_align.sh

@@ -34,18 +34,24 @@ tmpdir=$1
 dir=$2
 outdir=$3
 
+for f in  $dir/phones/optional_silence.txt $dir/phones.txt $dir/words.txt ; do
+  [ ! -f $f ] &&  echo "$0: The file $f must exist!" exit 1
+fi
+
 silphone=`cat $dir/phones/optional_silence.txt` || exit 1;
 
+if [ ! -f $tmpdir/lexicon.txt ] && [ ! -f $tmpdir/lexiconp.txt ] ; then
+  echo "$0: At least one of the files $tmpdir/lexicon.txt or $tmpdir/lexiconp.txt must exist" >&2
+  exit 1
+fi
+
 # Create lexicon with alignment info
 if  [ -f $tmpdir/lexicon.txt ] ; then
   cat $tmpdir/lexicon.txt | \
     awk '{printf("%s #1 ", $1); for (n=2; n <= NF; n++) { printf("%s ", $n); } print "#2"; }'
-elif [ -f $tmpdir/lexiconp.txt ] ;  then
+else
   cat $tmpdir/lexiconp.txt | \
     awk '{printf("%s #1 ", $1); for (n=3; n <= NF; n++) { printf("%s ", $n); } print "#2"; }'
-else
-  echo "Neither $tmpdir/lexicon.txt nor $tmpdir/lexiconp.txt does not exist"
-  exit 1
 fi | utils/make_lexicon_fst.pl - 0.5 $silphone | \
 fstcompile --isymbols=$dir/phones.txt --osymbols=$dir/words.txt \
   --keep_isymbols=false --keep_osymbols=false | \

egs/fisher_callhome_spanish/s5/local/merge_lexicons.py

@@ -1,11 +1,11 @@
-#!/usr/bin/env python
 # Copyright 2014  Gaurav Kumar.   Apache 2.0
+#    2018  Saikiran Valluri, GoVivace inc., Avaaya
+#!/usr/bin/env python
 # -*- coding: utf-8 -*-
 #
 # Merges unique words from Spanish Fisher, Gigaword and the LDC spanish lexicon
-
 from __future__ import print_function
-import sys
+import sys, re
 import json
 import codecs
 import operator
@@ -17,6 +17,7 @@
 uw_gigaword = tmpdir + "/es_wordlist.json"
 uw_LDC = ldc_lexicon + "/callhome_spanish_lexicon_970908/preferences"
 
+filtered_letters = re.compile(u'[¡¥ª°º¿àçèëìîôö0123456789]')
 merged_lexicon = []
 # All three lexicons are in different formats
 # First add the data from lexicon_fisher (A) into the dictionary
@@ -55,7 +56,8 @@
 ltuples = sorted(merged_lexicon)
 
 for item in ltuples:
-    lf.write(item + "\n")
+    if not item==u'ñ' and not re.search(filtered_letters, item):
+        lf.write(item + "\n")
 
 lf.close()
 

egs/gale_arabic/s5b/RESULTS

@@ -2,13 +2,7 @@
 # This file is generated using local/split_wer.sh $galeData  //galeData is a local folder to keep intermediate gale data
 # look at the end of run.sh in the same folder 
 ##
-##### RESULTS generated by amali at 2017-01-01-08-05-59
-
 Report Results WER:
-%WER 9.50 [ 2124 / 22363, 160 ins, 275 del, 1689 sub ] exp/chain_cleaned/tdnn_lstm1a_sp_bi/decode/wer_report_9
-%WER 10.72 [ 2398 / 22363, 163 ins, 313 del, 1922 sub ] exp/chain_cleaned/tdnn1b_sp_bi/decode/wer_report_9
-%WER 12.04 [ 2693 / 22363, 226 ins, 271 del, 2196 sub ] exp/nnet3_cleaned/lstm_ld5_sp/decode/wer_report_9
-%WER 12.29 [ 2749 / 22363, 273 ins, 266 del, 2210 sub ] exp/nnet3_cleaned/tdnn_sp/decode/wer_report_10
 %WER 17.82 [ 3986 / 22363, 315 ins, 618 del, 3053 sub ] exp/sgmm_5a_mmi_b0.1/decode/wer_report_12
 %WER 18.15 [ 4059 / 22363, 335 ins, 589 del, 3135 sub ] exp/sgmm_5a_mmi_b0.1/decode4/wer_report_11
 %WER 18.42 [ 4119 / 22363, 346 ins, 590 del, 3183 sub ] exp/sgmm_5a_mmi_b0.1/decode3/wer_report_11
@@ -27,10 +21,6 @@ Report Results WER:
 %WER 25.66 [ 5738 / 22363, 478 ins, 838 del, 4422 sub ] exp/tri2a/decode/wer_report_14
 %WER 26.38 [ 5900 / 22363, 435 ins, 929 del, 4536 sub ] exp/tri1/decode/wer_report_15
 Conversational Results WER:
-%WER 21.59 [ 10213 / 47305, 944 ins, 3092 del, 6177 sub ] exp/chain_cleaned/tdnn_lstm1a_sp_bi/decode/wer_conversational_9
-%WER 24.77 [ 11716 / 47305, 1098 ins, 3579 del, 7039 sub ] exp/chain_cleaned/tdnn1b_sp_bi/decode/wer_conversational_9
-%WER 26.78 [ 12670 / 47305, 1741 ins, 2434 del, 8495 sub ] exp/nnet3_cleaned/lstm_ld5_sp/decode/wer_conversational_9
-%WER 27.55 [ 13032 / 47305, 1800 ins, 2666 del, 8566 sub ] exp/nnet3_cleaned/tdnn_sp/decode/wer_conversational_11
 %WER 34.10 [ 16133 / 47305, 1903 ins, 3245 del, 10985 sub ] exp/sgmm_5a_mmi_b0.1/decode/wer_conversational_11
 %WER 34.81 [ 16466 / 47305, 2077 ins, 3037 del, 11352 sub ] exp/sgmm_5a_mmi_b0.1/decode4/wer_conversational_10
 %WER 35.19 [ 16648 / 47305, 1933 ins, 3264 del, 11451 sub ] exp/sgmm_5a_mmi_b0.1/decode3/wer_conversational_11
@@ -49,10 +39,6 @@ Conversational Results WER:
 %WER 45.92 [ 21724 / 47305, 1995 ins, 5213 del, 14516 sub ] exp/tri2a/decode/wer_conversational_14
 %WER 46.86 [ 22166 / 47305, 2212 ins, 4819 del, 15135 sub ] exp/tri1/decode/wer_conversational_13
 Combined Results for Reports and Conversational WER:
-%WER 17.64 [ 12286 / 69668, 1310 ins, 2807 del, 8169 sub ] exp/chain_cleaned/tdnn_lstm1a_sp_bi/decode/wer_8
-%WER 20.26 [ 14114 / 69668, 1261 ins, 3892 del, 8961 sub ] exp/chain_cleaned/tdnn1b_sp_bi/decode/wer_9
-%WER 22.05 [ 15363 / 69668, 1967 ins, 2705 del, 10691 sub ] exp/nnet3_cleaned/lstm_ld5_sp/decode/wer_9
-%WER 22.66 [ 15786 / 69668, 2047 ins, 2955 del, 10784 sub ] exp/nnet3_cleaned/tdnn_sp/decode/wer_11
 %WER 28.89 [ 20127 / 69668, 2244 ins, 3829 del, 14054 sub ] exp/sgmm_5a_mmi_b0.1/decode/wer_11
 %WER 29.48 [ 20541 / 69668, 2243 ins, 3860 del, 14438 sub ] exp/sgmm_5a_mmi_b0.1/decode4/wer_11
 %WER 29.81 [ 20767 / 69668, 2279 ins, 3854 del, 14634 sub ] exp/sgmm_5a_mmi_b0.1/decode3/wer_11
@@ -65,8 +51,30 @@ Combined Results for Reports and Conversational WER:
 %WER 32.36 [ 22542 / 69668, 2156 ins, 4184 del, 16202 sub ] exp/tri2b_mmi/decode_it4/wer_11
 %WER 32.50 [ 22640 / 69668, 2393 ins, 3956 del, 16291 sub ] exp/tri2b_mmi/decode_it3/wer_11
 %WER 32.79 [ 22847 / 69668, 2407 ins, 4760 del, 15680 sub ] exp/tri2b_mpe/decode_it3/wer_13
+# WER with train_sat_basis
+%WER 33.35 [ 23233 / 69668, 2385 ins, 5274 del, 15574 sub ] exp/tri3b/decode/wer_16_0.5
+# WER with train_sat
 %WER 33.61 [ 23413 / 69668, 2817 ins, 4577 del, 16019 sub ] exp/tri3b/decode/wer_17
 %WER 35.73 [ 24894 / 69668, 2630 ins, 4944 del, 17320 sub ] exp/tri3b/decode.si/wer_15
 %WER 36.17 [ 25196 / 69668, 2429 ins, 5393 del, 17374 sub ] exp/tri2b/decode/wer_16
 %WER 39.42 [ 27462 / 69668, 2473 ins, 6051 del, 18938 sub ] exp/tri2a/decode/wer_14
 %WER 40.35 [ 28113 / 69668, 2713 ins, 5635 del, 19765 sub ] exp/tri1/decode/wer_13
+
+
+# Effect of GMM seed model (tri2b instead of tri3b).  Using tri3b give a slightly better result
+# as compared to using tri2b as seed.
+%WER 16.66 [ 11610 / 69668, 1233 ins, 2747 del, 7630 sub ] exp/chain/tdnn_1a_3b_sp/decode_test/wer_10_0.0
+%WER 16.71 [ 11642 / 69668, 1145 ins, 2908 del, 7589 sub ] exp/chain/tdnn_1a_2b_sp/decode_test/wer_9_0.0
+
+# Effect of Tree-size (3500, 4500, 7000, 11000)
+%WER 16.66 [ 11610 / 69668, 1233 ins, 2747 del, 7630 sub ] exp/chain/tdnn_1a_3500_sp/decode_test/wer_10_0.0
+%WER 16.59 [ 11557 / 69668, 1234 ins, 2646 del, 7677 sub ] exp/chain/tdnn_1a_4500_sp/decode_test/wer_10_0.0
+%WER 16.47 [ 11474 / 69668, 1421 ins, 2207 del, 7846 sub ] exp/chain/tdnn_1a_7000_sp/decode_test/wer_9_0.0
+%WER 16.62 [ 11580 / 69668, 1164 ins, 2789 del, 7627 sub ] exp/chain/tdnn_1a_11000_sp/decode_test/wer_10_0.0
+
+# Effect of l2-regularization on the output with tree-size=7000. l2 on the output (0.005,0.002)
+%WER 16.54 [ 11522 / 69668, 1123 ins, 2739 del, 7660 sub ] exp/chain/tdnn_1a_7000_005_sp/decode_test/wer_9_0.5
+%WER 16.47 [ 11474 / 69668, 1421 ins, 2207 del, 7846 sub ] exp/chain/tdnn_1a_7000_002_sp/decode_test/wer_9_0.0
+
+#current best 'chain' models  (see local/chain/tuning/run_tdnn_1a.sh)
+%WER 16.47 [ 11474 / 69668, 1421 ins, 2207 del, 7846 sub ] exp/chain/tdnn_1a_sp/decode_test/wer_9_0.0

egs/gale_arabic/s5b/cmd.sh

@@ -10,6 +10,6 @@
 # conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
 # or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
 
-export train_cmd="queue.pl --mem 2G"
-export decode_cmd="queue.pl --mem 4G"
-export mkgraph_cmd="queue.pl --mem 8G"
+export train_cmd="retry.pl queue.pl --mem 2G"
+export decode_cmd="retry.pl queue.pl --mem 4G"
+export mkgraph_cmd="retry.pl queue.pl --mem 8G"

egs/gale_arabic/s5b/local/chain/compare_wer.sh

@@ -0,0 +1,72 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/cnn{1a,1b}
+
+# ./local/chain/compare_wer.sh exp/chain/cnn1a
+# System                          cnn1a
+# WER                              0.61
+# CER                              0.15
+# Final train prob              -0.0377
+# Final valid prob              -0.0380
+# Final train prob (xent)       -0.0830
+# Final valid prob (xent)       -0.0838
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/cnn{1a,1b}"
+  exit 1
+fi
+
+echo "# $0 $*"
+used_epochs=false
+
+echo -n "# System                     "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+echo -n "# WER                        "
+for x in $*; do
+  wer=$(cat $x/decode_test/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "# CER                        "
+for x in $*; do
+  cer=$(cat $x/decode_test/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
+if $used_epochs; then
+  exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
+fi
+
+echo -n "# Final train prob           "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob           "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train prob (xent)    "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob (xent)    "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo

egs/gale_arabic/s5b/local/chain/run_chain_common.sh

@@ -0,0 +1,82 @@
+#!/bin/bash
+
+# this script has common stages shared across librispeech chain recipes.
+# It generates a new topology in a new lang directory, gets the alignments as
+# lattices, and builds a tree for the new topology
+set -e
+
+stage=11
+
+# input directory names. These options are actually compulsory, and they have
+# been named for convenience
+gmm_dir=
+ali_dir=
+lores_train_data_dir=
+
+num_leaves=6000
+
+# output directory names. They are also compulsory.
+lang=
+lat_dir=
+tree_dir=
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+[ -z $lang ] && echo "Set --lang, this specifies the new lang directory which will have the new topology" && exit 1;
+[ -z $lat_dir ] && echo "Set --lat-dir, this specifies the experiment directory to store lattice" && exit 1;
+[ -z $tree_dir ] && echo "Set --tree-dir, this specifies the directory to store new tree " && exit 1;
+
+for f in $gmm_dir/final.mdl $ali_dir/ali.1.gz $lores_train_data_dir/feats.scp; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 11 ]; then
+  echo "$0: creating lang directory with one state per phone."
+  # 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 [ $stage -le 12 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat ${ali_dir}/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" ${lores_train_data_dir} \
+    $lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 13 ]; 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.
+  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" $num_leaves ${lores_train_data_dir} $lang $ali_dir $tree_dir
+fi
+
+exit 0;