wsjcam0_format_data.sh 2.89 KB
#!/bin/bash

# Copyright 2013 MERL (author: Felix Weninger)
# Contains some code by Microsoft Corporation, Johns Hopkins University (author: Daniel Povey)

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#  http://www.apache.org/licenses/LICENSE-2.0
#
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
# MERCHANTABLITY OR NON-INFRINGEMENT.
# See the Apache 2 License for the specific language governing permissions and
# limitations under the License.


. ./path.sh || exit 1;

echo "Preparing train and test data"
srcdir=data/local/data
lmdir=data/local/nist_lm
tmpdir=data/local/lm_tmp
lexicon=data/local/lang_tmp/lexiconp.txt
mkdir -p $tmpdir

for x in si_tr si_dt; do
  mkdir -p data/$x
  cp $srcdir/${x}_wav.scp data/$x/wav.scp || exit 1;
  cp $srcdir/$x.txt data/$x/text || exit 1;
  cp $srcdir/$x.spk2utt data/$x/spk2utt || exit 1;
  cp $srcdir/$x.utt2spk data/$x/utt2spk || exit 1;
  #utils/filter_scp.pl data/$x/spk2utt $srcdir/spk2gender > data/$x/spk2gender || exit 1;
done


# Next, for each type of language model, create the corresponding FST
# and the corresponding lang_test_* directory.

echo Preparing language models for test

#for lm_suffix in bg tgpr tg bg_5k tgpr_5k tg_5k; do
for lm_suffix in bg_5k tg_5k; do
  test=data/lang_test_${lm_suffix}
  mkdir -p $test
  for f in phones.txt words.txt phones.txt L.fst L_disambig.fst \
     phones/; do
    cp -r data/lang/$f $test
  done
  gunzip -c $lmdir/lm_${lm_suffix}.arpa.gz | \
    arpa2fst --disambig-symbol=#0 \
             --read-symbol-table=$test/words.txt - $test/G.fst
  fstisstochastic $test/G.fst
 # The output is like:
 # 9.14233e-05 -0.259833
 # we do expect the first of these 2 numbers to be close to zero (the second is
 # nonzero because the backoff weights make the states sum to >1).
 # Because of the <s> fiasco for these particular LMs, the first number is not
 # as close to zero as it could be.

  # Everything below is only for diagnostic.
  # Checking that G has no cycles with empty words on them (e.g. <s>, </s>);
  # this might cause determinization failure of CLG.
  # #0 is treated as an empty word.
  mkdir -p $tmpdir/g
  awk '{if(NF==1){ printf("0 0 %s %s\n", $1,$1); }} END{print "0 0 #0 #0"; print "0";}' \
    < "$lexicon"  >$tmpdir/g/select_empty.fst.txt
  fstcompile --isymbols=$test/words.txt --osymbols=$test/words.txt $tmpdir/g/select_empty.fst.txt | \
   fstarcsort --sort_type=olabel | fstcompose - $test/G.fst > $tmpdir/g/empty_words.fst
  fstinfo $tmpdir/g/empty_words.fst | grep cyclic | grep -w 'y' &&
    echo "Language model has cycles with empty words" && exit 1
  rm -r $tmpdir/g
done

rm -r $tmpdir
echo "Succeeded in formatting data."