Blame view
egs/tedlium/s5_r3/local/prepare_data.sh
2.98 KB
8dcb6dfcb
first commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 |
#!/bin/bash
#
# Copyright 2014 Nickolay V. Shmyrev
# 2014 Brno University of Technology (Author: Karel Vesely)
# 2016 Johns Hopkins University (Author: Daniel Povey)
# 2018 François Hernandez
#
# Apache 2.0
# To be run from one directory above this script.
. ./path.sh
# Prepare: test, train,
for set in dev test train; do
dir=data/$set.orig
mkdir -p $dir
# Merge transcripts into a single 'stm' file, do some mappings:
# - <F0_M> -> <o,f0,male> : map dev stm labels to be coherent with train + test,
# - <F0_F> -> <o,f0,female> : --||--
# - (2) -> null : remove pronunciation variants in transcripts, keep in dictionary
# - <sil> -> null : remove marked <sil>, it is modelled implicitly (in kaldi)
# - (...) -> null : remove utterance names from end-lines of train
# - it 's -> it's : merge words that contain apostrophe (if compound in dictionary, local/join_suffix.py)
{ # Add STM header, so sclite can prepare the '.lur' file
echo ';;
;; LABEL "o" "Overall" "Overall results"
;; LABEL "f0" "f0" "Wideband channel"
;; LABEL "f2" "f2" "Telephone channel"
;; LABEL "male" "Male" "Male Talkers"
;; LABEL "female" "Female" "Female Talkers"
;;'
# Process the STMs
cat db/TEDLIUM_release-3/legacy/$set/stm/*.stm | sort -k1,1 -k2,2 -k4,4n | \
sed -e 's:([^ ]*)$::' | \
awk '{ $2 = "A"; print $0; }'
} | local/join_suffix.py > data/$set.orig/stm
# Prepare 'text' file
# - {NOISE} -> [NOISE] : map the tags to match symbols in dictionary
cat $dir/stm | grep -v -e 'ignore_time_segment_in_scoring' -e ';;' | \
awk '{ printf ("%s-%07d-%07d", $1, $4*100, $5*100);
for (i=7;i<=NF;i++) { printf(" %s", $i); }
printf("
");
}' | tr '{}' '[]' | sort -k1,1 > $dir/text || exit 1
# Prepare 'segments', 'utt2spk', 'spk2utt'
cat $dir/text | cut -d" " -f 1 | awk -F"-" '{printf("%s %s %07.2f %07.2f
", $0, $1, $2/100.0, $3/100.0)}' > $dir/segments
cat $dir/segments | awk '{print $1, $2}' > $dir/utt2spk
cat $dir/utt2spk | utils/utt2spk_to_spk2utt.pl > $dir/spk2utt
# Prepare 'wav.scp', 'reco2file_and_channel'
cat $dir/spk2utt | awk -v set=$set -v pwd=$PWD '{ printf("%s sph2pipe -f wav -p %s/db/TEDLIUM_release-3/legacy/%s/sph/%s.sph |
", $1, pwd, set, $1); }' > $dir/wav.scp
cat $dir/wav.scp | awk '{ print $1, $1, "A"; }' > $dir/reco2file_and_channel
# Create empty 'glm' file
echo ';; empty.glm
[FAKE] => %HESITATION / [ ] __ [ ] ;; hesitation token
' > data/$set.orig/glm
# The training set seems to not have enough silence padding in the segmentations,
# especially at the beginning of segments. Extend the times.
if [ $set == "train" ]; then
mv data/$set.orig/segments data/$set.orig/segments.temp
utils/data/extend_segment_times.py --start-padding=0.15 \
--end-padding=0.1 <data/$set.orig/segments.temp >data/$set.orig/segments || exit 1
rm data/$set.orig/segments.temp
fi
# Check that data dirs are okay!
utils/validate_data_dir.sh --no-feats $dir || exit 1
done
|