aishell_train_lms.sh
3.32 KB
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
77
78
79
80
81
82
83
84
85
86
87
88
#!/bin/bash
# To be run from one directory above this script.
. ./path.sh
text=data/local/train/text
lexicon=data/local/dict/lexicon.txt
for f in "$text" "$lexicon"; do
[ ! -f $x ] && echo "$0: No such file $f" && exit 1;
done
# This script takes no arguments. It assumes you have already run
# aishell_data_prep.sh.
# It takes as input the files
# data/local/train/text
# data/local/dict/lexicon.txt
dir=data/local/lm
mkdir -p $dir
kaldi_lm=`which train_lm.sh`
if [ -z $kaldi_lm ]; then
echo "$0: train_lm.sh is not found. That might mean it's not installed"
echo "$0: or it is not added to PATH"
echo "$0: Use the script tools/extras/install_kaldi_lm.sh to install it"
exit 1
fi
cleantext=$dir/text.no_oov
cat $text | awk -v lex=$lexicon 'BEGIN{while((getline<lex) >0){ seen[$1]=1; } }
{for(n=1; n<=NF;n++) { if (seen[$n]) { printf("%s ", $n); } else {printf("<SPOKEN_NOISE> ");} } printf("\n");}' \
> $cleantext || exit 1;
cat $cleantext | awk '{for(n=2;n<=NF;n++) print $n; }' | sort | uniq -c | \
sort -nr > $dir/word.counts || exit 1;
# Get counts from acoustic training transcripts, and add one-count
# for each word in the lexicon (but not silence, we don't want it
# in the LM-- we'll add it optionally later).
cat $cleantext | awk '{for(n=2;n<=NF;n++) print $n; }' | \
cat - <(grep -w -v '!SIL' $lexicon | awk '{print $1}') | \
sort | uniq -c | sort -nr > $dir/unigram.counts || exit 1;
# note: we probably won't really make use of <SPOKEN_NOISE> as there aren't any OOVs
cat $dir/unigram.counts | awk '{print $2}' | get_word_map.pl "<s>" "</s>" "<SPOKEN_NOISE>" > $dir/word_map \
|| exit 1;
# note: ignore 1st field of train.txt, it's the utterance-id.
cat $cleantext | awk -v wmap=$dir/word_map 'BEGIN{while((getline<wmap)>0)map[$1]=$2;}
{ for(n=2;n<=NF;n++) { printf map[$n]; if(n<NF){ printf " "; } else { print ""; }}}' | gzip -c >$dir/train.gz \
|| exit 1;
train_lm.sh --arpa --lmtype 3gram-mincount $dir || exit 1;
# LM is small enough that we don't need to prune it (only about 0.7M N-grams).
# Perplexity over 128254.000000 words is 90.446690
# note: output is
# data/local/lm/3gram-mincount/lm_unpruned.gz
exit 0
# From here is some commands to do a baseline with SRILM (assuming
# you have it installed).
heldout_sent=10000 # Don't change this if you want result to be comparable with
# kaldi_lm results
sdir=$dir/srilm # in case we want to use SRILM to double-check perplexities.
mkdir -p $sdir
cat $cleantext | awk '{for(n=2;n<=NF;n++){ printf $n; if(n<NF) printf " "; else print ""; }}' | \
head -$heldout_sent > $sdir/heldout
cat $cleantext | awk '{for(n=2;n<=NF;n++){ printf $n; if(n<NF) printf " "; else print ""; }}' | \
tail -n +$heldout_sent > $sdir/train
cat $dir/word_map | awk '{print $1}' | cat - <(echo "<s>"; echo "</s>" ) > $sdir/wordlist
ngram-count -text $sdir/train -order 3 -limit-vocab -vocab $sdir/wordlist -unk \
-map-unk "<SPOKEN_NOISE>" -kndiscount -interpolate -lm $sdir/srilm.o3g.kn.gz
ngram -lm $sdir/srilm.o3g.kn.gz -ppl $sdir/heldout
# 0 zeroprobs, logprob= -250954 ppl= 90.5091 ppl1= 132.482
# Note: perplexity SRILM gives to Kaldi-LM model is same as kaldi-lm reports above.
# Difference in WSJ must have been due to different treatment of <SPOKEN_NOISE>.
ngram -lm $dir/3gram-mincount/lm_unpruned.gz -ppl $sdir/heldout
# 0 zeroprobs, logprob= -250913 ppl= 90.4439 ppl1= 132.379