librispeech_lm_decode.sh
964 Bytes
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
#!/bin/bash
# Copyright 2016 Allen Guo
# Apache 2.0
# This script decodes the librispeech test set using a librispeech LM, which is assumed
# to be prepared already using the librispeech recipe.
. ./cmd.sh
. ./path.sh
set -e
stage=0
lib=../../librispeech/s5
lang=data/lang_libri_tg_small
srilm_opts="-subset -prune-lowprobs -unk -tolower -order 3"
step=tri5
. utils/parse_options.sh
# You do not need to redo this stage when changing the "step" argument
if [ $stage -le 0 ]; then
utils/format_lm_sri.sh --srilm-opts "$srilm_opts" \
data/lang $lib/data/local/lm/lm_tgsmall.arpa.gz \
data/local/dict/lexicon.txt $lang
fi
graph_dir=exp/multi_a/$step/graph_libri_tg
if [ $stage -le 1 ]; then
utils/mkgraph.sh $lang \
exp/multi_a/$step $graph_dir
fi
if [ $stage -le 2 ]; then
steps/decode_fmllr.sh --nj 45 --cmd "$decode_cmd" --config conf/decode.config $graph_dir \
data/librispeech/test exp/multi_a/$step/decode_libri_tg_librispeech
fi