decode procedure

1  test data preparation

1>  select representative data voice to match real application scenario

    test device  :containing noise

    read by text : relatively clean

    tts : synthesis

2> get right label for test data

　　i. call other asr api to do the labeling

　　ii.use our asr  model to do this

　　draw a parallel between the two results above

2 get standard label

python get_label_ref.py $test_dir  $ref

3 decoding

for file in $test_dir/*
do
 run.pl $decode_dir/decode.`basename $file`.log \
 online2-wav-nnet3-latgen-faster --online=true --do-endpointing=false --frame-subsampling-factor=3 \
 --config=conf/online.conf --max-active=7000 --beam=15.0 --lattice-beam=6.0 --acoustic-scale=1.0 \
 --word-symbol-table=exp/chain/tdnn/graph/words.txt exp/chain/tdnn/final.mdl exp/chain/tdnn/graph/HCLG.fst \
 "ark:echo utterance-id1 utterance_`basename $file`|" "scp:echo utterance_`basename $file` $file |" ark:/dev/null 
done

4 result deal

#get test_set's decoding result
for file in $decode_dir/*
do
     name=`basename $file`
     label=`cat $file | grep '^utterance'`
     label_1=`echo ${label/utterance_/}`
     label_2=`echo ${label_1/.wav/}`
     echo $label_2 > $text/${name}
done

5 get wer

if you want to get CER, you can set a flag to do and get cer result

if [ $flag == 'cer' ] || [ $flag == 'CER' ] ;then
echo "do CER estimation:"
PYTHONIOENCODING=utf8 python3 handle_file.py $text $ref
else
echo "do WER estimation:"
fi

and the default performance means WER

#each file to compute WER
for file in ${text}/*
do
    cat $file >> $total_content
    compute-wer --text --mode=present  ark:$ref ark:$file > $log_wer/wer_`basename $file` 
done 
#all files to compute a overall WER 
compute-wer --text --mode=present ark:$ref ark:$total_content