phonetisaurus script

little-test.sh

#!/bin/bash

if [ $# -ne 2 ]
then
    echo "USAGE: $0 <lexicon> <order>"
    echo "  Recommended: order=6~9 for English"
    exit
fi

#Triage the lexicon a bit to overcome my poor design
cat ${1} \
    | perl -e'while(<>){
                 chomp;
                 @_ = split(/\s+/);
                 $w = shift(@_);
                 $p = join(" ",@_);
                 $p =~ s/_/,/g;
                 print $w."\t".$p."\n";
               }' \
    > ${1}.triaged.lex

#Run the aligner with some reasonable params (for English)
phonetisaurus-align --input=${1}.triaged.lex --ofile=${1}.triaged.corpus --seq1_del=false
echo ""

#Train an n-gram model
# Note: You can use *any* tool that outputs a valid ARPA-format LM
# Recommended: Interpolate a Kneser-Ney model with a MaxEnt model (latest SRILM),
#   rescore with an RNNLM (if you want to get fancy)
# If possible results can be further improved by pruning pronunciations with
# a further forced alignment step with your recognizer
estimate-ngram -o ${2} -t ${1}.triaged.corpus -s FixModKN -wl ${1}.triaged.${2}g.arpa
echo ""

#Convert the model
phonetisaurus-arpa2wfst-omega --lm=${1}.triaged.${2}g.arpa --ofile=${1}.triaged.${2}g.fst
echo ""

#Test an input word, and fix the phoneme we triaged
phonetisaurus-g2p-omega --model=${1}.triaged.${2}g.fst --input=A-bomb \
    | perl -e'while(<>){
                 chomp;
                 @_ = split(/\t/);
                 $_[2] =~ s/,/_/g;
                 print join("\t",@_)."\n";
              }'
echo ""

#Test a list of input words, get the 5-best, and fixe the phoneme we triaged
for w in A-frame A-line; do echo $w; done > short.wlist

phonetisaurus-g2p-omega --model=${1}.triaged.${2}g.fst --input=short.wlist --isfile=true \
    --nbest=5 --decoder_type=fst_phi \
    | perl -e'while(<>){
                 chomp;
                 @_ = split(/\t/);
                 $_[2] =~ s/,/_/g;
                 print join("\t",@_)."\n";
              }'

echo ""

Try running this with something like, with just a couple thousand lines at first:

$ ./little-test.sh sampa.spaced.lex 3

If it works, turn up the order to 7 or so.

Assuming the '_' is the only character-related issue, I think this should suffice

Also note that you only need to run the aligner step one time.
You can use the corpus after that directly, and just play around with LM training/tuning, which is typically really fast.