Run phonetisaurus on TIMIT
You have to provide the TIMIT lexicon yourself
The first few cells set up phonetisaurus; they are adapted from the instructions in the git README.
%%capture
!apt-get -y install git g++ autoconf-archive make libtool
# Python bindings
!apt-get -y install python-setuptools python-dev
# mitlm (to build a quick play model)
!apt-get -y install gfortran
%%capture
!wget http://www.openfst.org/twiki/pub/FST/FstDownload/openfst-1.6.2.tar.gz
!tar -xvzf openfst-1.6.2.tar.gz
%cd openfst-1.6.2
# Minimal configure, compatible with current defaults for Kaldi
!./configure --enable-static --enable-shared --enable-far --enable-ngram-fsts
!make -j 4
# Now wait a while...
!make install
import os
ldlibpath = os.environ['LD_LIBRARY_PATH']
#_STORED_LD = "/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
newld = f'{ldlibpath}:/usr/local/lib:/usr/local/lib/fst'
os.environ['LD_LIBRARY_PATH']=newld
%env LD_LIBRARY_PATH
%%capture
%cd /content
!git clone https://github.com/AdolfVonKleist/Phonetisaurus.git
%cd Phonetisaurus
!./configure
!make
!make install
%cd /content/
We also need MITLM
%%capture
!git clone https://github.com/mitlm/mitlm
%cd mitlm
!autoreconf -i
!./configure
!make
!make install
%cd /content
The TIMIT dictionary is relatively clean, so there are only a few small changes that are needed for phonetisaurus.
!cat TIMITDIC.txt|grep -v '^;'|tr -d '/'|sed -e 's/ */ /g;s/~adj//;s/~v_past//;s/~v_pres//;s/~v//;s/~n//;' > TIMIT.cleaned
!cat TIMIT.cleaned | perl -pe 's/\s+/ /g; s/^\s+//; s/\s+$//; @_ = split (/\s+/); $w = shift (@_); $_ = $w."\t".join (" ", @_)."\n";' > TIMIT.clean
!phonetisaurus-align --input=TIMIT.clean --ofile=TIMIT.clean.corpus --seq1_del=false
That thing I just said about the TIMIT dictionary being relatively clean? Nah. There are some errors, particularly with 'c' being transcribed as 'ao' (which is a vowel sound). Also, the default output of phonetisaurus-align only does 1:1, 1:0, 0:1, 2:1, and 1:2 mappings of graphemes and phonemes, which means some of the alignments look quite strange.
%%writefile clean_ngrams.pl
#!/usr/bin/perl
# Fix some of the alignments from phonetisaurus-align to be more recognisable to humans
# Also fixes some transcription errors in the TIMIT dictionary (mostly c -> ao)
use warnings;
use strict;
use utf8;
my $raw_replacements = <<_HERE_;
e}_ l}el e|l}el
e}_ d}ed e|d}ed
e}_ d}d e|d}d
e}iy1 e}_ e|e}iy1
i}ix o|n}n i|o}ix n}n
r}_ t|-}r t}t r}r t|-|t}t
-|k}n n|a}ae1 -}_ k|n}n a}ae1
a|c}ax c}k a}ax c|c}k
c}k h}_ c|h}k
c}k q|u}w c|q}k u}w
n}n|t c}s n}n c}t|s
i|c}ih1 k|-}k i}ih1 c|k}k -}_
a|k}ey1 e|-}k a}ey1 k}k e|-}_
-|k}n n|a}ae2 -}_ k|n}n a}ae2
a|t}ax e}_ -|e}t y}ay1 a}ax t}t e}_ -}_ e|y}ay1
t|u}ch r}axr t}ch u|r}axr
e}_ d}d e|d}d
a}ae1 e}_ a|e}ae1
a}ih e}_ a|e}ih
-|c}ao -}_ c}k
x}eh1|k -}s x}eh1|k|s -}_
e}_ l|l}el e|l|l}el
w|h}hh y}w|ay1 w|h}hh|w y}ay1
a|d}ax j|o}jh u|r}er1 a}ax d|j}jh o|u|r}er1
a|d}ae2 u}jh|uw a}ae2 d}jh u}uw
u}y|uh a|b}b u|a}y|uh b}b
x}k -}s x}k|s -}_
u|r}er1 r}_ u|r|r}er1
o|r}axr r}_ o|r|r}axr
u|r}axr r}_ u|r|r}axr
e|r}axr r}_ e|r|r}axr
a|r}axr r}_ h|o}iy1 e}_ a|r|r|h}axr o|e}iy1
e|r}er r}_ e|r|r}er
i|r}er1 r}_ i|r|r}er1
u}_ a}aa1 u|a}aa1
w|h}hh i}w|er1 r}_ w|h}hh|w i|r}er1
b|o}b r}r b}b o|r}r
e}_ a|r}er1 e|a|r}er1
q|u}k a}w|ey2 q}k u}w a}ey2
q|u}k a}w|ao1 q}k u}w a}ao1
w|h}hh a}w|ax w|h}hh|w a}ax
t|u}ch r}axr t}ch u|r}axr
d|u}jh a}uw|ax d}jh u}uw a}ax
c|i}sh a}iy|ey2 c}sh i}iy a}ey2
i}ix a|t}t i|a}ix t}t
w|h}hh e}w|iy1 a|t}t w|h}hh|w e|a}iy1 t}t
q|u}k a}w|aa1 q}k u}w a}aa1
q|u}k a}w|ao2 q}k u}w a}ao2
q|u}k a}w|ae1 q}k u}w a}ae1
w|h}hh a}w|aa1 w|h}hh|w a}aa1
w|h}hh a}w|aa2 w|h}hh|w a}aa2
w|h}hh e}w|ae1 w|h}hh|w a}ae1
w|h}hh e}w|ae2 w|h}hh|w a}ae2
w|h}hh i}w|ay1 w|h}hh|w a}ay1
w|h}hh o}w|aa1 w|h}hh|w o}aa1
y|a}y c|h}aa1 t}t y}y a}aa1 c|h}_ t}t
i}iy1 e}_ i|e}iy1
m|a}m '}_ a}ae1 m}m a|'|a}ae1
g|u}g e}_ g}g u|e}_
r}r h}_ r|h}r
s}z s|a}ix s|s}z a}ix
_HERE_
my %replacements = ();
for my $rl (split('\n', $raw_replacements)) {
next if($rl !~ /\t/);
my @tmp = split(/\t/, $rl);
$replacements{$tmp[0]} = $tmp[1];
}
my $regex_inner = join('|', map { quotemeta $_ } keys %replacements);
while(<>) {
chomp;
while(/(?:^| )($regex_inner)(?:$| )/g) {
my $m = $1;
my $qm = quotemeta($m);
s/$qm/$replacements{$m}/;
}
my @phns = split/ /;
my @out = ();
for my $phn (@phns) {
if($phn =~ /^([-'])\|/) {
my $ch = $1;
push @out, "$ch}_";
push @out, substr($phn,2);
} elsif($phn =~ /^([^\|])\|([-'])\}(.*)$/) {
my $ch1 = $1;
my $ch2 = $2;
my $ch3 = $3;
push @out, "$ch1}$ch3";
push @out, "$ch2}_";
} elsif($phn eq 'c}ao') {
if($phns[0] eq 'n}n') {
push @out, 'c}s';
} else {
push @out, 'c}k';
}
} else {
push @out, $phn;
}
}
print join(' ', @out) . "\n";
}
!cat TIMIT.clean.corpus | perl clean_ngrams.pl > TIMIT.cleaner.corpus
!estimate-ngram -o 8 -t TIMIT.cleaner.corpus -wl timit.arpa
# Convert to OpenFst format (10s-20s):
!phonetisaurus-arpa2wfst --lm=timit.arpa --ofile=timit.fst