This repository contains scripts suitable for training, evaluating and using grapheme-to-phoneme models for speech recognition using the OpenFst framework. The current build requires OpenFst version 1.6.0 or later, and the examples below use version 1.6.2.
The repository includes C++ binaries suitable for training, compiling, and evaluating G2P models. It also some simple python bindings which may be used to extract individual multigram scores, alignments, and to dump the raw lattices in .fst format for each word.
Standalone distributions related to previous INTERSPEECH papers, as well as the complete, exported
final version of the old google-code repository are available via git-lfs in a separate
repository:
This build was tested via AWS EC2 with a fresh Ubuntu 14.04 and 16.04 base, and m4.large instance.
$ sudo apt-get update
# Basics
$ sudo apt-get install git g++ autoconf-archive make libtool
# Python bindings
$ sudo apt-get install python-setuptools python-dev
# mitlm (to build a quick play model)
$ sudo apt-get install gfortran
Next grab and install OpenFst-1.6.2 (10m-15m):
$ wget http://www.openfst.org/twiki/pub/FST/FstDownload/openfst-1.6.2.tar.gz
$ tar -xvzf openfst-1.6.1.tar.gz
$ cd openfst-1.6.1
# 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...
$ sudo make install
$ cd
# Extend your LD_LIBRARY_PATH .bashrc:
$ echo 'export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib:/usr/local/lib/fst' \
>> ~/.bashrc
$ source ~/.bashrc
Checkout the latest Phonetisaurus from master
$ git clone https://github.com/AdolfVonKleist/Phonetisaurus.git
$ cd Phonetisaurus
$ ./configure
$ make
$ sudo make install
or, if you want to compile with python bindings
$ git clone https://github.com/AdolfVonKleist/Phonetisaurus.git
$ cd Phonetisaurus
$ ./configure --enable-python
$ make
$ sudo make install
$ cd python
$ cp ../.libs/Phonetisaurus.so .
$ sudo python setup.py install
$ cd
Grab and install mitlm to build a quick test model with the cmudict (5m):
$ git clone https://github.com/mitlm/mitlm.git
$ cd mitlm/
$ ./autogen.sh
$ make
$ sudo make install
$ cd
Grab a copy of the latest version of CMUdict and clean it up a bit:
$ mkdir example
$ cd example
$ wget https://raw.githubusercontent.com/cmusphinx/cmudict/master/cmudict.dict
# Clean it up a bit and reformat:
$ cat cmudict.dict \
| perl -pe 's/\([0-9]+\)//;
s/\s+/ /g; s/^\s+//;
s/\s+$//; @_ = split (/\s+/);
$w = shift (@_);
$_ = $w."\t".join (" ", @_)."\n";' \
> cmudict.formatted.dict
Train a complete model with default parameters using the wrapper script:
$ phonetisaurus-train --lexicon cmudict.formatted.dict --seq2_del
INFO:phonetisaurus-train:2017-07-09 16:35:31: Checking command configuration...
INFO:phonetisaurus-train:2017-07-09 16:35:31: Checking lexicon for reserved characters: '}', '|', '_'...
INFO:phonetisaurus-train:2017-07-09 16:35:31: Aligning lexicon...
INFO:phonetisaurus-train:2017-07-09 16:37:44: Training joint ngram model...
INFO:phonetisaurus-train:2017-07-09 16:37:46: Converting ARPA format joint n-gram model to WFST format...
INFO:phonetisaurus-train:2017-07-09 16:37:59: G2P training succeeded: train/model.fst
Generate pronunciations for a word list using the wrapper script:
$ phonetisaurus-apply --model train/model.fst --word_list test.wlist
test T EH1 S T
jumbotron JH AH1 M B OW0 T R AA0 N
excellent EH1 K S AH0 L AH0 N T
eggselent EH1 G S L AH0 N T
Generate pronunciations for a word list using the wrapper script. Filter against a reference lexicon, add n-best, and run in verbose mode, and generate :
$ phonetisaurus-apply --model train/model.fst --word_list test.wlist -n 2 -g -v -l cmudict.formatted.dict
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: Checking command configuration...
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: beam: 10000
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: greedy: True
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: lexicon_file: cmudict.formatted.dict
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: model: train/model.fst
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: nbest: 2
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: thresh: 99.0
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: verbose: True
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: Loading lexicon from file...
DEBUG:phonetisaurus-apply:2017-07-09 16:48:22: Applying G2P model...
GitRevision: kaldi-1-g5028ba-dirty
eggselent 26.85 EH1 G S L AH0 N T
eggselent 28.12 EH1 G Z L AH0 N T
excellent 0.00 EH1 K S AH0 L AH0 N T
excellent 19.28 EH1 K S L EH1 N T
jumbotron 0.00 JH AH1 M B OW0 T R AA0 N
jumbotron 17.30 JH AH1 M B OW0 T R AA2 N
test 0.00 T EH1 S T
test 11.56 T EH2 S T
Generate pronunciations using the alternative % of total probability mass constraint, and print the resulting scores as human readable, normalized probabilities rather than raw negative log scores:
phonetisaurus-apply --model train/model.fst --word_list Phonetisaurus/script/words.list -v -a -p 0.85 -pr
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: Checking command configuration...
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: accumulate: True
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: beam: 10000
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: greedy: False
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: lexicon_file: None
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: logger: <logging.Logger object at 0x7fdaa93d2410>
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: model: train/model.fst
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: nbest: 100
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: pmass: 0.85
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: probs: True
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: thresh: 99.0
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: verbose: True
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: phonetisaurus-g2pfst --model=train/model.fst --nbest=100 --beam=10000 --thresh=99.0 --accumulate=true --pmass=0.85 --nlog_probs=false --wordlist=Phonetisaurus/script/words.list
DEBUG:phonetisaurus-apply:2017-07-30 11:55:58: Applying G2P model...
GitRevision: kaldi-2-g6e7c04-dirty
test 0.68 T EH1 S T
test 0.21 T EH2 S T
right 0.81 R AY1 T
right 0.13 R AY0 T
junkify 0.64 JH AH1 NG K AH0 F AY2
junkify 0.23 JH AH1 NG K IH0 F AY2
Align, estimate, and convert a joint n-gram model step-by-step:
# Align the dictionary (5m-10m)
$ phonetisaurus-align --input=cmudict.formatted.dict \
--ofile=cmudict.formatted.corpus --seq1_del=false
# Train an n-gram model (5s-10s):
$ estimate-ngram -o 8 -t cmudict.formatted.corpus \
-wl cmudict.o8.arpa
# Convert to OpenFst format (10s-20s):
$ phonetisaurus-arpa2wfst --lm=cmudict.o8.arpa --ofile=cmudict.o8.fst
$ cd
Test the manual model with the wrapper script:
$ cd Phonetisaurus/script
$ ./phoneticize.py -m ~/example/cmudict.o8.fst -w testing
11.24 T EH1 S T IH0 NG
-------
t:T:3.31
e:EH1:2.26
s:S:2.61
t:T:0.21
i:IH0:2.66
n|g:NG:0.16
<eps>:<eps>:0.01
Test the G2P servlet [requires compilation of bindings and module install]:
$ nohup script/g2pserver.py -m ~/train/model.fst -l ~/cmudict.formatted.dict &
$ curl -s -F "wordlist=@words.list" http://localhost:8080/phoneticize/list
test T EH1 S T
right R AY1 T
junkify JH AH1 NG K AH0 F AY2
junkify JH AH1 NG K IH0 F AY2
Use a special location for OpenFst, parallel build with 2 cores
$ ./configure --with-openfst-libs=/home/ubuntu/openfst-1.6.2/lib \
--with-openfst-includes=/home/ubuntu/openfst-1.6.2/include
$ make -j 2 all
Use custom g++ under OSX (Note: OpenFst must also be compiled with this custom g++ alternative [untested with v1.6.2])
$ ./configure --with-openfst-libs=/home/osx/openfst-1.6.2gcc/lib \
--with-openfst-includes=/home/osx/openfst-1.6.2gcc/include \
CXX=g++-4.9
$ make -j 2 all
If you need to rebuild the configure script you can do so:
$ autoreconf -i
$ sudo make install
$ sudo make uninstall
$ bin/phonetisaurus-align --help
$ bin/phonetisaurus-arpa2wfst --help
$ bin/phonetisaurus-g2prnn --help
$ bin/phonetisaurus-g2pfst --help
cpplint command:
$ ./cpplint.py --filter=-whitespace/parens,-whitespace/braces,\
-legal/copyright,-build/namespaces,-runtime/references\
src/include/util.h