CMU Pronunciation Dictionary Tools

• Dependencies
• Building
• Usage
  • Example: Porter Stemmer
  • Example: Phonetisaurus
  • Example: Sphinx Dictionary
  • Example: Festival Dictionary
  • Example: Git Merge Resolution
• VIM Syntax File
• CMU Pronunciation Dictionary File Format
  • Metadata
• File-Based Metadata
• Phone Table File Format
• Metadata Description File Format
  • CSV Metadata
  • RDF Metadata
• Configuration Options
  • ACCENT
  • FORMAT
  • METADATA
  • PHONESET
  • SORT
  • TAGSET
  • WARNING
• License

This is a collection of tools for working with the CMU Pronunciation Dictionary.

Dependencies

To install these libraries on Debian-based machines (including Ubuntu and Mint), you can run:

sudo apt-get install python-LIBRARY

On other operating systems, you can run:

sudo pip install LIBRARY

This will install the LIBRARY python package.

Building

The cmudict-tools project uses the python setup.py script to build and install the package. This requires the README.rst file, which can be built by running:

make README.rst

This build step depends on pandoc to create the RST file from README.md.

The Makefile respects the following environment variables:

If you use a custom prefix, you may need to adjust your PYTHONPATH environment variable when running cmudict-tools, so it can find the python module.

The Makefile provides the following standard GNU commands that make it easier to run the setup script:

Usage

The cmudict-tools program has the following command-line structure:

cmudict-tools [OPTIONS] COMMAND DICTIONARY
cmudict-tools [OPTIONS] COMMAND YOURS THEIRS
cmudict-tools [OPTIONS] COMMAND YOURS THEIRS BASE

The supported OPTIONS are:

NOTE: The --remove-stress option will remove any duplicate entries that result from removing the stress markers.

COMMAND can be one of:

The DICTIONARY file is auto-detected according to one of the supported input FORMAT values.

The SELECTOR value can be:

For the diff and merge commands, the following usage modes are supported:

Example: Porter Stemmer

The select command can be used to extract the data used to test a Porter stemmer algorithm:

./cmudict-tools select=word cmudict > in.txt
./cmudict-tools select="@stem|word" cmudict > out.txt

This examples entries to have metadata like:

BURN  B ER1 N
BURNING  B ER1 N IH0 NG #@@ stem=BURN @@

Example: Phonetisaurus

The dictionary can be converted to a form that is usable with phonetisaurus-align by running:

./cmudict-tools --format=sphinx --remove-context-entries \
	print cmudict > cmudict.lex

Or, if you don't want stresses on vowels:

./cmudict-tools --format=sphinx --remove-context-entries --remove-stress \
	print cmudict > cmudict.lex

This can then be passed to phonetisaurus-align using:

phonetisaurus-align --input=cmudict.lex --ofile=cmudict.corpus --seq1_del=false

Example: Sphinx Dictionary

To generate a sphinx4 dictionary you can run:

./cmudict-tools --format=sphinx --remove-stress \
	print cmudict > cmudict_SPHINX_40

Example: Festival Dictionary

To generate a festival dictionary you can run:

./cmudict-tools --format=festlex --output-context=festlex --remove-duplicate-contexts \
	print cmudict > cmudict.scm

Example: Git Merge Resolution

To configure git to support merging cmudict files, run:

git config --global merge.tool cmudict
git config --global mergetool.cmudict.cmd 'cmudict-tools merge $LOCAL $REMOTE $BASE > $MERGED'
git config --global mergetool.cmudict.trustExitCode true
git config --global mergetool.cmudict.keepBackup false

Now, when there are merge conflicts from a cmudict format file, you can run:

git mergetool -t cmudict

You will still need to check the file for conflicts in the case where an entry has been modified by the local and remote versions. These are denoted by standard conflict markers, so searching for <<<<< will work.

VIM Syntax File

The cmudict-tools project provides a syntax highlighting file for cmudict-style dictionaries.

You can install the files to a VIM install by running:

make VIMDIR=<path-to-vim> vim

Alternatively, if your system supports VIM addons (e.g. Debian Linux), you can install the files by running:

make vim_plugin

which installs the files to /usr/share/vim. If the addon files are not in this location, you need to point VIMDIR to the addons directory and VIMPLUGINDIR to the registry directory.

Once installed, it will automatically highlight files named cmudict. You can explicitly enable highlighting by using the VIM command:

set ft=cmudict

The following variables and file-based metadata are supported:

NOTE: The file-based metadata must occur within the first 5 lines of the file to be supported by the VIM syntax file.

If no FORMAT is specified, its value is determined based on the content. This logic has rules for the old Weide format as well as the new formats (both the currently maintained dictionary and cmudict-new formats). Additionally, it detects files containing Pronunciation Dictionary within a ;;; comment as cmudict documents.

NOTE: You need to set the variables before setting the filetype. For example:

let b:cmudict_phoneset="arpabet"
set ft=cmudict

Alternatively, you can set this information as file-based metadata. For example:

;;;@@ phoneset=arpabet @@

CMU Pronunciation Dictionary File Format

A line comment starts with ;;; and spans until the end of the current line. In the cmudict-weide format, a line comment starts with ##.

An entry has the form ENTRY PRONUNCIATION, where ENTRY consists of WORD for the primary entry for the word, or WORD(VARIANT) for an alternate entry. The ENTRY and PRONUNCIATION are separated by two space ( ) characters and ENTRY is in upper case. In the cmudict-new format, ENTRY and PRONUNCIATION are separated by a single space ( ) character and ENTRY is in lower case.

The VARIANT consists of a number from 1 to 9 that denotes an alternate entry. These are numbered consecutively from 1 in the current cmudict format and from 2 in the cmudict-weide and cmudict-new formats.

The PRONUNCIATION section consists of Arpabet-based phones separated by a single space ( ). The casing of the phones depends on the phoneset being used. The vowels have an additional stress marker, which can be:

• 0 for an unstressed vowel;
• 1 for a primary stressed vowel;
• 2 for a secondary stressed vowel.

An entry may have a comment. These comments start with # and span to the end of the line.

NOTE: The various releases of the cmudict contain various formatting errors in several entries, where those entries deviate from the format described here. These will show up as validation warnings when those dictionary versions are run through cmudict-tools with the appropriate validation warnings enabled.

Metadata

Metadata is not supported in the official CMU dictionary format. However, the cmudict-tools project interprets specifically formatted comments as metadata. This allows additional information to be provided in a way that is compatible with existing cmudict tools.

Metadata occurs in line comments for file-based metadata, or entry comments for entry-based metadata. The metadata section of the comment starts and ends with @@. The @@ must be at the start of the comment (i.e. no spaces or other characters) for it to be recognised as metadata. Any text after the metadata is treated as a regular comment.

The content within the metadata block may be in one of the following formats:

1. key-value

   The content within the metadata block is a sequence of space-separated key=value pairs. A key can occur multiple times, in which case the key will have both values.

2. json

   The metadata is a JSON string.

File-Based Metadata

This is metadata on line comments in the given dictionary format. This must be in the key-value format.

The file-based metadata that cmudict-tools understands is:

Phone Table File Format

This is a CSV document with the first line containing the titles of each field. At a minimum, it needs to support the following fields:

• Arpabet is the phone using an upper-case Arpabet transcription, excluding the stress marker;
• Normalized is the optional canonical form for phonesets that use a different transcription for a given phone;
• IPA is the International Phonetic Alphabet (IPA) transcription for the phone excluding stress markers;
• Type is the phone type (see below) of the given phone;
• Phone Sets is a semi-colon (;) separated list of phonesets that support this phone.

The Type field supports any values that can be separated using a ;. The following values are given special meanings:

• schwa -- A weak vowel that can have an optional unstressed (0) stress marker in pronunciations.

• vowel -- A vowel that must have a stress marker in pronunciations.

For example:

Arpabet,Normalized,IPA,Type,Phone Sets
AA,,ɑ,vowel,arpabet;ipa;cmu;festvox;cepstral;timit
AE,,æ,vowel,arpabet;ipa;cmu;festvox;cepstral;timit

Metadata Description File Format

CSV Metadata

This is a CSV document with the first line containing the titles of each field. At a minimum, it needs to support the following fields:

• Key is the metadata key;
• Value is an allowed value for the metadata key.

For example:

Key,Value
key1,value1
...
keyN,valueN

RDF Metadata

This is an RDF document using the SKOS ontology.

A key is defined as a skos:ConceptScheme and a value as a skos:Concept. The labels are defined using skos:notation predicates. A value is associated with a key using the skos:inScheme predicate. All other metadata triples are currently ignored.

For example, to support key=value a minimal RDF Turtle file is:

@prefix skos: <http://www.w3.org/2004/02/skos/core#> .

<#> a skos:ConceptScheme ; skos:notation "key" .

<#val> a skos:Concept ; skos:notation "value" ; skos:inScheme <#> .

Configuration Options

These are the various valid values used by the command-line options, VIM syntax file and file-based metadata. See the appropriate sections on how to specify these values.

ACCENT

The supported ACCENT values are:

• en-US to use the American English phone table;
• en-GB-x-rp to use the Received Pronunciation British English phone table;
• a CSV file to use phonesets defined in that CSV file (see Phone Table File Format for a description of this file format).

FORMAT

The supported FORMAT values are:

METADATA

The METADATA value points to a metadata description file containing the valid (key,value) pairs for entry-based metadata.

To test the (key,value) extraction for a metadata description file, you can run:

python metadata.py <metadata-description-file>

This will output JSON text, for example:

{"key1": ["value1", "value2"], "key2": ["value3"]}

Additionally, @type:key can be used to specify keys that have arbitrary values of a regular format. Valid types are:

• s for string values;
• i for integer numbers (e.g. 274);
• f for floating point numbers (e.g. 56.46325).

PHONESET

The supported PHONESET values depend on the phone table used. For the en-US and en-GB-x-rp phone tables defined by cmudict-tools, the supported phonesets are:

NOTE: The VIM syntax file does not support ipa phoneset validation.

SORT

The supported SORT values are:

• air to use the new-style sort order (group variants next to their root entry);

• none to leave the entries in the order they are in the dictionary;

• unicode to use the new-style sort order (group variants next to their root entry) with words sorted using the Unicode Collation Algorithm (UCA) to provide a more natural grouping of accented characters, etc.;

• weide to use the old-style sort order (simple ASCII character ordering).

TAGSET

The supported TAGSET values are:

WARNING

The supported WARNING values are:

NOTE: Currently, the unsorted check does not recognise the weide sort order.

If warn is used, the option is enabled. If no-warn is used, the option is disabled.

The following values have a special behaviour, and cannot be used with the no- prefix:

The order is important, as the warning set is tracked incrementally. This allows things like the following combinations:

It is also possible to locally disable warnings when using the JSON-based line metadata format using the disable-warnings metadata key as a list of the warnings to disable. For example:

IS  IH0 Z #@@{ "disable-warnings": ["missing-primary-stress"] }@@

License

The CMU Pronunciation Dictionary Tools are released under the GPL version 3 or later license.

Copyright (C) 2015-2016 Reece H. Dunn