International Speech Lexicon Project Page (ISLEX)

Mark Hasegawa-Johnson and Margaret Fleck

Download

• Interim version May 22, 2008. Changes: coverage almost doubled by incorporating the nmoby dictionary; phone codes changed to worldbet; all information with any restrictions is now removed, so that you can now freely redistribute the dictionary without a license file. Caveat: the dictionary is way too big, you should not use this dictionary for any real ASR task; subdivided smaller versions will (hopefully) be posted here soon. Download worldbet version here. The StarChallenge phone code is a kind of stripped-down worldbet. Download StarChallenge dictionary here.
• Version 0.2.0, June 6, 2007: zip, tgz

Features

1. Redistributable. All sources for this dictionary allow redistribution (see SOURCES.bib for a list of sources, partly obsolete). All entries in this dictionary may be edited and redistributed.
2. Words with pronunciation but no part of speech tags are words that have not been verified or could not be verified. Most of these words will eventually be verified, but some may be unverifiable, either because they are neologisms from some corpus or other, or because they are mistakes. They remain in the dictionary for now, in case you find them useful.
3. Words marked "(fw misspelling)" are words known to be misspelled, usually from the Switchboard dictionary. They remain in the dictionary because they may be useful in Switchboard training.
4. Useful out-of-the-box. This dictionary includes every entry in the Mississippi State Switchboard dictionary, including fragments, digit strings, neologisms, mispronunciations, and misspellings. Therefore you should be able to train a Switchboard recognizer using this dictionary without any special manipulation of fragments, digit strings, etc., etc. Note that this property will change soon: fragments and neologisms will be consigned to auxiliary dictionaries, downloadable but separate.
5. Feature-rich. The dictionary includes the following information. Pronunciation, syllabification, and lexical stress are available for every word. Part of speech and named entity tags (when appropriate) exist for most words. Morphology exists for only a few words.
6. Part of speech is derived from the Edinburgh WSJ corpus, from the SIL KTagger program, and from named entity files: see README-POS.txt.
7. Named entity tags are labeled with "nnp" or "nnps", followed by the named entity type. The following tags currently exist.
   • (nnp (surname|boyname|girlname) \d\.\d\d\d) specifies that the entry was logged as a surname, male given name, or female given name in the 1990 U.S. Census, or was listed as one of these name types in a relevant wikipedia entry, or (in the case of manual verification) elsewhere on the web. The number at the end of the entry specifies the percent frequency of the surname, male name, or female name in 1990 U.S. Census data, rounded to the nearest thousandth of one percent.
   • (nnp city) tags include cities with more than 100,000 inhabitants, U.S. state capitols, and other cities.
   • (nnp country) includes member states of the United Nations, and an assortment of other country-like entities (dependent protectorates, historical nations, etc.).
   • (nnp continent) and (nnp state) are used to mark continents and U.S. states.
   • (nnp place) is used for other places including rivers, mountains, planets and moons, and mythical places.
   • (nnp product) is used generically for registered trademarks, generic names of medicines, types of music, and names of philosophies (e.g., dixieland (nnp product)).
   • (nnp company) is a for-profit corporation; (nnp organization) is any non-country, non-company organization.
8. Morphology: two types of morphological information are provided by the SIL KTagger program for the words it knows (with some extra labels provided by hand):
   1. ( root:) specifies the root word
   2. (morphology:...+...) specifies orthographic morpheme decomposition for words with known non-trivial decomposition
   Mark Hasegawa-Johnson, June 1, 2007