File: timitdic.doc, updated 10/11/90 TIMIT Lexicon Documentation The dictionary found in the file "timitdic.txt" contains entries for all of the words in the TIMIT prompts. There are a total of 6229 entries in the lexicon. The lexicon was derived in part from the MIT adapted version of the Merriam-Webster Pocket Dictionary of 1964 ("Pocket") and a preliminary version of a general English lexicon under development at CMU. The pronunciations in the MIT pocket lexicon have been verified and modified over the years by various people. However, many of the words in the TIMIT scripts did not appear in the pocket lexicon, and needed to be added. These include other forms of words found in "Pocket" and words not found in any form. Rules were used to generate pronunciations in the former case and the derived pronunciations were hand-checked. In the latter case, consisting mainly of proper names and abbreviated forms (such as "takin'" instead of "taking" or "'em"), the pronunciations were added by hand. The symbols in the lexical representation are abstract, quasi-phonemic marks representing the underlying sounds and typically correspond to a variety of different sounds in the actual recordings. The term quasi-phonemic is used because some differences represented in the lexicon are often found in speech, even though they are not really phonemically distinctive in English, such as the /er/~/axr/ in which /er/ co-occurs with stress. Format of the Lexicon --------------------- All entries have been converted to lower case. Stress markers are given as 1 for primary stress and 2 for secondary stress, tacked on to the end of the vowel symbol. Hyphenated words such as "head-in-the-clouds" can be found as a single entry and as the individual words "head", "in", "the", and "clouds". One pronunciation is provided per entry except in the case where the same orthography corresponds to different parts of speech with different pronunciations, and both forms exist in the TIMIT scripts. To differentiate these words, multiple entries are given, with the syntactic class following the symbol ~. The classes found in the lexicon are: ~n noun ~v verb ~adj adjective ~pres present tense ~past past tense. An example is the word "live", with the entries: live~v l ih v live~adj l ay v Pronunciation Conventions ------------------------- The pronunciation is specified using the the CMU symbol set (see "phoncode.doc" for a description of the symbols). While it is realized that one pronunciation is often not sufficient, many of the alternate pronunciations may be predicted by use of phonological rules and may be highly dialect dependent. Using only one pronunciation per word forced the somewhat vexing decision of which one to use. We did not put extensive study into such issues, and do not make any claims of the theoretical correctness of our decisions on particular words. Our tendency has been to use the more marked alternate because we think it is harder to predict. We tried to make the pronunciations as consistent as possible. In a number of cases we referred to the authorities Kenyon and Knott (1953) and Webster's Third New International Dictionary (1966). 1) Vowel variability: Many of the pronunciation differences for vowels occur in semi-vowel environments and in unstressed syllables. - The vowel in words like "for", "pour", and "more" are often transcribed using either the vowel /ow/ or the vowel /ao/. This lexicon uses /ao/. - The vowel in words like "air" and "care" has been transcribed using /ae/, so as to differentiate this vowel from the /eh/ in "berry". Some speakers actually make a three-way distinction ("Mary" "merry" "marry"), with the vowel in Mary being somewhat in between an /eh/, /ae/, and /ey/. These speakers may use the same vowel in words like "care". - The vowel /ih/ (as opposed to /iy/) has been systematically used in the transcription of words like "fear" and "year". - unstressed schwa alternation /ix/~/ax/: /ix/ is usually used for schwas between 2 alveolars ("roses" /r ow1 z ix z/), otherwise /ax/ is used ("ahead" /ax hh eh1 d/). - /r/ following the diphthongs /aw/ ("hour") and /ay/ ("fire") have been transcribed as /axr/, except where the /r/ is syllable-initial as in words like "irate" and "virus". - vowel reduction: In some cases the pronunciation of a word may alternate between a full vowel and a highly reduced one. In these cases preference was given to the pronunciation with the more marked vowel instead of the schwa. For example, the pronunciation of "accept" is given as /ae1 k s eh p t/, not /ax1 k s eh p t/. 2) Stress differences: - /er/~/axr/ alternation -- /axr/ is used in unstressed syllables and /er/ in stressed syllables. - /ih/~/ix/, /ah/~/ax/ -- once again the distinction is based on stress. The forms /ix/ and /ax/ are used in reduced syllables. - /y uw/~/y uh/ -- the tendency is to use /y uw/ in stressed positions as in "attribution" /ae t r ih b y uw1 sh ix n/, and /y uh/ in unstressed position as in "attribute~v" /ax t r ih2 b y uh t/. 3) Syllabics: The transcriptions are fairly liberal in the use of the syllabics /em/, /en/, and /el/, even though they may be pronounced as a sequence of a schwa followed by /m/, /n/, or /l/. For example, words ending in "-ism" are transcribed as /ih z em/ even though a short schwa often appears in the transition from the /z/ to the /em/. - /en/ must follow a coronal, except in rare occurrences such as "cap'n" /k ae1 p en/ and "haven't" /h ae1 v en t/. - in general, /el/ is used instead of /ax l/ except before a stressed vowel. Some exceptions are found in words ending in the "-ly" suffix. For example, "angrily" is transcribed /ae1 ng g r ax l iy/, not /ae1 ng g r el iy/. The only occurrences of /el l/ are found in compound words such as "jungle-like" and "liberal-led". References ---------- John S. Kenyon and Thomas A. Knott (1953), "A Pronouncing Dictionary of American English," Merriam-Webster Inc., Springfield, Ma. "Webster's Third New International Dictionary, Unabridged" (1966), Philip B. Gove, Ed., William Benton, Publ., Encyclopedia Britannica, Inc., Chicago.