8.2.1 Dictionary Lookup
Phonetic dictionaries were introduced in Sec. ?? of Ch. 8. One of the most widely-used for TTS is the freely available CMU Pronouncing Dictionary (CMU, 1993), which has pronunciations for about 120,000 words. The pronunciations are roughly phonemic, from a 39-phone ARPAbet-derived phoneme set. Phonemic transcriptions means that instead of marking surface reductions like the reduced vowels [ax] or [ix], CMUdict marks each vowel with a stress tag, 0 (unstressed), 1 (stressed), or 2 (secondary stress). Thus (non-diphthong) vowels with 0 stress generally correspond to [ax] or [ix]. Most words have only a single pronunciation, but about 8,000 of the words have two or even three pronunciations, and so some kinds of phonetic reductions are marked in these pronunciations. The dictionary is not syllabified, although the nucleus is implicitly marked by the (numbered) vowel. Fig. 8.6 shows some sample pronunciations.
The CMU dictionary was designed for speech recognition rather than synthesis uses; thus it does not specify which of the multiple pronunciations to use for synthesis, does not mark syllable boundaries, and because it capitalizes the dictionary headwords, does not distinguish between e.g., US and us (the form US has the two pronunciations [AH1 S] and [Y UW1 EH1 S]).
| ANTECEDENTS | AE2 N T IHO S IY1 D AHO N T S | PAKISTANI | P AE2 K IHO S T AE1 N IYO |
| CHANG | CH AE1 NG | TABLE | T EY1 B AHO L |
| DICTIONARY | D IH1 K SH AHO N EH2 R IYO | TROTSKY | T R AA1 T S K IY2 |
| DINNER | D IH1 N ERO | WALTER | W AO1 L T ERO |
| LUNCH | L AH1 N CH | WALTZING | W AO1 L T S IHO NG |
| MCFARLAND | M AHO K F AA1 R L AHO N D | WALTZING(2) | W AO1 L S IHO NG |
The 110,000 word UNISYN dictionary, freely available for research purposes, resolves many of these issues as it was designed specifically for synthesis (Fitt, 2002). UNISYN gives syllabifications, stress, and some morphological boundaries. Furthermore, pronunciations in UNISYN can also be read off in any of dozens of dialects of English, including General American, RP British, Australia, and so on. The UNISYN uses a slightly different phone set; here are some examples:
going: { g * ou }.> i ng >
antecedents: { * a n . t ^ i . s ~ ii . d n! t }> s >
dictionary: { d * i k . sh @ . n ~ e . r ii }