Skip to Content

Department of Linguistics

ROBERT MANNELL - RESEARCH PAGES

Development of word stress rules

Word stress patterns are explicitly stored in TTS system dictionaries. Word stress for words not found in the TTS dictionary need to be generated by explicitly coded rules or by applying machine learning algorithms to large word databases. Grapheme-to-phoneme (GTP) rules tend to poorly handle word stress. In the string of phonemes output by the present GTP system in MU-Talk, word stress is only indicated by the presence of full vowels or by the schwa vowel. GTP rules tend to greatly under produce reduced vowels because only the most complex rules deal with orthographic strings that consist of more than one syllable. Single syllable rules are effectively monosyllabic word rules and single syllable words in citation form generally contain fully realised vowels rather than reduced vowels (schwa).

Word stress rules require a word-stress post-processor module which applies vowel reduction rules once the syllabic structure of the word and the vowel in each syllable is known. Additionally, word stress rules require knowledge of whether a word is a noun or a verb (etc.) for the application of certain rules. Also, in continuous speech, function word reduction rules also need to be applied. Words generated by the GTP rule system are unknown words (ie. not in the dictionary), so part-of-speech and function/content words status would be unknown (although such words will always be assumed to be content words as the set of function words is a closed set and all such words are in the dictionary). It is sometimes possible to predict whether a word is a noun or a verb by examining the word (particularly the end of the word) for certain grapheme sequences, but this results in probabilistic output as most such patterns have exception words.

The word-stress post-processor will apply vowel reduction rules to the strings output by the GTP rules. Since many rules are probabilistic, this module will very often produce more than one version of a word, with different vowels reduced in each version. Note that the GTP rules produce only one pronunciation, but that the word-stress post processor may produce more than one pronunciation differing only in stress patterns.

Stress rules sub-modules:-
i. word stress for bisyllabic and polysyllabic words
ii. compound word stress
iii. stress alternations caused by appending certain suffixes

Project Status

Some progress occurred on this project in 1999 and 2000 with some prototype software being written and with a general overview of the stress rules being formulated. Unfortunately, little progress has occurred since then.

Relevant Papers

Mannell R.H., & Clark J.E., (1987) "Text-to-speech rule and dictionary development", Speech Communication 6, 1987, pp 317- 324.

Mannell R.H., & Clark J.E., (1986) "Text-to-speech rule and dictionary development", Proceedings of the First Australian Conference on Speech Science and Technology, Canberra, Nov. 1986. pp 14-19