| Paper: | 03-P1.2 (ICASSP 2003 Paper) | ||
| Session: | ICASSP 2003 Papers | ||
| Time: | Tuesday, May 18, 13:00 - 15:00 | ||
| Presentation: | Poster (ICASSP 2003 Presentation) | ||
| Topic: | ICASSP 2003 Papers: ICASSP 2003 Papers | ||
| Title: | UNSUPERVISED, LANGUAGE-INDEPENDENT GRAPHEME--TO--PHONEME CONVERSION BY LATENT ANALOGY | ||
| Authors: | Jerome Bellegarda; Apple Computer, Inc. | ||
| Abstract: | Automatic, data-driven grapheme-to-phoneme conversion is a challenging but often necessary task. The top-down strategy implicitlyfollowed by traditional inductive learning techniques tends to dismissrelevant contexts when they have been seen too infrequently in thetraining data. The bottom-up philosophy inherent in, e.g., pronunciationby analogy, allows for a markedly better handling of rarer contexts, butproves nonetheless equally dependent on local, language-dependentalignments between letters and phonemes. This paper proposes analternative bottom-up approach, dubbed pronunciation by latentanalogy, which adopts a global definition of analogy, more amenable toobviate such supervision. For each out-of-vocabulary word, aneighborhood of globally relevant pronunciations is constructedthrough an appropriate data-driven mapping of its graphemic form.Phoneme transcription then proceeds via locally optimal sequencealignment and maximum likelihood position scoring. This method wassuccessfully applied to the speech synthesis of proper names with alarge diversity of origin. | ||
| Back | |||
Home -||-
Organizing Committee -||-
Technical Committee -||-
Technical Program -||-
Plenaries
Paper Submission -||-
Special Sessions -||-
ITT -||-
Paper Review -||-
Exhibits -||-
Tutorials
Information -||-
Registration -||-
Travel Insurance -||-
Housing -||-
Workshops