This yields a five-module system performing

1. morphological segmentation
2. graphemic parsing
3. grapheme-phoneme conversion
4. syllabification
5. stress assignment

Actually, this is what the unknown-word-pronunciation module of MITalk (Allen, Hunnicutt, and Klatt, 1987) does.

By training inductive learning algorithms on each of the modules of these possible word pronunciation systems, and coupling the resultant modules, one can test the hypothesis whether modularisation is a good idea.

When IGTree is used for this purpose, the following generalisation accuracy numbers are obtained (% correct test instances).

(click here or on the image above to start moving)

Last update: 25 January 2000