We have transferred our lexicon learning technology for English developed in WERNICKE and SPRACH to both the Portuguese and French systems, as described in the attached report, Automatic learning of lexica (Deliverable D2.5). This system provides for probabilistic learning of pronunciation variants, as well as the learning of probabilities of phonological rules.
The Broadcast News report (D7.2) illustrates how our English system was also extended by adding the automatic induction of phonological rules using decision trees. In particular, these methods have lead to improved static dictionaries for our evaluation Broadcast News system. We have also investigated promising avenues for incorporating the dynamic features of speaking rate and word predictability into pronunciation models.
Our lexica work resulted in a submission to ICASSP-99  and a submission to Speech Communications .