Paper: Joint Versus Independent Phonological Feature Models within CRF Phone Recognition

ACL ID N07-2004
Title Joint Versus Independent Phonological Feature Models within CRF Phone Recognition
Venue Human Language Technologies
Session Short Paper
Year 2007
Authors

We compare the effect of joint modeling of phonological features to independent feature detectors in a Conditional Random Fields framework. Joint modeling of fea- tures is achieved by deriving phonological feature posteriors from the posterior prob- abilities of the phonemes. We find that joint modeling provides superior perfor- mance to the independent models on the TIMIT phone recognition task. We ex- plore the effects of varying relationships between phonological features, and sug- gest that in an ASR system, phonological features should be handled as correlated, rather than independent.