Abstract:
Grapheme-to-Phoneme (G2P) conversion is a
necessary step for speech synthesis and speech
recognition. This paper addresses the problem of
grapheme to phoneme conversion for the Myanmar
language. In our method, we propose four simple
Myanmar syllable pronunciation patterns as features
that can be used to augment the models in a Conditional
Random Field (CRF) approach to G2P conversion. Our
results show that our additional features are able to
improve a strong baseline model that does not include
them. We found that combination of all four features
gave rise to the highest performancefor Myanmar
language G2P conversion.
