Modified Grapheme Encoding and Phonemic Rule to Improve PNNR-Based Indonesian G2P

Abstract

A grapheme-to-phoneme conversion (G2P) is very important in both speech recognition and synthesis. The existing Indonesian G2P based on pseudo nearest neighbour rule (PNNR) has two drawbacks: the grapheme encoding does not adapt all Indonesian phonemic rules and the PNNR should select a best phoneme from all possible conversions even though they can be filtered by some phonemic rules. In this paper, a modified partial orthogonal binary grapheme encoding and a phonemic-based rule are proposed to improve the performance of PNNR-based Indonesian G2P. Evaluating on 5-fold cross-validation, contain 40K words to develop the model and 10K words to evaluation each, shows that both proposed concepts reduce the relative phoneme error rate (PER) by 13.07%. A more detail analysis shows the most errors are from grapheme ⟨e⟩ that can be dynamically converted into either /E/ or /@/ since four prefixes, 'ber', 'me', 'per', and 'ter', produce many ambiguous conversions with basic words and also from some similar compound words with both different pronunciations for the grapheme ⟨e⟩. A stemming procedure can be applied to reduce those errors.