Nearest Neighbor-Based Indonesian G2P Conversion

Abstract

Grapheme-to-phoneme conversion (G2P), also known as letter-to-sound conversion, is an important module in both speech synthesis and speech recognition. The methods of G2P give varying accuracies for different languages although they are designed to be language independent. This paper discusses a new model based on pseudo nearest neighbor rule (PNNR) for Indonesian G2P. In this model, partial orthogonal binary code for graphemes, contextual weighting, and neighborhood weighting are introduced. Testing to 9,604 unseen words shows that the model parameters are easy to be tuned to reach high accuracy. Testing to 123 sentences containing homographs shows that the model could disambiguate homographs if it uses long graphemic context. Compare to information gain tree, PNNR gives slightly higher phoneme error rate, but it could disambiguate homographs.