NN-KOG2P: A Novel Grapheme-to-Phoneme Model for Korean Language

Abstract

With the development of text-to-speech technology, high-quality voices can be heard in AI speaker responses, car navigation guidance, and news article-reading services. As services become more diverse, domains are expanded, requiring fast and high-performance grapheme-to-phoneme (G2P) technology. In this paper, we propose a novel Korean G2P model architecture, reflecting the characteristics of Korean pronunciation, called neural network-based Korean G2P (NN-KoG2P). Our proposed method achieves high accuracy in an open-domain dataset and a fast inference speed that can generate pronunciation sequences in real-time services.