Abstract:

[Application Background] In resource construction for Korean phonetic information processing, automatic phonetic transcription technology plays a crucial role. At present, there are two main approaches for grapheme-to-phoneme(G2P) conversion: knowledge-based and data-based. [Objective] The purpose of this paper is to solve the problems existing in these two approaches. The knowledge-driven method cannot easily adapt to the real situation of large volume data information, which results in complex models and difficult computations. The data-driven method relies on high-quality data and has difficulty in determining the input variables, which requires adequate model features and accurate selection. [Methods] This paper proposes a method which integrates the knowledge-driven and data-driven approaches for Korean phonetic transcription. Firstly, this method extracts accurate feature attributes based on the variation pattern of Korean speech to obtain high-quality data; then it trains the machine learning model for automatic pronunciation of Korean by taking the advantages of the data-driven approach in fitting the input and output variables. [Results] The proposed method takes the phonological changes in continuous Korean speech such as syllable weakening, disambiguation, augmentation and dissimilation into account, and can accurately obtain the phonemes corresponding to the graphemes. A cross-testing shows that this method can improve the performance of the prediction model, and the average correct rate of grapheme-phoneme conversion reaches 94.63%. [Conclusions] The automatic Korean phonetic transcription method proposed in this paper can effectively establish an accurate Korean pronunciation dictionary, which is expected to provide technical support for systems such as Korean speech recognition and synthesis.

Key words: knowledge-driven, data-driven, Korean, phonetic variation, grapheme-to-phoneme