Input Encoding for Sequence-to-Sequence Learning of Romanian Grapheme-to-Phoneme Conversion

This paper evaluates the use of sequence-to-sequence learning models for the Romanian grapheme-to-phoneme conversion. The strategies explore the use of different input feature encoding: one-hot letter encoding, additional embedding layer and grapheme embeddings learned from a large corpus of Romanian text. Additional lexical information, such as syllabification and lexical stress is also taken into consideration for augmenting the orthographic form of the word and providing more accurate phonetic transcriptions.The sequence-to-sequence models are also compared to a baseline decision tree algorithm in terms of both phone- and word-level accuracy. The best results are achieved by the model which uses grapheme embeddings and all additional linguistic information. Its accuracy is 97.90% at word-level, and 99.62% at phone-level. However, only minor differences exist between the tested systems.