Using Auto-Encoder BiLSTM Neural Network for Czech Grapheme-to-Phoneme Conversion

The crucial part of almost all current TTS systems is a grapheme-to-phoneme (G2P) conversion, i.e. the transcription of any input grapheme sequence into the correct sequence of phonemes in the given language. Unfortunately, the preparation of transcription rules and pronunciation dictionaries is not an easy process for new languages in TTS systems. For that reason, in the presented paper, we focus on the creation of an automatic G2P model, based on neural networks (NN). But, contrary to the majority of related works in G2P field, using only separate words as an input, we consider a whole phrase the input of our proposed NN model. That approach should, in our opinion, lead to more precise phonetic transcription output because the pronunciation of a word can depend on the surrounding words. The results of the trained G2P model are presented on the Czech language where the cross-word-boundary phenomena occur quite often, and they are compared to the rule-based approach.