A hybrid approach for grapheme-to-phoneme conversion based on a combination of partial string matching and a neural network

The quality of a text-to-speech (TTS) system heavily depends on the transcription quality of the words to be spoken. Obviously the best transcription can be found in a phonetic dictionary. But for out of vocabulary (OOV) words fall back routines have to be developed. This paper proposes a fall back routine that combines the correctness of a phonetic dictionary with the exibility of a neural network. In the rst step parts of the OOV word are looked up in the dictionary. They are then connected with the additional feature that the last phoneme of the rst part is re-estimated using a neural network and a special phonetic dictionary. In the second step the word stress is determined either from the dictionary or using a second neural network.