Neural TTS in French: Comparing Graphemic and Phonetic Inputs Using the SynPaFlex-Corpus and Tacotron2

Abstract

The SynPaFlex-Corpus is a publicly available TTS-oriented dataset, which provides phonetic transcriptions automatically produced by the JTrans transcriber, with a Phoneme Error Rate (PER) of 6.1%. In this paper, we analyze two mono-speaker Tacotron2 models trained on graphemic and phonetic inputs, provided by the SynPaFlex-Corpus. Through three subjective listening tests, we compare their pronunciation accuracy, sound quality and naturalness. Our results show significantly better pronunciation accuracy and prosody naturalness for the phoneme-based model, but no significant difference in terms of perceived sound quality. They demonstrate that a PER of 6.1% is sufficient to enhance pronunciation control by using phonetic transcripts instead of graphemes with 83 hours of recorded French read speech. They suggest that the SynPaFlex-Corpus is suitable for pre-training a model in mono-speaker fine-tuning approaches.