The GIPSA-Lab Text-To-Speech System for the Blizzard Challenge 2023

This paper describes the GIPSA-Lab submission to the Blizzard Challenge 2023. The Text-To-Speech system trained for this challenge is a Transformer-based non-autoregressive encoder-decoder architecture based on FastSpeech2. Updates of the FastSpeech2 framework were provided to specifically train the model on orthographic inputs, which is our main focus for this edition of the challenge. This model was trained with both orthographic and phonetic transcriptions of the same dataset. An additional phonetic prediction layer was added to the model. This additional layer enables to train the text encoder on phonetic prediction alone, without the need for audio recordings.