LIUM-TTS entry for Blizzard 2023

Félix Saget (1) , Thibault Gaudier (1) , Meysam Shamsi (1) , Marie Tahon (1)

1 LIUM - Laboratoire d'Informatique de l'Université du Mans

This paper presents the LIUM-TTS entry for Blizzard 2023. It is the first participation of the LIUM in the Blizzard Challenge. The Blizzard Challenge 2023 focused on French language in two tasks. The Hub task was provided with 50 audio hours (with partially aligned annotation), and the Spoke task only 2 hours. The proposed TTS for the Hub task consists of a Transformer-based grapheme-to-phoneme, a FastSpeech 2based acoustic model, and a fine-tuned Waveglow vocoder. The output of this system has been fed through a voice conversion module from Hub to Spoke voice. The perceptual evaluation of our system in comparison with other Blizzard participants shows its weaknesses and highlights future working axes to deal with upcoming challenges.

Mots clés: Text-to-speech; voice conversion; Blizzard Challenge 2023 Text-to-speech voice conversion; Blizzard Challenge 2023

Domaines: Informatique [cs]; Intelligence artificielle [cs.AI]

Félix Saget, Thibault Gaudier, Meysam Shamsi, Marie Tahon. LIUM-TTS entry for Blizzard 2023. Blizzard Challenge Workshop, Aug 2023, Grenoble, France. pp.28-33, ⟨10.21437/Blizzard.2023-2⟩. ⟨hal-04188761⟩