VO2Lium : Voxygen and LIUM contribution for Blizzard 2025
Résumé
This paper presents the Vo2Lium contribution to Blizzard Challenge 2025. The challenge focuses on low resource data, and aims at synthesizing speech in Bildts, a unique language variety from Netherlands. Our contribution addresses both the mono-speaker hub task and the zero-shot task. It relies on the fine-tuning of pre-trained models with Bildts speech, a careful selection of audio prompts, and the development of automatic metrics and scores that guide the development and selection of TTS systems at multiple stages. The perceptual evaluation of our system highlights strong likeness, appropriateness and naturalness from Bildts audience, thus showing the relevance of our approach.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |