Xpress: The 10AI Speech Synthesis System for Blizzard Challenge 2023
Yaohua Bu, Yang Sheng Zhao · 2023
This paper presents the 10AI text-to-speech (TTS) system "Xpress" for the Blizzard Challenge 2023.Xpress is an endto-end TTS system trained on the provided publicly available French speech data for Hub task 2023-FH1.The Xpress system is constructed with a character-level prosody modeling component, a conditional Flow Variational AutoEncoder (FVAE) based acoustic model and a BigVGAN vocoder.Comparing with conventional TTS systems, no conventional front-end analysis component is employed in the proposed Xpress, thus significantly simplifying the overall construction of the TTS system.Moreover, with the simplified FVAE acoustic model design, the Xpress requires less computation resource in training and inference, donates three times faster in inference than transformer-style system.Evaluation results demonstrate the proposed system with competitive effectiveness in both MOS quality and MUSHRA evaluation, and leading quality in both SUS and HOMOS intelligibility assessments.