Speech Synthesis of Valencian Using a Conditional Variational Autoencoder with Adversarial Learning

The growing demand for high-quality speech synthesis systems in minority languages presents a notable challenge for researchers. In response, this study focuses on synthesizing Valencian speech to develop an effective text-to-speech system for this linguistic variety. A meticulously recorded corpus, comprising 7 hours of speech data, was utilised to train a model based on a conditional variational autoencoder with adversarial learning, specifically Variational Inference with adversarial learning for end-to-end Text-to-Speech (VITS). Additionally, a pretrained multispeaker model was fine-tuned using 30 minutes, and the entire corpus. Perceptual testing was conducted to evaluate the synthesised speech quality, revealing promising results. Notably, the proposed model demonstrated competitiveness compared to the recently released Valencian model by the Aina project, indicating its efficacy in generating natural and fluent Valencian speech. These findings contribute to advancing the field of Valencian text-to-speech synthesis and carry implications for the development of speech synthesis systems in other minority languages.