ZeSTA: Zero-Shot TTS Augmentation with Domain-Conditioned Training for Data-Efficient Personalized Speech Synthesis

Researchers propose ZeSTA, a domain-conditioned training framework that improves personalized speech synthesis by better integrating synthetic and real speech data. The method addresses speaker similarity degradation issues when using zero-shot text-to-speech augmentation with limited real recordings.

• →ZeSTA framework uses domain embeddings to distinguish between real and synthetic speech during training
• →The approach improves speaker similarity over naive synthetic augmentation methods
• →Real-data oversampling helps stabilize adaptation when target data is extremely limited
• →Experiments on LibriTTS and proprietary datasets validate the framework's effectiveness
• →The method preserves speech intelligibility and perceptual quality while enhancing personalization