Meta-TTS: Meta-Learning for Few-Shot Speaker Adaptive Text-to-Speech

Personalizing a speech synthesis system is highly desired application, where the can generate with user’s voice rare enrolled recordings. There are two main approaches to build such in recent works: speaker adaptation and encoding. On one hand, methods fine-tune trained multi-speaker text-to-speech (TTS) model few samples. However, they require at least thousands of fine-tuning steps for high-quality adaptation, making it hard apply on devices. other encoding encode enrollment utterances into embedding. The TTS synthesize conditioned corresponding Nevertheless, encoder suffers from generalization gap between seen unseen speakers. In this paper, we propose applying meta-learning algorithm method. More specifically, use Model Agnostic Meta-Learning (MAML) as training model, which aims find great meta-initialization adapt any few-shot tasks quickly. Therefore, also meta-trained speakers efficiently. Our experiments compare proposed method (Meta-TTS) baselines: baseline baseline. evaluation results show that Meta-TTS high speaker-similarity samples fewer than outperforms under same scheme. When pre-trained extra 8371 data, still outperform LibriTTS dataset achieve comparable VCTK dataset.