SNAC: Speaker-Normalized Affine Coupling Layer in Flow-Based Architecture for Zero-Shot Multi-Speaker Text-to-Speech

Zero-shot multi-speaker text-to-speech (ZSM-TTS) models aim to generate a speech sample with the voice characteristic of an unseen speaker. The main challenge ZSM-TTS is increase overall speaker similarity for speakers. One most successful conditioning methods flow-based (TTS) utilize functions which predict scale and bias parameters affine coupling layers according given embedding vector. In this letter, we improve on previous method by introducing speaker-normalized (SNAC) layer allows synthesis in zero-shot manner leveraging normalization-based technique. newly designed explicitly normalizes input predicted from vector while training, enabling inverse process denormalizing new at inference. proposed scheme yields state-of-the-art performance terms quality setting.