SwanVoice: Expressive Long-Form Zero-Shot Speech Synthesis for Both Monologue and Dialogue
Researchers have developed SwanVoice, a novel zero-shot text-to-speech system capable of generating expressive, long-form dialogue for multiple speakers. The system combines VAE, flow-matching DiT, and diffusion post-training techniques, building upon a new dataset called SwanData-Speech. SwanVoice aims to overcome limitations in acoustic consistency and affective continuity across dialogue turns, outperforming existing open-source baselines in richness and hierarchy on the SwanBench-Speech benchmark, though content accuracy is noted as a remaining challenge. AI
IMPACT Introduces a new method for more natural and coherent multi-speaker dialogue synthesis, potentially improving conversational AI agents.