Alibaba recently open-sourced CosyVoice3, an advanced text-to-speech system based on large language models that excels in content consistency, speaker similarity, and prosodic naturalness. It supports 9 common languages and 18+ Chinese dialects, enabling multilingual zero-shot voice cloning. A developer has created a Windows local TTS tool based on this model that requires only 4GB of VRAM to run and supports four modes: zero-shot cloning, fine control, instruction control, and speech repair. The tool is fully locally deployed with no API calls required, featuring a clean and user-friendly interface suitable for various applications including video dubbing, game NPC dialogue, and audiobook production. Performance comparisons show that CosyVoice3 outperforms similar open-source models across multiple test metrics, demonstrating the latest advancements in AI speech synthesis technology.
