CosyVoice is a multilingual text-to-speech model family from Alibaba’s FunAudioLLM team — capable of zero-shot voice cloning, cross-lingual synthesis, and fine-grained control over emotion and prosody.
CosyVoice is an open-source project on GitHub with around 24,000 stars and active development in 2026, now at its third generation. The models generate natural speech across nine languages including Chinese, English, Japanese, Korean, and Cantonese, plus multiple Chinese dialects. Given a short reference clip, CosyVoice can clone a speaker’s voice and read new text in that voice — even in a different language — while pronunciation inpainting and instruct-style controls adjust emotion, speed, and style. The repository provides the full stack: inference, training and fine-tuning code, and deployment support including streaming inference.
CosyVoice belongs to Alibaba’s FunAudioLLM family alongside the SenseVoice speech-understanding models, and the third generation expands language coverage and instruction following — the model responds to natural-language style prompts as well as speaker references. For production use, the repository includes optimised serving runtimes (including a TensorRT-LLM path) and recipes for fine-tuning on custom voices, so a cloned voice can be improved with your own data.
CosyVoice is free and open source under the Apache-2.0 license. You can download the models and run them on your own hardware; no API fees or hosted service are required.
Developers and researchers building voice interfaces, audiobook and video narration, accessibility tools, or virtual characters — anyone who needs high-quality, controllable, multilingual speech synthesis they can run themselves.