FunAudioLLM releases Fun-CosyVoice3-0.5B-2512, an Apache 2.0 voice cloning TTS
Fun-CosyVoice3 is a 0.5B zero-shot cloning TTS for nine languages and 18+ Chinese dialects, with latency down to 150 ms.
FunAudioLLM published Fun-CosyVoice3-0.5B-2512, a 0.5B-parameter text-to-speech model. The card dates the 2512 base and RL models to December 2025. It supports nine languages: Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, plus 18-plus Chinese dialects and accents. Features include zero-shot multilingual and cross-lingual voice cloning, pronunciation control with Pinyin and English CMU phonemes, instruction control for emotion, speed and volume, and bi-streaming with latency as low as 150 ms.
The card's own benchmarks: on test-zh the character error rate is 1.21 for the base model and 0.81 for the RL variant, against 1.45 for CosyVoice2; on test-en the word error rate is 2.24 and 1.68 against 2.57. These are the developers' own results, and several comparison models are closed source.
The licence is Apache-2.0. The card says the content is "for academic purposes only" and meant to demonstrate technical capabilities, which sits oddly next to Apache 2.0. Weights come in Safetensors and ONNX.
The date used is the Hugging Face repo creation date, 11 December 2025, consistent with the card's December 2025 roadmap entry.
Good numbers and a permissive licence. Read the academic-use note too.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0 weights in Safetensors and ONNX, with a base and an RL variant.
- Useful 4: Cloning, dialects and pronunciation control suit Chinese-English work.
- Notable 3.5: Strong self-reported error rates and a permissive licence.
- Novel 3: An incremental update to CosyVoice2 with reinforcement-learning tuning.
Comments