Zero-Shot Cloning
π Reading Time: ~13 min read
π
Updated 2026-03
CosyVoice 2 & F5-TTS: Zero-Shot Voice Cloning, Streaming Low-Latency Synthesis & WebUI Setup
Comparing CosyVoice 2 and F5-TTS non-autoregressive flow matching speech architectures. 3-second reference voice cloning, cross-lingual synthesis, and streaming WebUI.
ποΈ Zero-sample voice cloning duo: CosyVoice 2 vs F5-TTS
| model framework | Generate schema | Time to First Packet (TTFT) | Cross-language cloning capabilities |
|---|---|---|---|
| CosyVoice 2 (0.5B) | Autoregressive semantics Token + Flow Matching (Flow Matching) | About 150ms (excellent streaming) | Supports multiple dialects of Chinese, English, Japanese, Cantonese, and Korean |
| F5-TTS | Pure non-autoregressive flow matching (Diffusion / Flow) | Complete sentence generation is extremely fast (RTF < 0.15) | High-fidelity bilingual reproduction in Chinese and English |
β‘ Image pulling and quick sound reproduction
export HF_ENDPOINT="https://hf-mirror.net"
# Pull CosyVoice2 official model
huggingface-cli download FunAudioLLM/CosyVoice2-0.5B \
--local-dir /data/models/CosyVoice2-0.5B \
--local-dir-use-symlinks False
# Pull F5-TTS weights
huggingface-cli download SWivid/F5-TTS \
--local-dir /data/models/F5-TTS \
--local-dir-use-symlinks False