About CosyVoice Models
CosyVoice is a family of open-source TTS models by FunAudioLLM that delivers high-quality speech synthesis, zero-shot voice cloning, and low-latency streaming from v1.0 to v3.0.
All CosyVoice Models
alibaba/cosyvoice-design
[Core Function] CosyVoice Design creates a temporary voice from a natural-language voice_prompt and synthesizes speech in one async request; only the final audio is returned. [Strengths] No voice enrollment management; language_hint (zh/en) applies to both design and synthesis; same prosody and format controls as CosyVoice TTS. [Best For] One-off designed speech without storing enrolled voices. [Limitations] Do NOT use this to obtain a reusable voice library entry; the designed voice is temporary and is not returned. model must be cosyvoice-v3.5-plus or cosyvoice-v3.5-flash; voice_prompt and text are required.
alibaba/cosyvoice-clone
[Core Function] CosyVoice Clone clones a speaker from a public reference audio URL and synthesizes new speech in one async request; only the final audio is returned. [Strengths] No voice enrollment management; language_hint applies to both cloning and synthesis; SSML, hot_fix, and prosody controls. [Best For] One-off cloned narration and demos where a lasting voice library is not needed. [Limitations] Do NOT use this to obtain a reusable voice library entry; the cloned voice is temporary and is not returned. Reference URL must be publicly accessible. model must be cosyvoice-v3.5-plus or cosyvoice-v3.5-flash. [Routing] Choose cosyvoice-v3.5-plus for higher speech quality; cosyvoice-v3.5-flash for lower latency.
alibaba/cosyvoice-v3-flash
[Core Function] CosyVoice v3 Flash is Alibaba's low-latency text-to-speech model. [Strengths] Rich system voice catalog, fixed-format instruction on Instruct-capable system voices, SSML, hot_fix, AIGC watermark, Markdown filter (cloned voices only), and multiple audio formats with faster turnaround than Plus. [Best For] Voice assistants, interactive prompts, IVR, short announcements, dialect system voices, and latency-sensitive batch TTS. [Limitations] Do NOT use when maximum speech quality or long-form audiobook fidelity is the priority (use Plus). System-voice instruction must follow CosyVoice voice-list fixed Chinese formats. [Routing] Choose Flash when speed or a richer system-voice catalog matters most. Choose Plus for premium narration quality.
alibaba/cosyvoice-v3-plus
[Core Function] CosyVoice v3 Plus is Alibaba's high-quality text-to-speech model. [Strengths] System voices (e.g. longanyang, longanhuan), SSML and LaTeX input, hot_fix pronunciation correction, AIGC watermark, and output in mp3, pcm, wav, or opus. System-voice instruction must use the fixed Chinese formats in the CosyVoice voice list. [Best For] Brand voiceovers, audiobooks, high-quality narration, marketing clips, and scenarios where speech quality matters more than minimum latency. [Limitations] Do NOT use for real-time streaming or word-level timestamps. Fewer system voices than Flash. [Routing] Choose Plus when quality or narration fidelity matters most. Choose cosyvoice-v3-flash for lower latency or a richer system-voice catalog.
