最后更新：11/10/2026最后核验：2026-11-10
<p>ZONOS3-Mini 不只有模型权重：官方提供可操作的 Zonos Voice Studio 与轻量声音克隆页面，能选择声音、上传参考音频、输入文本并导出语音，也可通过 MLX、CUDA 或 C++/ggml 在自己的设备上运行。工具支持从参考音频克隆声音，以及调整年龄、口音、音色、情绪、语速和词级表达，适合游戏角色、旁白与语音助手开发。模型权重采用 Apache-2.0，运行时代码采用 MIT，本地使用软件免费但需自行承担硬件成本；Zyphra 托管 API 单独按每百万 UTF-8 字节 10 美元计费，不等于免费云端额度。浏览器版要求 WebGPU，首次约需下载 1.9 GB 模型并在当前浏览器缓存。声音和情绪控制属于生成目标，不能保证每次一致，需要试听复核。本文依据官方产品页、定价、模型卡、仓库及公开体验页核验，未下载模型、生成声音、注册账号或调用付费 API。</p><ul><li>多语言文本转语音与参考音频声音克隆</li><li>Zonos Voice Studio 和轻量 WebGPU 浏览器体验页</li><li>声音设计、情绪、语速及词级表达控制</li><li>本地多运行时与独立计费的托管语音 API</li></ul>
详情
平台
网页版macOSWindowsLinuxAPI
功能特性
- Generate multilingual speech and clone voices from a reference audio sample
- Use Zonos Voice Studio or the lightweight voice-cloning page in a WebGPU browser
- Design voice identity and steer emotion, pace, and word-level delivery
- Run local MLX, CUDA, or C++/ggml runtimes, or use the separately billed hosted speech API
支持语言
enzhjakodeesfritpt
适合人群
Developers building local or self-hosted speech and voice-assistant workflowsGame and narration creators experimenting with character voices and expressive speechTeams comparing an open local TTS stack with a separately priced hosted API
已知限制
- Local software is free, with Apache-2.0 model weights and MIT runtime code; hosted inference is separately billed at $10 per 1M UTF-8 bytes
- The browser playground needs WebGPU and an approximately 1.9 GB initial model download, cached in that browser
- Voice identity and expression controls are generation targets, not guarantees; outputs need listening and review
- Local execution uses your own hardware; published device latency figures are vendor measurements, not our own tests
替代工具
ElevenLabs — Hosted voice generation, cloning, and dubbing platform
VoiceStudio — Broader local desktop workflow for dubbing, transcription, and audiobooks
