voice-pack 2.0：音色包 skill 能单独装了voice-pack 2.0: the voice pack skill now installs on its own
2026-10-21——voice-pack 2.0.0 发布。它是给 agent 用的 skill：一个声音的参考集和配音指引装成一个包，谁拿到都能用。这是它第一个能单独装的正式版：和任何平台都解耦，最多两条命令就装进你的 agent。2026-10-21 — voice-pack 2.0.0 is out. It is a skill for your agent: a voice's reference clips and its dubbing guidance, sealed in one pack that anyone can use. This is its first formal release that installs on its own: decoupled from any platform, and in your agent with at most two commands.
这个 skill 正在从零重做（2026-10-30），旧版的安装入口已经撤下。已发布的签名包和各版标签都还在，已有的音色包照常能用。This skill is being rebuilt from scratch (2026-10-30), and the install entry for the old version has been withdrawn. The signed releases and the version tags are still there, and existing voice packs keep working.
这一版带来什么What this release brings
- 能单独装了。voice-pack 有了自己的公开仓库：别的 agent 用 skills 命令行一条命令装上，Claude Code 从插件市场两条命令装上，不用先装别的产品。It installs on its own. voice-pack has its own public repository: other agents install it with one skills CLI command, Claude Code with two plugin marketplace commands, and nothing else has to be installed first.
- 和平台解耦。流程按所需能力写，不点名模型，也不点名平台。官方的适配器只剩通用的 process：任何本机命令都能接成能力；没有现成的程序，骨架生成器几步写出你自己的适配器。Decoupled from any platform. Flows are written by the capability they need and name no model and no platform. The only official adapter left is the generic process adapter, which turns any local command into capabilities; with no program yet, the skeleton generator writes your own adapter in a few steps.
- 协议 2.0：包里写的是自己的名字。新做的音色包和分享文件只写 voice-pack 这个名字。新程序照旧读 1.x 的包，检查、补全、拆开都和以前一样；1.x 的程序读不了 2.0 的包，所以这是一个大版本。Protocol 2.0: packs carry their own name. New voice packs and share files carry only the voice-pack name. The new tools still read 1.x packs, and check, complete and unpack them as before; 1.x tools cannot read 2.0 packs, which is why this is a major version.
- 官方发布在产品自己的地址上。补全时取的官方版本索引和签名包，换到了 voice-pack 自己的仓库；签名的公钥没变。Official releases live at the product's own address. The official release index and signed packages that completion fetches moved to voice-pack's own repository; the signing key is unchanged.
- 做包更稳。做包时给片段打情绪、风格标签的模型暂时出错，在同一次做包里再试，每段最多 3 次，和转写用同一条规则；多出来的次数，事先算进你确认的费用里。Steadier pack making. When the model that tags clips with emotion and style fails with a transient error, it is retried within the same run, at most 3 times per clip, under the same rule as transcription; the extra attempts are counted in the cost you confirm beforehand.
为什么用它Why use it
一个声音不只是一段录音。参考集挑好了、响度对齐了，哪一句该用哪段参考、哪些词怎么念、配音时要注意什么，都写在包里。把包交给别人、交给别的 agent，拿到就能配出同一个声音，不用再来问你。A voice is more than one recording. In a pack, the references are chosen and loudness-matched, and it records which reference suits which line, how the tricky words are said, and what to watch for when dubbing. Hand the pack to someone else, or to another agent, and they can dub in the same voice without coming back to ask you.
skill 本身不绑模型：今天接本机的一个程序，明天换成别家的服务，台本、音色包和流程都不用改。And the skill is not tied to a model: plug in a program on your machine today and someone's service tomorrow, without changing your scripts, your packs or your flows.
听一听Listen
A 面 · 设计Side A · Designed
照一段文字描述设计出来的声音（voice-design），没有用到任何真人的录音。A voice designed from a written description (voice-design); no real person's recording went into it.
B 面 · 克隆Side B · Cloned
拿 A 面当参考克隆出来的同一个声音（voice-cloning），念的是英文。The same voice, cloned with side A as its reference (voice-cloning), speaking English.
两段念的是什么、怎么做出来的，在首页。What the two clips say, and how they were made, is on the home page.
装上Install
这个 skill 正在从零重做（2026-10-30），旧版的安装入口已经撤下。已发布的签名包和各版标签都还在，已有的音色包照常能用。This skill is being rebuilt from scratch (2026-10-30), and the install entry for the old version has been withdrawn. The signed releases and the version tags are still there, and existing voice packs keep working.
关于 voice-packAbout voice-pack
voice-pack 是给 agent 用的 skill：音色包的格式，和造声音、配音、做包、验包的一整套工具。一个声音的参考集和配音指引装成一个包，谁拿到都能用。流程按所需能力写，模型经适配器接进来。许可是 Apache-2.0，官方发布带 Ed25519 签名。voice-pack is a skill for your agent: the voice pack format, plus the tools to make voices, dub scripts, and build and check packs. A voice's reference clips and dubbing guidance go into one pack that anyone can use. Flows are written by the capability they need, and models come in through adapters. It is licensed under Apache-2.0, and official releases carry an Ed25519 signature.