Herta 0.1.5
黑塔 v0.1.5 — 实时语音、仓库卡片、立体设备卡片
本次更新
实时语音 黑塔的回复可以边出现边朗读：文字随语音同步显示，被她撤回的那一句由她自己的声音收回；开场白与提起设备时的那句话同样发声。语音由本地神经网络模型合成，模型不含在安装包内，在「设置 → 语音」中按需下载；也可使用自己的 MiniMax 密钥接入云端音色。
仓库卡片 侧栏显示工作区仓库的当前状态：分支与上游、未提交的改动、最近的提交，实时更新。点击一处改动查看差异，点击一个提交查看它改了什么；历史标签页可分页浏览、搜索提交信息、只读查看任一分支；尚未推送的提交带有标记。
设备卡片 差分协处理器的设备卡片改为物理渲染：随本地时钟从清晨走到深夜，按主题折叠光照，工作时呼吸，出错时闪动。场景构建期间显示毛玻璃预览；没有可用 GPU 的机器直接显示平面卡片。
协处理器 协处理器会读取自己写入的项目记忆。DeepSeek 已将 Flash 更名为 deepseek-flash 并内置图片输入，协处理器默认使用它，原「Flash 视觉版」选项随之移除；关闭推理时不再实际推理。
更新与稳定性 更新页在无法连接更新源时会说明原因，并提供网盘下载入口。修复了云端语音在控制请求上的无限等待、仓库 .git 目录被删除时的监听风暴、无 GPU 机器上设备卡片先毛玻璃后硬切的显示、语音合成初始化的超时处理，以及启动失败时的静默退出。
Downloads
| 平台 | 文件 |
|---|---|
| Windows | Herta-Setup-0.1.5.exe |
| macOS (Apple silicon) | Herta-0.1.5-arm64.dmg |
| macOS (Intel) | Herta-0.1.5-x64.dmg |
macOS 构建已签名并通过 Apple 公证。已安装 0.1.4 的用户将自动收到更新。
使用前需配置 DeepSeek API 密钥，密钥仅保存在本机。语音模型在首次启用实时语音时下载。
第三方声明 安装包内的语音合成运行时是 sherpa-onnx 1.13.6 的预编译包，其中静态链接了 espeak-ng（GPL-3.0-or-later）与 piper-phonemize（MIT）。对应源码：k2-fsa/sherpa-onnx v1.13.6 的构建树，以及它所拉取的 csukuangfj/espeak-ng@ed530aa 与 csukuangfj/piper-phonemize@f3ff95a。完整的许可证文本随应用安装在资源目录的 THIRD-PARTY-NOTICES.md 中。
In English
Real-time voice — Herta's replies can be spoken as they appear: the text is revealed in step with her voice, and a line she takes back, she takes back in her own voice; the opening and the line she says when the device is lifted are spoken too. Speech is synthesized by a local neural model, which is not in the installer and downloads on demand from Settings → Voice; a MiniMax cloud voice can be used with your own key.
Repository card — the rail shows the workspace repository's current state: branch and upstream, uncommitted changes, recent commits, kept live. Click a change for its diff, a commit for what it touched; the history tab pages, searches commit messages, and reads any branch without checking out; unpushed commits are marked.
Device card — the Coprocessor's device card is physically rendered: it follows the local clock from morning to night, folds the lighting per theme, breathes while working and flashes on failure. A frosted preview shows while the scene builds; machines without a usable GPU show the flat card directly.
Coprocessor — the Coprocessor reads the project memory it writes. DeepSeek has renamed Flash to deepseek-flash with image input built in; the Coprocessor uses it by default, and the former "Flash Vision" option is removed. Turning reasoning off now actually turns it off.
Updates and stability — the update pane explains when the update feed cannot be reached and offers the netdisk download. Fixed the cloud voice's unbounded waits on control requests, a watcher storm when a repository's .git directory is deleted, the device card's frost-then-snap on machines without a GPU, timeouts during speech-synthesis start-up, and a silent exit on a start-up failure.
macOS builds are signed and notarized by Apple. Existing 0.1.4 installs update automatically. A DeepSeek API key is required; it is stored only on your machine. The voice model downloads the first time real-time voice is enabled.
Third-party notices — the speech-synthesis runtime in the installer is sherpa-onnx 1.13.6's prebuilt package, which statically links espeak-ng (GPL-3.0-or-later) and piper-phonemize (MIT). Corresponding source: the k2-fsa/sherpa-onnx v1.13.6 build tree and the csukuangfj/espeak-ng@ed530aa and csukuangfj/piper-phonemize@f3ff95a commits it fetches. The full license texts are installed with the app in THIRD-PARTY-NOTICES.md under its resources directory.