dsh-speech-input
用于 DeepSeek Harness 的麦克风按钮，可将浏览器语音识别结果写入编写器草稿。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:liznee/dsh-speech-input#1224ad888ba6d9b2e378d757ddb4f5b17624634a
Speech Input 以 dsh-speech-input 发布，当前版本为 0.2.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
dsh-speech-input 为 DeepSeek Harness Web 输入框增加一个原生麦克风按钮。点击开始听写，识别结果实时写入当前草稿；再次点击停止并补全句末标点。插件不会自动发送消息。
image
image
conversation.input.right 插槽，不查询或改写 Harness DOMsrc/client/index.js 的 DEFAULT_SILENCE_TIMEOUT_MS 调整，设为 0 可关闭自动停止完整历史见 CHANGELOG.md。
直接从 GitHub 安装到 Harness Web profile：
dsh plugin --profile web add github:liznee/dsh-speech-input
仓库已提交预构建的 lib/，Git 安装不会执行构建脚本，也不需要在
pnpm-workspace.yaml 中授权 allowBuilds。安装或升级后重启 dsh web。
已发布到 npm，也可以直接安装：
dsh plugin --profile web add dsh-speech-input
本地 tarball：
npm ci
npm test
npm pack
dsh plugin --profile web add ./dsh-speech-input-0.2.1.tgz
插件本身不保存音频、不写日志，也不把音频发送给 Harness 或 DeepSeek。为了显示真实音量，插件会额外打开一条本地麦克风流，只连接到 Web Audio AnalyserNode 计算 RMS；不会播放、录制或上传该流，停止或取消时会立即关闭音轨。Web Speech API 的识别处理路径仍由浏览器决定：Edge/Chrome 通常会把识别音频交给浏览器厂商的在线语音服务。因此它不是离线识别；若不能接受该数据路径，请不要授权麦克风。
SpeechRecognition 实现，按钮会禁用。识别文字只通过 Harness 的公开 inputActions.setDraft() 写入普通草稿。模型看不到音频、识别状态或任何额外系统提示；只有用户最终发送的文字会进入模型上下文。
dsh-speech-input adds a native microphone control to the DeepSeek Harness Web composer. Click once to dictate into the current draft and again to stop. It never submits the message automatically.
It uses the public conversation.input.right, inputActions.setDraft(), and composer-block APIs. While listening it shows a gray circular cancel control, a full-width 24-segment meter driven by locally measured microphone RMS, and a square stop control; the host send action is disabled. Cancel removes the current dictation while preserving the pre-existing draft and recognizable manual suffix edits. It also supports interim-result replacement, reports permission and network failures, releases recognition and local metering tracks on teardown, and includes reduced-motion and screen-reader behavior. After 5 seconds without detected speech it auto-stops (like a manual stop, keeping the dictation and showing a brief notice); tune DEFAULT_SILENCE_TIMEOUT_MS in src/client/index.js, or set it to 0 to disable auto-stop.
Install directly from GitHub:
dsh plugin --profile web add github:liznee/dsh-speech-input
Prebuilt lib/ artifacts are committed, so Git installation does not execute a build script or require a pnpm allowBuilds grant. Restart dsh web after installing or updating.
Published on npm; install directly with:
dsh plugin --profile web add dsh-speech-input
The plugin uses the browser Web Speech API. Edge and Chrome commonly send microphone audio to the browser vendor's online speech service. The plugin does not store audio or send it to Harness/DeepSeek, but it is not an offline recognizer. Press Enter while listening to stop immediately and keep the dictation; press Enter again to send.
See CHANGELOG.md for the full history. v0.2.1 reworked the listening pill (oval ring + shadow, end-cap curvature matching the circular buttons, thicker strokes, a five-bar volume stop icon, symmetric outlined cancel/stop buttons), removed the waveform baseline, and lowered the default silence timeout to 5 seconds; v0.2.0 introduced auto-stop on silence.
npm ci
npm test
npm run test:coverage
npm run pack:check
Node.js 20 or newer is required for development. The built client is emitted in Harness's lazy CommonJS module-factory format.