Voice Input · 语音输入插件

A browser-side speech-to-text plugin for the DeepSeek Harness
chat interface: click the mic, speak, and the transcript lands in the message box — no more typing
long messages.

Everything happens in your browser: microphone permission, recording, recognition, and writing
into the input box. No server relay, no audio upload (unless you explicitly configure the optional
Whisper endpoint).

✨ Features

• A mic button injected right next to the send button (≤ 30px, rounded, tooltip "语音输入").
• Click to start listening: button turns red with a pulse animation, live interim results and a
  recognized-character counter appear in a preview strip under the composer.
• Click again to stop: the final text is appended to the end of the input box (existing content
  kept, fully editable, input refocused). No auto-send unless you enable it.
• Continuous dictation (Web Speech API continuous + interimResults) — long monologues keep
  committing sentence by sentence, no 60s limit.
• Cancel anytime (web-native, benchmarked against WeChat): while listening, a 「✕ 取消」
  button appears in the preview strip and Esc discards the current session — including an
  automatic rollback of the text this session already appended.
• Smart punctuation: questions end with「？」, exclamations with「！」, otherwise「。」.
• Optional hold-to-talk mode (touch/trackpad style: press and hold to record, release to
  commit, slide up to cancel) — default stays click-to-toggle for desktop.
• Full error handling: permission denied → guidance; no-speech → one auto retry; service
  unavailable; no microphone; unexpected end → one auto-restart.
• Survives re-renders / session switches; unloading removes the button, stops recognition and
  releases the microphone.
• Optional enhanced tier: fill a local Whisper endpoint and it switches to
  MediaRecorder → chunked upload → transcription.

⚙️ Configuration (Settings → 语音输入)

🔌 Installation (DeepSeek Harness)

This is the Client half of a dynamic Cordis plugin. Define it with your harness
plugin tooling (plain JavaScript, React via React.createElement, no JSX), register the three
slots, and approve the client package when prompted.

// code.client = the contents of client.js

Slot usage (official catalog seats):

Requires a secure context (https or localhost, e.g. http://127.0.0.1:3080) and
Chrome/Edge (Web Speech API). On unsupported browsers the button is disabled and the settings page
shows a capability report.

🔒 Privacy

• Default path uses the browser's built-in speech recognition: no recording, no saving, no
  upload.
• Only when you configure whisperEndpoint does the plugin send 4-second audio chunks to that
  endpoint (your own local service).

📄 License

MIT

中文说明

为 DeepSeek Harness 聊天界面增加「点击说话 → 自动转文字」能力。全部在浏览器端完成：
麦克风权限、录音、识别、写入输入框；无服务端中转、不上传音频（除非显式配置增强档）。

功能

• 输入框右侧注入麦克风按钮（≤30px），点击变红 + 脉冲动画，实时中间结果与字数显示在输入框下方预览条
• 停止后最终文字追加到输入框末尾（保留已有内容、可编辑），默认不自动发送
• continuous 连续听写，长段口述逐句落定，不受 60 秒限制；识别中不劫持键盘
• 完整错误处理：权限拒绝指引 / 无语音自动重试 / 服务不可用 / 无麦克风 / 意外中断自动重连
• 卸载/禁用即清理：停止识别、释放麦克风、移除按钮与样式

配置（设置 → 语音输入）

language（默认 zh-CN）、autoSend（说完即发，默认关）、punctuation（自动补标点，默认开）、
autoRestart（意外中断自动重连，默认开）、whisperEndpoint（增强档，默认空 = 纯浏览器识别）。

浏览器兼容性

Chrome / Edge（webkitSpeechRecognition）；需安全上下文（https 或 localhost）。
不支持时按钮置灰，设置页显示能力检测报告。

隐私

默认方案不录制、不保存、不上传音频；只有填写 whisperEndpoint 后才会把 4 秒一段的
录音发送到你指定的本地端点。

文件

• client.js — 插件 Client 半区源码（识别引擎 + 三个 Slot 组件 + 配置页）
• docs/语音输入插件·创造模式提示词.md — 项目原始提示词（设计规范）
• 验证报告.md — 能力检测 + 手动测试清单