Whisper Right Ctrl 是一款為 Windows 11 設計的開源「按住說話」工具。
把游標放在 LINE、瀏覽器、Word、Notion、Obsidian 或其他文字欄位，按住鍵盤右側的 Ctrl 說話，放開後便會使用 faster-whisper 在本機辨識，轉成繁體中文並貼到游標位置。
Windows 內建的 Win + H 不能替換成 Whisper，也需要使用 Microsoft 的語音辨識服務。這個工具提供另一個選擇：
- 錄音在記憶體中交給本機模型處理
- 支援 NVIDIA CUDA
float16加速 - 按住 Right Ctrl 錄音，放開後自動貼入
- OpenCC 自動轉換為繁體中文
- 可從工作列圖示直接切換麥克風
- 依麥克風名稱自動找回裝置，不受 Windows 重排編號影響
- 若目前的 Windows 音訊介面暫時失效，會自動嘗試同一支麥克風的其他介面
- 可設定登入 Windows 後靜默啟動
- 不會留下黑色命令視窗
- 支援瀏覽器、通訊軟體、文件編輯器及筆記工具
- 原始碼、安裝流程與測試全部公開
- 前往 Releases 下載最新版 ZIP。
- 解壓縮到固定資料夾。
- 執行
INSTALL.cmd。 - 從清單選擇正確的麥克風。
- 首次啟動等待
turbo模型下載完成。 - 按住 Right Ctrl 說話，放開後等待文字出現。
若要更換輸入裝置，請在工作列通知區的綠色麥克風圖示按滑鼠右鍵，開啟「Microphone」選單。選擇會立即生效並保存；下次登入時，即使 Windows 改變裝置編號，程式也會依名稱找回原本的麥克風。
預設設定以 NVIDIA GPU 為目標。參考測試硬體為 RTX 4060 Ti 16 GB。
- 快速回覆 Email、LINE 與社群訊息
- 在 ChatGPT、Codex 或其他 AI 工具中用語音輸入提示
- 會議後補充筆記與工作事項
- 工地巡檢、設計討論及施工紀錄
- 長篇 Word、Notion 或 Obsidian 草稿
- 希望錄音不送到雲端的工作環境
模型下載完成後，語音轉文字在本機執行。本工具不會把錄音上傳到 OpenAI，也不會儲存錄音或逐字稿。首次安裝套件與下載模型時仍需要網路連線。
Local push-to-talk voice input for Windows 11, powered by faster-whisper.
Hold the keyboard's Right Ctrl key to record. Release it to transcribe, convert Chinese output to Traditional Chinese, and paste the result at the active caret.
- Local transcription after the model has been downloaded
- NVIDIA CUDA
float16acceleration - Right Ctrl push-to-talk
- Traditional Chinese conversion through OpenCC
- Works in browsers, chat applications, editors, Word, Notion, and Obsidian
- Hidden startup with no command window
- Restores the previous text clipboard after pasting
- Tray icon for ready, recording, transcribing, pause, and exit states
- Runtime microphone selection from the tray menu
- Automatic recovery when Windows changes a microphone's numeric device index
- Automatic fallback to another interface for the same microphone when the current Windows audio interface fails
Audio is captured in memory and passed directly to the local transcription model. This application does not upload recordings to OpenAI.
The first launch downloads the selected Whisper-compatible model from its model host. Package installation also requires internet access. Review your network and model-host policies before use.
The project does not save microphone recordings or transcripts. Diagnostic logs and configuration are stored under:
%LOCALAPPDATA%\WhisperRightCtrl
- Windows 11
- Python 3.11, or the
uvpackage manager with access to Python 3.11 - NVIDIA GPU with a compatible driver for the default CUDA configuration
- A working microphone
- Internet access during installation and first model download
The tested reference hardware is an NVIDIA RTX 4060 Ti 16 GB. Other supported NVIDIA GPUs may work. CPU mode can be selected manually in the configuration file, but is not the default release target.
- Download and extract the release ZIP to a permanent folder.
- Double-click
INSTALL.cmd. - Select the microphone number that receives your voice.
- Wait for installation and tests to complete.
- On first launch, allow time for the
turbomodel to download.
The installer creates a project-local .venv and a per-user Windows Startup entry. Administrator rights are not required for the Startup entry.
- Place the caret in a text field.
- Hold Right Ctrl and speak.
- Release Right Ctrl.
- Wait for the recognized text to appear.
Tray colors:
- Green: ready
- Red: recording
- Orange: transcribing
- Gray: paused
Right-click the tray icon to pause or exit.
Run CONFIGURE_MICROPHONE.cmd, choose another microphone number, and restart the application from the tray.
The selected device and other settings are stored in:
%LOCALAPPDATA%\WhisperRightCtrl\config.json
Available settings include:
{
"input_device": 2,
"language": "zh",
"model": "turbo",
"device": "cuda",
"compute_type": "float16",
"minimum_seconds": 0.25,
"restore_clipboard": true
}Device numbers can change after audio hardware or drivers are changed. Run microphone configuration again if recordings become silent.
Open the log:
%LOCALAPPDATA%\WhisperRightCtrl\voice-input.log
If the VAD reports that it removed the entire recording, select the correct microphone with CONFIGURE_MICROPHONE.cmd.
Check the log for missing CUDA libraries, an invalid microphone, or model download errors. Re-run INSTALL.cmd to repair project-local packages.
Windows may block input injection into an elevated application when Whisper Right Ctrl is running without elevation. Run both at the same integrity level.
Run UNINSTALL.cmd. It removes automatic startup and stops the resident process. It deliberately leaves the extracted project, model cache, and local configuration in place so removal is recoverable.
Delete those folders manually if you also want to remove downloaded files.
The release builder creates its own validation environment if needed, installs the project, and runs the full test suite:
BUILD_RELEASE.cmd
This runs the tests, scans the source tree for the original developer's personal path, excludes generated environments and logs, and creates a versioned ZIP under dist.
This tool globally observes the Right Ctrl key and uses the clipboard to paste recognized text. Review the source before use. Do not download unsigned release packages from untrusted mirrors.
Project source code is available under the MIT License. Dependencies and downloaded models retain their own licenses; see THIRD_PARTY_NOTICES.md.