Step 1. Get the app
Download the unsigned Windows x64 portable ZIP for Windows 10 or 11. Quit any older Saydraft instance from its tray, extract the whole ZIP and run Saydraft.exe. The portable includes its Python runtime. First-run setup guides you through the choices below; no installer is required.
Step 2. Choose your microphone
On Microphone, keep Default Windows input or choose Refresh inputs to find a named device. Test input runs a ten-second, level-only check without transcription or saved audio. Check microphone privacy reads Windows permission status without opening an input. If access is blocked or unknown, follow the microphone-privacy link to review desktop-app access in Windows settings.
Step 3. Review and install local models
The Models step also offers optional local AI cleanup. Leave it off to use rules immediately, or review a cleanup model’s size/license and explicitly download it. Finish saves the staged choices; enabling AI or choosing a model starts no transfer.
Check this PC gives a provisional capability recommendation. Use speech recommendation selects the suggested speech model; it does not benchmark it, enable AI cleanup or download anything.
Review the download size and licenses, then choose Download / Resume. Progress shows transfers and checksum verification; Pause is available. Installed packages are checked in the background and reused on later launches.
After installation, Check model performance measures the selected speech model with bundled audio, without a microphone or download. Preparation is separate from three timed samples. Stop check waits for native work to finish. Unavailable AI cleanup speed stays unmeasured, and a fast sample is not an everyday latency guarantee. See model choices.
For Arabic, Bengali, Chinese, Hindi, Indonesian, Japanese, Korean, Malay, Thai or Urdu, add the optional shared All languages model. One download serves those languages; selecting a language starts no transfer. Portuguese, including Brazil, is also covered by the default model. See language setup.
Next requires the selected model to be ready. Catalog packages must pass their file checks. An existing local Whisper directory can instead pass an explicit offline performance check while its files remain unchanged. Finish checks the current model again before saving. Cancelling setup preserves saved preferences; downloaded files remain separate.
Step 4. Choose a shortcut and practice
Existing custom bindings are preserved. A fresh profile uses Ctrl + Win to dictate, with F9 as the alternate and Ctrl + Win + Space for hands-free. Learn shortcut listens while setup stays focused, for up to ten seconds; Escape cancels. Manual entry and conflict checks are also available.
Try it has chat and email fields inside setup. Preparing an installed model opens no microphone. Hold your shortcut or choose Record, then release or choose Stop and preview. Each practice recording is limited to one minute. Review raw words, cleaned text and edit chips. Practice audio stays in memory, creates no History entry and uses no external paste or clipboard. You can continue without recording.
If a stop or Settings change ends practice, choose Prepare practice explicitly when ready. If setup cannot confirm a stop, cancel and restart Saydraft as prompted.
Step 5. Make it sound like you
Style quick pick offers Casual or Formal for Messages and Email. Its rules preview uses your last practice transcript or a labelled example. Finish saves changed choices and opens Home after practice stops; existing cleanup intensity and AI settings stay unchanged.
Later, use Hub → Style for app-group tones and intensity, Dictionary for names, and Snippets for frequently used text. AI cleanup is optional; rules work without a text model.
Step 6. Try it offline
With compatible speech files installed, open an ordinary text editor and put the cursor in an editable field. Hold your shortcut, speak, then release. Disconnect Wi-Fi and try another sentence. Shift + Alt + Z pastes the last saved dictation. Review offline verification for what the check proves and its boundaries.
If something feels wrong
Check that the model is installed and the microphone meter responds. Start in a non-elevated text editor; password and protected fields are excluded. History keeps a saved transcript when an app cannot accept insertion. Continue with troubleshooting and keep private audio or text out of public issue reports.