Unlimited local speech-to-text dictation for Windows 11.
Hold a shortcut, speak, and let go. The text lands at your cursor. No account, no cloud, no weekly limit.
Grab the newest release from the button above or from releases/latest. Every release ships three files:
| File | What it is |
|---|---|
FreeDictate-<version>-Setup.exe |
The installer. Per-user by default, no admin prompt. |
FreeDictate-<version>-Portable.zip |
Portable build. Extract anywhere and run FreeDictate.exe. |
SHA256SUMS |
SHA-256 checksums for the files above. |
- Download
FreeDictate-<version>-Setup.exe(or the Portable zip). Both are frozen executables that bundle everything, the Visual C++ runtime included: no Python, no dependencies, nothing else to install. - Run the Setup exe. It installs to your profile by default (no admin needed; you can pick all users on the first page), offers a startup shortcut and an optional desktop icon, and adds an entry to Apps › Installed apps for clean removal.
- On the first launch, pick a voice model in Settings. It downloads once, then dictation is fully offline.
The builds run on 64-bit Windows 10 1809 or later; the window materials need Windows 11 22H2. GPU acceleration is optional: on a PC with an NVIDIA card, Settings → Speech model downloads NVIDIA's libraries in one click (only the NVIDIA driver is needed).
Updating is built in: Settings → General → Check for updates. FreeDictate downloads the new build from GitHub Releases, verifies it against SHA256SUMS, and restarts itself. An update whose release has no checksum for it is refused, not installed.
The portable setup is its own section: Portable.
- Click any text field.
- Hold Ctrl+Win (the default hotkey) and speak.
- Release. The transcript is pasted at your cursor.
The stop key (Esc by default, changeable in Settings) drops a take at any point: while you talk, or while it's being written out. In toggle mode a short press starts and a second press stops; in both mode a tap latches and holding pushes to talk. A small overlay near your cursor shows the input level and a timer while you speak.
Extract the zip wherever you like, a USB stick works, and run FreeDictate.exe. A portable.txt marker next to the exe tells FreeDictate to keep everything (settings, history, downloaded models) in a data\ folder beside it, so the whole folder moves with you and nothing is written to the registry or AppData. The portable build can update itself the same way, keeping your data\ folder.
Settings, page by page
Everything lives in one window with seven pages:
- Dictation: hotkey and mode, microphone with a live level test, sounds (start, stop, and a tone when the text arrives) and volume, auto-stop on silence for press-to-start takes, vocabulary hints passed to the model, pause, keeping other apps off the mic while a take runs, and muting or turning down other apps' sound during a take.
- Speech model: pick a model, from Whisper Tiny (75 MB, fastest) to Large v3, or NVIDIA's Parakeet v3 for English and 2 more languages by CPU design. CPU or NVIDIA GPU. One list picks the model; a second deletes models you no longer use; Verify model files re-checks each byte against the manifest. A fallback model keeps dictation working while a new one downloads.
- Text and pasting: filler removal, repeated-word removal ("the the" becomes "the"), capitalization, end punctuation, spoken punctuation and lists ("comma", "full stop", "question mark", "hyphen", "open bracket", "dot point", "new line", "new paragraph"; say "literal" first to type the word), an optional AI tidy-up, find-and-replace rules, "submit" ending a take with Enter, output method (paste / Shift+Insert / type / clipboard only) and clipboard restore.
- Appearance. Ten themes plus Match Windows, the Windows accent colour, any installed font, and the window material: Mica Alt, Mica, Acrylic, Blur or solid, with a tint slider. Every slider has a value box you can type a number into. Backdrops need Windows 11 22H2 and transparency turned on; otherwise you get the solid look automatically. The title bar can be FreeDictate's own or the standard Windows one; with the Windows title bar the material runs under it and the tint covers the body.
- History: searchable dictation history in SQLite on your PC, retention in days, clear or export. When the AI tidy-up changed a take, you can copy what you originally said. Transcribe audio and video files here too.
- General. Start at login (normal or elevated), update settings, prerelease channel, export/import settings.
- About. Version, licences, links.
Speech models and accuracy
| Model | Size | Good for |
|---|---|---|
| Tiny | 75 MB | Short notes, slow PCs |
| Base | 145 MB | Fast and decent |
| Small | 484 MB | The default; good balance |
| Medium | 1.5 GB | More accuracy |
| Distil Large v3.5 | 1.5 GB | English. About as accurate as Large v3, and faster |
| Parakeet v2 | 661 MB | English. The most accurate here, and fast on the CPU alone. Doesn't use vocabulary hints |
| Parakeet v3 | 670 MB | 25 languages, more accurate than v2, and it runs on the processor by design |
| Large v3 Turbo | 1.6 GB | Close to Large v3, much faster; best with a GPU |
| Large v3 | 3.1 GB | The most accurate Whisper model, any language; best with a GPU |
| multilingual variants | varies | 30+ languages via Detect automatically |
With an NVIDIA GPU, Settings offers a one-click CUDA support download. Without it, everything runs on CPU.
Accuracy, measured on 30 LibriSpeech clips and 21 accented clips (Scottish, Irish, southern British, Indian English) from EdAcc, as word error rate (lower is better):
| Model | Clean speech | Accented speech |
|---|---|---|
| Parakeet v2 (CPU, int8) | 2.6% | 9.5% |
| Large v3 (GPU) | 4.2% | 10.2% |
| Distil Large v3.5 (GPU) | 5.1% | 10.9% |
AI tidy-up (optional)
Optional, under Text and pasting. A small writing model on your PC removes stumbles, false starts and repeats, and punctuates run-on takes, keeping your words and their order. If the answer drops much of what you said, changes a number, or reads like a reply rather than your words, FreeDictate pastes the plain transcript instead.
FreeDictate runs the model itself with llama.cpp (MIT licence). Settings downloads the engine (32 MB, checked against a pinned SHA-256) together with one of two models: Light (1.1 GB, for graphics cards under 6 GB) or Standard (2.5 GB, about a second per take on a recent card). It needs a graphics card from any maker; on the processor alone a clean-up takes about 11 seconds, so the setting says so and stays off. The writing model loads only for takes that need it and leaves graphics memory after a few minutes without one (5 by default, adjustable under Speech model › Memory).
Recording, transcription, history and settings stay on your PC. The AI tidy-up runs on your own PC. The only network traffic is the model downloads you choose and the optional update check against the GitHub Releases API, both visible and switchable in Settings. No telemetry, no accounts.
Quit Wispr Flow or rebind its push-to-talk key before starting FreeDictate. Both listen for the same hotkey, so running both pastes twice.
Building from source
pwsh -File scripts\build.ps1runs the test suite, freezes the app with PyInstaller, self-tests the frozen exe, compiles the Inno Setup installer, and zips the portable build into dist\. Pass -SkipInstaller, -SkipTests or -Python <path> as needed; see the header of scripts\build.ps1. Source layout: src\freedictate (core), src\freedictate\ui (settings window and overlay), tests\, packaging\.
Bug reports with a log file from General › Your data, feature ideas and pull requests are all welcome. See CONTRIBUTING.md for setup, checks and what to include.
Thank you to everyone using FreeDictate, reporting bugs and sending fixes.
- Author: Myself, DavidHiFi
- FreeDictate is open source under the MIT licence.
| Who | What |
|---|---|
| OpenAI (Whisper) | The Whisper speech models, from Tiny to Large v3 |
| SYSTRAN (faster-whisper) | The Whisper runtime that transcribes the Whisper models, on its CTranslate2 engine |
| NVIDIA (Parakeet) | The Parakeet v2 and v3 speech models, loaded with onnx-asr |
| VoxType (peteonrails/voxtype) | Text-processing details the options mirror: filler removal, punctuation spacing and the "submit" auto-submit |
| Qwen (Qwen3) | The writing models behind the optional AI tidy-up |
| ggml-org (llama.cpp) | Runs the writing models on your PC; FreeDictate downloads its Vulkan build |
| Qt project (Qt for Python) | The toolkit behind every window, tray icon and page |
| sounddevice (python-sounddevice) | Microphone capture, over PortAudio |
| PyAV (PyAV) | Reads audio and video files so History can transcribe them |
| pycaw (pycaw) | Windows audio controls that mute or turn down other apps while you dictate, over comtypes |
| dofuuz (python-soxr) | Resamples every capture down to 16 kHz for the models |
| psutil (psutil) | Core and thread counts behind the CPU limit slider |
Built on Python, with PyInstaller freezing the builds and Inno Setup compiling the installer.
Missing or wrong credit? Open an issue and I'll fix it.
Most of this project's code was developed with AI coding agents, mainly Claude Code. No single AI agent made the project, and every change was tested before it shipped in a release.
The optional AI tidy-up is a feature that runs on your own PC. It is off until you turn it on.
Not affiliated with OpenAI, NVIDIA or Wispr. Requires 64-bit Windows 10 1809 or later; the window materials need Windows 11 22H2.