Offline voice dictation on Windows

A practical guide to installing WritHer, turning speech into text in your apps and adding a local assistant when you need one.

Updated 24 October 2026

What you need

• Windows 10 or 11 and a working microphone.
• Internet access for setup. WritHer downloads a speech model on first launch. If you add the assistant, download its language model too.
• A model that fits your PC. CPU dictation is supported. A compatible NVIDIA GPU can accelerate inference when its runtime is configured, but is not required.

Dictation and the voice assistant are separate. You can start transcribing without installing Ollama. The Windows download includes the application; Python 3.11+ is required when you choose to run the source instead.

Install WritHer on Windows

1. Open the latest GitHub release and download its Windows ZIP.
2. Extract the entire archive to a folder. Run WritHer.exe from the extracted folder.
3. Leave the internet connection available while the speech model downloads and loads. The first start can take longer than later launches.
4. When the app is ready, use the system-tray menu to open Settings. Check the microphone and recognition language.

For source installation and optional GPU configuration, follow the repository installation instructions. Model choices and packaging can change between releases; the release notes describe the version you download.

Dictate into your first app

1. Open a simple text editor or an email draft and click the text field.
2. Hold AltGr and say a short sentence.
3. Release the key. WritHer transcribes the audio and pastes the result into the focused app.

The floating widget shows the recording and processing state. After the speech model is available locally, this dictation step does not need internet access. The target application must accept pasted text.

In Settings you can change the shortcut, select a microphone, choose a speech model, and switch between hold and toggle recording. In toggle mode, one press starts recording and the next stops it. Use the recording-duration limit as a safeguard.

Make the vocabulary yours

Add spoken-to-written vocabulary entries for names, abbreviations and specialist terms. Priming terms provide additional context to speech recognition. If you dictate symbols, enable Symbol & spelling mode and try “forward slash” or “semicolon”.

Recognition language is independent of the interface language. Use Auto for mixed-language recordings, or choose an explicit recognition language when that better fits your work. A clear microphone signal matters as much as the model choice.

Add a local voice assistant

The assistant can create notes, reminders and appointments. Install Ollama and download a compatible language model, or use an OpenAI-compatible local server such as llama.cpp or LM Studio. Configure the provider and model in WritHer Settings.

Hold Ctrl + Alt + R, say “Save a note: review the first draft”, then release. Model capabilities and speed vary; use a model suited to the assistant's tool-calling workflow and to your hardware.

If something does not work

The microphone is silent

Check Windows microphone permissions and the microphone selected in WritHer Settings. Confirm that Windows receives audio from that device. If you changed or disconnected devices, select the intended microphone again and restart WritHer if necessary.

Text does not appear

Click the target text field before recording. Test in a simple editor and check whether ordinary clipboard paste works. If the target app runs as administrator, a privilege mismatch can affect input injection. Review WritHer's log or recovery output before repeating a long dictation.

The first launch or transcription is slow

The first launch includes downloading and loading a model. Later transcription speed depends on the model and hardware. Try a smaller model or check your compute settings. GPU execution needs the compatible NVIDIA runtime; CPU mode is a useful starting point.

The assistant cannot connect

Start the local model server, check its address and model name in Settings, and confirm the model is downloaded. A successful dictation does not by itself mean the assistant server is running.

Help make the next release better

If you find a problem, open an issue with your Windows version, WritHer version, selected model and clear steps to reproduce it. Remove private text from logs before sharing them. Ideas, translations and documentation improvements are welcome too.

Read the WritHer FAQ for more answers, or see the contribution guide to make the project your own.