Hands-free control of your Windows PC by voice.
Wheelhouse is free, open-source voice control for everyone — dictation into any application, voice commands, and clicking things by name, often faster and more comfortable than reaching for the keyboard and mouse. It is equally serious assistive technology: if using a keyboard and mouse is painful, difficult, or impossible, Wheelhouse aims to give you the whole computer by voice. It runs entirely on your machine by default: no cloud account, no subscription, no telemetry.
- Dictate anywhere. Speak into the focused application and watch the words stream in as you talk — the first word typically lands in under two seconds and the rest flow continuously, instead of appearing all at once after you stop. Spoken punctuation ("comma", "new line") becomes symbols.
- Voice commands. Switch windows, press keys, launch programs — driven by a pattern catalog you can extend with your own commands through the built-in Pattern Manager.
- Click by voice. Say
click cancelorclick the submit buttonand Wheelhouse finds the control in the focused window and clicks it. When names are ambiguous or unlabeled, sayapply numbersto badge every clickable control with a number andclick 5to pick one. - Offline by default. The default speech engine (NVIDIA Parakeet, running locally on your CPU) never sends audio or text anywhere.
- Careful about where text goes. Before typing a word, Wheelhouse checks that the focused control actually accepts text, so dictation does not spray keystrokes into the wrong place.
One command, in any PowerShell window:
irm https://github.com/wheelhouse-project/Wheelhouse/releases/latest/download/install-wheelhouse.ps1 | iexThe installer checks your hardware, installs its own Python environment (nothing global), downloads the offline speech model, and puts Wheelhouse in your Start menu. Details, prerequisites for the optional speech engines, troubleshooting, and uninstall instructions are in INSTALL.md.
Requirements: Windows 10 or 11 (64-bit), a microphone, and a few GB of disk space for the speech model. See INSTALL.md for the exact hardware guidance.
A note on security warnings: Wheelhouse releases are digitally signed; the installer's publisher shows as David Chesley Hite III, the project author. Windows SmartScreen may still warn for a while after each new release, until it has seen that exact file often enough — click More info, check the publisher name, then click Run anyway. INSTALL.md explains each warning, and the entire source code is in this repository if you would rather read the code and install from source (see CONTRIBUTING.md).
| Engine | Where speech is processed | When to choose it |
|---|---|---|
| Parakeet (default) | On your machine, CPU | No account, no cloud, works offline. The default for everyone. |
| Distil-Whisper (opt-in) | On your machine, NVIDIA GPU | You have a CUDA-capable GPU and want lower latency. |
| Google Cloud STT (opt-in) | Google's servers | You have a Google Cloud account and prefer its recognition quality; audio streams to Google while you dictate. |
Privacy is a safety property for a voice-control system — dictation can include passwords and medical text. The short version:
- No telemetry. Nothing is reported to the project or anyone else.
- Offline by default. With the default engine, audio and transcripts never leave your machine. Only the engines and AI features you opt into make network connections, and PRIVACY.md states exactly what each one sends.
- Logs don't contain what you dictate. By default, log lines record placeholders instead of recognized speech, at every log level.
- Broad local powers, disclosed plainly. Hands-free control requires the microphone, global input listeners, clipboard access, synthetic input, and reading the UI of the focused window. PRIVACY.md lists each power, why it is needed, and its limits.
| Document | What's in it |
|---|---|
| INSTALL.md | Installation in detail, optional engines, troubleshooting, uninstall |
| User help | Every voice command and feature, for daily use |
| llm/README.md | Load the user help into your own AI chat (ChatGPT, Gemini, Claude, Perplexity) |
| PRIVACY.md | Data flow, logging, and the capability disclosure |
| ARCHITECTURE.md | Process model, IPC, and the speech pipeline |
| CONTRIBUTING.md | Development setup, tests, and the contribution workflow |
| SECURITY.md | Reporting vulnerabilities |
The fastest path is the official Wheelhouse Assistant on ChatGPT, which always answers from the latest help document. A free ChatGPT account is enough — if ChatGPT says you "do not have access to GPT interactions", click Sign up for free in the upper right corner (no credit card required).
Prefer your own AI service? The user help is written so any AI chat service can answer questions from it. If you already use ChatGPT, Gemini, Claude, or Perplexity, you can turn the help document into a personal Wheelhouse support assistant: you need only one file — the help document to upload. Its assistant rules are embedded at the top, so there is nothing to paste. The llm/ folder explains the setup, and the steps for each service live on the project site: ChatGPT · Gemini · Claude · Perplexity.
Wheelhouse is a young open-source project with a single primary author. It has been the author's daily driver for years and reliability is the project's first value — but it has so far been validated on a small set of machines, so expect rough edges on hardware and applications it has not met yet. Bug reports are genuinely welcome, especially from users who depend on hands-free input: if Wheelhouse fails you, that is exactly the report the project needs.
Questions, or stuck on something? Email help@wheelhouse-project.org or open a GitHub issue.
- The default speech model is NVIDIA Parakeet TDT 0.6B (CC-BY-4.0, NVIDIA NeMo), served through sherpa-onnx.
- Wake-word detection uses openWakeWord community models.
- Notification sounds are from Pixabay.
- Full third-party attribution lives in NOTICE and PROVENANCE.toml.
Apache License 2.0 — see LICENSE.