VoiceWave

Free and open source · Works offline after a one-time model download · English

How it works

1. Hold Ctrl + Win and talk. Double-tap the shortcut to go hands-free; Escape cancels.
2. Let go. Your speech is transcribed on your PC: on the graphics card through Vulkan when it qualifies, otherwise on the CPU.
3. The text lands at your cursor, in whatever app you were using.

What you get

Never leaves your PC

There is no VoiceWave server and no account. Audio stays in memory and is transcribed on your machine. VoiceWave goes online for three things only:

• Model downloads, once: the speech model at setup, and the AI polish model if you turn it on.
• Update checks against GitHub. Updates are signed and verified before they install.
• Anonymous usage and problem reports, on by default, which you can turn off during setup or in Settings → Diagnostics. What they send

AI polish, on your machine

Optional and off by default. A small local model (Qwen2.5 1.5B through llama.cpp) tidies fillers, casing and punctuation, with profiles for Standard, Coding and Writing (Literal skips it). A fidelity check rejects rewrites that drift from what you said and keeps the plain transcript instead. It is a one-time download of about 1 GB and needs a CPU with AVX2.

Install

1. Download VoiceWave.Local.Core_<version>_x64-setup.exe from the latest release and run it.
2. Windows may warn "Windows protected your PC" (Unknown publisher). The installer is not code-signed yet, so SmartScreen does not recognise it. Choose More info → Run anyway. Every installer is built in public by GitHub Actions from this repository, and in-app updates are signed and verified before they install.
3. First launch checks your microphone and downloads one speech model, chosen for your PC: about 488 MB (Small) up to 1.6 GB (Large Turbo). After that, dictation works offline.

Or install with winget install VoiceWave.LocalCore.

Requirements

	Needs
Windows	10 or 11, 64-bit
CPU	x64 with SSE4.1. AVX2 is needed for the GPU engine and AI polish.
GPU (optional)	A discrete NVIDIA, AMD or Intel graphics card with Vulkan 1.2. PCs with only integrated graphics use the CPU engine for now.
Language	English

Compared with cloud dictation

	VoiceWave	Wispr Flow
Price	Free	$15/month for Pro; free tier capped at 2,000 words/week
Where speech is transcribed	On your PC	On their servers
Works offline	Yes, after the model download	No
Account required	No	Yes
Open source	Yes (Apache-2.0)	No

Wispr Flow details from wisprflow.ai/pricing, April 2026.

Anonymous usage and problem reports

On by default for new installs. They show us whether setup and dictation work for you, and what breaks, so we can fix it. Reports go to PostHog on its EU Cloud, in a project set to discard IP addresses. If you shared the older usage count, VoiceWave asks you once; until you answer, reports wait on your PC and nothing is sent. One switch turns everything off, during setup or in Settings → Diagnostics. Details are in the privacy notice.

Every report carries a random install ID and a random ID for each launch, the app version, Windows 10 or 11, your graphics card maker, whether speech runs on the graphics card or the processor and which model family, how you installed VoiceWave (winget, a direct download, or unknown), which version you first installed and how many days ago (estimated for installs older than 0.7.7), and whether the PC is a virtual machine. On top of that, VoiceWave sends only:

• when the app starts: whether it started with Windows, and whether the last run ended normally, crashed, hit an internal error, or was cut off by a restart or an update, what it was doing then and how long it had run;
• setup steps, and whether setup was finished or skipped;
• once for each day you dictate: how many dictations were started, worked, came back empty, failed, heard no speech, were retried or were hands free; how fast text appeared; how long you spoke; how many words were inserted; how often speech ran on the graphics card; how the text got in (typed, pasted, left on the clipboard, kept in History only, or held back from an app running as administrator); how many went into each kind of app (browser, code editor, terminal, chat, office, notes, AI app, or other); and your formatting style;
• whether your first dictation ran on the graphics card or the processor;
• failed dictations by type, with the kind of app when the text couldn't be inserted;
• speech helper crashes by cause, internal errors by area, and graphics card to processor fallbacks;
• model downloads: which model, how they ended, where a failed one stopped, retries, how far they got and how long they took;
• failed updates: the step, the kind of error and the version;
• shortcut setup problems;
• your answer to "Where did you hear about VoiceWave?", asked once at the end of setup on a new install;
• the reason you pick in "Something wrong?", and where you opened it.

It never sends audio, transcripts, dictionary or snippet content, window titles, app names, file paths, device names, error messages, account details, or your IP address. An app's name is turned into a kind of app on your PC and dropped there. 0.7.7 also still sends the older anonymous usage count through TelemetryDeck, under the same switch; that stops in the next release.

Support file. Settings → Diagnostics → Save support file writes a zip with your diagnostics report and recent logs to your Downloads folder. It never leaves your PC unless you attach it to an issue yourself.

What's new in 0.7.0

• GPU dictation on any vendor. whisper.cpp runs on discrete NVIDIA, AMD and Intel graphics cards through Vulkan, in its own worker process. No CUDA download.
• Picked per PC. A one-time check confirms the GPU works before VoiceWave switches to it, and dictation steps down to the CPU automatically if the GPU misbehaves.
• Verified GPU models. whisper.cpp models are checked against pinned SHA-256 hashes before they replace anything.

Full history in the changelog.

Support

Something not working? In the app, Help → Report a problem opens a pre-filled GitHub issue with your app version, Windows version and speech engine. You can also open an issue directly. Logs never contain what you dictated.

Build from source

Layer	What it is
App shell	Tauri 2, Rust core (audio capture, hotkeys, insertion, storage)
Interface	React 18, TypeScript, Tailwind
GPU speech	whisper.cpp 1.9.5 worker processes (voicewave-asr-vk.exe, voicewave-asr-cpu.exe)
CPU speech	faster-whisper in a bundled Python runtime
AI polish	Qwen2.5 1.5B through llama-cpp-python, on the CPU

npm install
npm run tauri:dev     # desktop app (needs the Rust + Tauri prerequisites)
npm run dev           # interface only, in a browser
npm test              # frontend tests
npm run build         # type-check and build the interface

The release build stages the Python runtime and the whisper.cpp workers; see release.yml and the scripts under scripts/. More in CONTRIBUTING.md.

License

Apache-2.0. Contributions welcome: see CONTRIBUTING.md.

Speech models are downloaded separately under their own licenses:

• Parakeet TDT 0.6B v3 by NVIDIA, the default model, licensed under CC BY 4.0. VoiceWave uses the GGUF conversion by ggml-org as published, without changes.
• Whisper by OpenAI (Small and Large Turbo), MIT.