Use this command to install HoldToSpeak with WinGet:
winget install --id=rootMonsteR.HoldToSpeak -e
HoldToSpeak is push-to-talk dictation for Windows that runs entirely on your own machine.
Hold Ctrl+Win anywhere, speak, let go, and your words appear at the cursor in whatever app
you were already using.
Your voice never leaves your computer. There is no account, no cloud service, no telemetry
and no analytics. After a one-time speech-model download the app works with the network
cable unplugged, and it is built so you can verify that rather than take it on trust.
It works in the apps that usually break dictation - terminals, WSL, IDEs, browsers and chat
apps - by choosing a text-injection route per target rather than relying on clipboard paste.
Dictation drops to verbatim automatically in terminals and code editors, so commands are
never rewritten. Password fields are detected live via UI Automation and refused rather than
typed into. A plain-text personal dictionary lets you correct jargon permanently.
Speech recognition uses the Parakeet TDT 0.6B v2 model, which is downloaded on first run and
is licensed separately under CC-BY-4.0. Note that release binaries are not code-signed yet,
so Windows SmartScreen will show a warning on first run.
README
Hold Ctrl + Win anywhere in Windows, speak, let go — your words appear at the cursor in whatever
app you were already using. No account, no cloud, no telemetry. After a one-time model download it
works with the network cable unplugged, and it is built so you can verify that rather than take
our word for it (PRIVACY.md).
The overlay while you hold the keys — a live voiceprint, the active cleanup mode, a signal meter. Gone the moment you let go.
> Status: early. The dictation core works and is used daily by its author. A settings UI and
> code signing are still to come. Bug reports welcome; rough edges expected.
Why this exists
Cloud dictation is fast and good — and it ships your microphone to someone else's computer. That's
a non-starter in a lot of workplaces (legal, medical, finance, anywhere under an NDA), and it's
just unnecessary: a 2024-era speech model runs comfortably on a normal CPU.
What it does
Works in the apps that usually break. Terminals, WSL, IDEs, browsers, chat apps, remote
sessions — via a per-target injection chain rather than clipboard-paste-and-hope.
Never rewrites your commands. Focused on a terminal or code editor? Dictation drops to
verbatim automatically, so kubectl get pods -n kube-system stays exactly that.
Refuses password fields. Detected live via UI Automation — if the focused control is a
credential field, nothing is typed.
Learns your jargon.learn cube ctl => kubectl and it's fixed permanently, in plain text
you can edit.
Tells you when it can't help. Elevated window? You get "text kept, not inserted" — never a
silently swallowed sentence.
Install
Requires Windows 10/11 (x64) and a microphone.
Grab the installer or the portable ZIP from the
latest release. The installer is
per-user — it lands in %LOCALAPPDATA% and never asks for admin rights. The release notes carry
the SHA-256 of every artifact.
On first launch it downloads the speech model (~460 MB, once) and tells you it's doing so. After
that it needs no network at all.
> The binaries are not code-signed yet, so Windows SmartScreen will show a
> "Windows protected your PC" warning on first run — click More info → Run anyway, or verify
> the download against the published hash. Signing costs money and is on the list; until then,
> building from source below is the paranoid-friendly path.
A winget manifest is submitted and awaiting review
(microsoft/winget-pkgs#420408); once it
lands, winget install rootMonsteR.HoldToSpeak will work too.
Plus the tray icon (modes, overlay themes, quit), and in the console window: m to cycle mode,
learn => to teach it a word permanently, q to quit.
Rebind in %APPDATA%\HoldToSpeak\hotkeys.toml — off disables one:
ptt = "Ctrl+Win" # modifiers only
cycle_mode = "Ctrl+Alt+M"
cycle_style = "Ctrl+Alt+O"
quit = "Ctrl+Alt+Q"
Cleanup modes
Raw — exactly what you said, plus your dictionary. Forced automatically in terminals/IDEs.
Auto — light tidy: um/uh removed, sentence casing, terminal punctuation.
Polish — Auto plus conversational scaffolding removed: "I'm just testing this, like, you
know, testing the software and stuff, just to make sure…" becomes "I'm just testing this,
testing the software, just to make sure…"
All three are rule-based and run entirely offline, so they can only ever delete filler — they
can never invent or reword anything. Polish only removes a marker when it's comma-delimited, which
is what keeps "I like this" and "you know the answer" intact.
Overlay themes
Four looks for the push-to-talk overlay. Every one tells you the same three things — that the mic
is being heard, which cleanup mode is active, and how strong the signal is — so switching theme
changes how it looks, never what it tells you. Cycle live with Ctrl + Alt + O or from the
tray; set a default with overlay_style in %APPDATA%\HoldToSpeak\settings.toml.
<img src="assets/readme/overlay-hud.png" alt="HUD theme" width="100%" /><b>HUD</b> (default) — tactical comms: voiceprint, timecode, mode callout
<img src="assets/readme/overlay-volt.png" alt="Volt theme" width="100%" /><b>Volt</b> — an electric beam with lightning forks that ride your voice
<img src="assets/readme/overlay-wave.png" alt="Wave theme" width="100%" /><b>Wave</b> — flowing layered waveform, tinted by the active mode
<img src="assets/readme/overlay-bars.png" alt="Bars theme" width="100%" /><b>Bars</b> — frequency spectrum, the classic visualizer
Privacy you can check
The whole claim is in PRIVACY.md, including the one network request the app ever
makes. The short version is a firewall rule: after the one-time model download, block the app from
the network entirely — it keeps working.
Ctrl+Win ──► keyboard hook ──► always-on mic ring (400 ms look-back, so your first
word is never clipped)
│ release
▼
local ASR (Parakeet, CPU) ──► deterministic cleanup
│
▼
injected at the cursor (route chosen per target app)
Rust throughout, with every Win32 call confined to one crate behind a trait wall — enforced in CI,
which is also what makes a macOS port a swap rather than a rewrite.
Contributing
Issues and PRs welcome. cargo test --workspace and cargo run -p xtask -- check-layering must
pass; CI runs fmt, clippy (-D warnings), tests, and the layering check.
License
MIT. Third-party components and the speech model's required attribution are listed in
THIRD-PARTY-NOTICES.md.
