KeyReed 0.4.0MIT or Apache-2.0

Your voice.
Your words.
Your machine.

Open-source voice dictation for Windows. Speak naturally, turn your words into polished text, and use it anywhere you work.

  • Open source
  • Local-first
  • Windows x64
  • Six writing modes
  • No dedicated GPU required

Installer 17.8 MB. The speech model (about 460 MB) downloads once from Settings or the tray after install. The build is unsigned, so Windows SmartScreen may warn. What that means.

Illustration of the dictation flow, not a recording of the app.

MIT or Apache-2.0Dual licensed, source on GitHub
Local by defaultNo network call in the default profile
CPU onlySpeech and grammar models run on your processor
Pinned modelsSize and SHA-256 checked before load

How it works

Hold. Speak. Release. Paste.

One shortcut sits between your thought and the text field. Nothing to switch to, nothing to open.

  1. 01
    Right Ctrl

    Hold the shortcut

    Right Ctrl by default. The microphone opens only while you hold it, and closes when you let go.

  2. 02

    Speak naturally

    Say it the way you would say it out loud. Audio stays in memory for the length of the recording.

  3. 03

    KeyReed transcribes and processes

    A local speech model transcribes. Deterministic rules and an optional local grammar model clean it up.

  4. 04
    CtrlV

    Paste it anywhere

    The finished text lands on your clipboard. Turn on Auto Insert if you want it typed for you.

Product

Six writing modes. One voice.

Pick how KeyReed should write, or let it adapt to the app you are in. Choose a mode below to see an illustrative before and after.

Smart

You said

KeyReed writes

Illustrative example. Not a live connection to the app.

Smart, Professional, Casual and Technical use the local grammar model when it is installed. Without it, KeyReed applies deterministic rules only, so rewording will be lighter than shown. Exact and Terminal / Code never rephrase.

Features

Built for the way you already work.

Every capability below exists in the current source. Each one is here because of what it changes for you.

System-wide dictation

It follows your cursor, not a window.

KeyReed lives in the system tray and listens for a global shortcut, so dictation works in whatever you are typing into. With automatic app-context adaptation on, it picks a writing mode to match the app in front of you.

Output goes to the clipboard by default. Auto Insert types into the focused field instead, and falls back to the clipboard for password fields or elevated windows.

Active appWriting mode
MailProfessional
ChatCasual
Documents and notesSmart
Code editorTechnical
TerminalTerminal / Code
Password-protectedExact

Automatic adaptation can be switched off in Settings.

Push-to-talk and toggle

Hold to talk, or tap to start and stop.

Hold Right Ctrl, speak, release. Prefer hands-free for longer thoughts? Switch to toggle-to-talk and press once to start, once to finish. The shortcut is a preset you pick in Settings.

Escape cancels a recording. A lost key release or the 120 second cap discards the audio instead of pasting something half-captured.

Right CtrlHold to talk
EscCancel
CtrlShiftSpaceCycle writing mode
CtrlVPaste the result

Local transcription

Speech recognition that runs on your CPU.

KeyReed uses sherpa-onnx to run speech models on the processor you already have. NVIDIA Parakeet TDT is the English default. Lighter and multilingual models are one command away, and nothing downloads unless you ask.

Every model file is checked against a compile-time size and SHA-256 before it loads. A half-finished download is never activated.

ModelDownloadRAM
Parakeet TDT 0.6B v2English, default
≈460 MB≈690 MB
Parakeet TDT 0.6B v325 European languages
≈465 MB≈710 MB
Whisper small.enEnglish
≈606 MB≈440 MB
SenseVoice Smallen, zh, ja, ko, yue
≈999 MB≈290 MB
Moonshine BaseEnglish, light
≈239 MB≈165 MB
Moonshine TinyEnglish, low-RAM PCs
≈103 MB≈95 MB

Approximate figures from the model catalog in the source. Model licenses differ; see each model's page.

Grammar correction

Cleanup that is allowed to say no.

First, fast deterministic rules remove fillers and stutters, resolve spoken self-corrections, and fix casing. Then, if a local grammar model is installed, it polishes the sentence. Its output is checked against guards and discarded when it drifts from what you said.

Terminal / Code mode compiles spoken flags into commands and warns before a destructive one. It does not try to improve your command.

  1. Raw transcriptfrom the local speech model
  2. Tier 0 rulesfillers, stutters, self-repairs, casing, your vocabulary
  3. Local grammar modeloptional, Qwen3.5 0.8B, runs on CPU
  4. Guardsreject chatter, number changes and runaway length, then fall back to Tier 0
  5. Finished texton your clipboard

Your controls

Yours to tune.

Teach KeyReed your names and jargon with a personal vocabulary. Choose the microphone, the shortcut, how text is delivered and which speech engine runs. Choose how private you want it, from fully offline to a hybrid that can use your own cloud key.

Settings are stored in %LOCALAPPDATA%\Keyreed. Startup with Windows is off until you turn it on.

Privacy

Your voice doesn't need to leave your computer.

In the default profile it doesn't. Here is how audio moves, and where an optional cloud route would fit in if you turn it on.

Microphoneopens while you hold the key
Local speech processingsherpa-onnx on your CPU
Text correctionrules, then optional local model
Active applicationclipboard, or Auto Insert
Optional cloud route

Only if you switch to Balanced or Maximum Accuracy and store an API key. Speech or text can then be sent to the endpoint you configured. The default profile never uses it.

Local by default

The default profile, Maximum Privacy (100% Offline), makes no network call for audio or text. A unit test in the source holds it to that.

One setup download

The small installer fetches a speech model once, over HTTPS, from the tray. The full offline archive ships with it and needs no download.

Audio stays in memory

Recordings are kept in RAM, are not written to a file on the normal path, and are wiped after decoding. Cancelled and timed-out recordings are discarded.

Verified models

Model files are checked against compile-time size and SHA-256 pins before they load. A partial or mismatched download is not activated.

Network, precisely

Needed to download a model, and for the cloud profiles. Not needed to dictate once a model is installed. Cloud keys are stored DPAPI-encrypted.

Know the clipboard

Finished text stays on the clipboard until you replace it. You can exclude it from Windows Clipboard History in Settings. That option is off by default.

Don't take our word for it: the source is public. We have not published an independent network capture of the offline profile yet, so read the code or run your own.

Open source

Not a black box. Open source by design.

KeyReed gives you control over your voice input. Explore the source, inspect the implementation, contribute improvements and make it your own.

  • Public repository with an inspectable Rust workspace
  • Dual licensed MIT or Apache-2.0, your choice
  • Contributions welcome, small and tested
  • Release audits and known limits documented in the open
  • Local-first architecture you can read end to end
crates/vi-core/src/types.rsBrowse crates

Compared

Love voice typing? Want more control?

Meet the open-source alternative to Wispr Flow for Windows. Wispr Flow is a polished, mature product. This is how the two differ, and where it is ahead.

KeyReed compared with Wispr Flow
CapabilityKeyReedWispr Flow
Source codeOpen source, MIT or Apache-2.0Commercial service
Where speech is transcribedOn your PC by defaultIn Wispr's cloud, per its docs
Works offlineYes, once a model is installedNo, an internet connection is required
Audio leaves your deviceOnly if you enable a cloud profileYes, sent for transcription
Writing stylesSix selectable modes plus per-app adaptationAuto-edits and AI commands
CustomisationVocabulary, shortcut, model choice, build from sourceDictionary, snippets, custom shortcuts
PriceFree. Cloud profiles use your own keyFree for 2,000 words a week. Pro $15/mo, or $12/mo billed yearly
PlatformsWindows 11 x64 (Windows 10 intended)Mac, Windows, iPhone, Android
LanguagesEnglish by default, more via optional models100+ languages
Maturity and team featuresEarly 0.1.x release, individual useTeam plans, admin controls, SOC 2 Type II

Open source. Local-first. Windows-focused. Yours to inspect and improve. KeyReed does not claim feature parity with Wispr Flow. The last three rows are where Wispr Flow is the stronger choice today.

Wispr Flow details come from its own pages, checked in November 2026: pricing, privacy, what is Flow. KeyReed is an independent project and is not affiliated with Wispr.

Download

Get KeyReed for Windows.

Start with the small installer. Choose the full archive only if you want everything in one offline download.

Recommended

Windows x64 installer

Version 0.4.017.9 MBKeyReed-Setup-x64.exe

Installs KeyReed and its runtimes. It does not include speech models: after launch, the 4-step first-run setup guides you through downloading the default speech model (about 460 MB, once, verified against compile-time SHA-256 pins).

Download installer
SHA-256f81da5b3530d7581d01f3f41fdda41231b4f2a68d7e22e2b4b291556d7722812

Full offline archive

1.56 GBkeyreed-0.4.0-windows-x64-full-offline.zip

Includes the speech models and the local grammar model, so it works with no setup download. Extract it and run keyreed.exe.

Download full archive
SHA-256a75ba1b69c752b37db517689b97d1e7279bb138bfe96b905464415680aa0f97e

About the SmartScreen warning

KeyReed's Windows files are not code-signed yet. Windows SmartScreen may show an unknown-publisher prompt or "Windows protected your PC". That is expected for an unsigned download and is not a verdict on the file, but it is a reason to verify it. Compare the SHA-256 above with your file, or build from source.

Get-FileHash .\KeyReed-Setup-x64.exe -Algorithm SHA256

Requirements

  • 64-bit Windows 11 (packaged and launched). Windows 10 is an intended target and was not re-tested for this release.
  • A microphone that desktop apps are allowed to use.
  • No dedicated GPU. Models run on the CPU.
  • Memory: on a 4 GB test machine the app and its local grammar runtime used about 1.6 GB together during a session. More free RAM is safer. An 8 GB PC has not been benchmarked.

FAQ

Questions, answered straight.

KeyReed is a Windows tray application for voice dictation. You hold a shortcut, speak, and release. It transcribes your speech, corrects it, and puts the finished text on your clipboard, or types it into the focused field if you turn on Auto Insert.

Yes. It is free, open-source software with no subscription. The optional cloud profiles use an API key you provide, and that provider's own pricing applies.

Yes. The whole application is a public Rust workspace on GitHub, dual licensed under MIT or Apache-2.0. You may use it under either license. The speech and grammar models it downloads have their own licenses.

KeyReed is open source, Windows-only, and transcribes on your own PC by default. Wispr Flow is a mature commercial product on Mac, Windows, iPhone and Android with 100+ languages, and its documentation says transcription happens in its cloud. If you need those platforms, languages or team features today, Wispr Flow is the stronger choice. If you want to inspect the code and keep audio local, KeyReed is built for that. See the comparison.

Yes, once a speech model is installed. The small installer needs a one-time internet connection to download the default speech model from the tray. The full offline archive already contains it. The default profile makes no network calls for dictation.

No. The packaged Windows build runs its speech and grammar models on the CPU. A dedicated GPU is not required, and GPU performance has not been measured.

Not by default. The default profile, Maximum Privacy (100% Offline), keeps audio and text on your computer. Only if you switch to Balanced or Maximum Accuracy and store an API key can audio or text go to the endpoint you configured. The endpoint fields are pre-filled with Groq's OpenAI-compatible API and are editable. KeyReed does not ship a key.

The package was built and launched on 64-bit Windows 11. Windows 10 (64-bit) is an intended target but was not re-tested for this release. There are no 32-bit, ARM, macOS or Linux builds.

Download KeyReed-Setup-x64.exe and run it. Windows may warn because the file is unsigned. Install, then open the tray icon and choose Download default speech model. Pick your microphone in Settings, hold Right Ctrl, speak, release, and press Ctrl+V. The full steps are in the installation guide.

Open Settings from the tray icon and choose a different option under Push-to-Talk Shortcut. Under Activation Behaviour you can switch between hold-to-talk and toggle-to-talk. Escape cancels a recording, and Ctrl+Shift+Space cycles the writing mode.

Read CONTRIBUTING.md, then open an issue or a small, tested pull request. Open an issue first for anything large. Bug reports with your Windows version and steps are very welcome.

This is an early release. The Windows files are unsigned. A full dictation loop on a physical desktop with a real microphone has not been recorded in the repository yet. Hindi accuracy is unscored, and Auto Insert was not replayed into Notepad on a physical desktop. The details are in the release documentation.

Your keyboard can take a break.

Download KeyReed and experience a more open way to turn speech into text.