Windows prerelease

Private dictation on your Windows PC

Hold Right Alt, speak, and release. Presspeech transcribes on your PC and normally pastes into the window where you started; if it cannot verify that destination, it keeps the result on the clipboard for manual paste. There is no account, subscription, telemetry, or cloud transcription.

Download Windows 0.1.11 Release notes and checksum First dictation guide Install on macOS

This prerelease is not code-signed. Windows may show Unknown publisher, and managed PCs or Windows 11 Smart App Control may block it. Verify that the download came from this repository and compare its SHA-256 checksum before deciding whether to run it. Never bypass a warning for a file from another source.

Before you download

Requirements and download size

Windows 11A 64-bit x64 PC. Windows 10 remains compatible but needs Extended Security Updates or an edition that remains supported; Windows on Arm is not supported.
4.4 GB installedThe installer includes the Python and speech runtimes; Python is not required separately.
NVIDIA CUDA optionalA current NVIDIA driver gives the fastest Parakeet path. CPU-only PCs use the smaller Whisper base.en model.
First-run modelAbout 2.5 GB with CUDA Parakeet, or about 141 MiB for the CPU default.
Install safely

Download, verify, and run

Presspeech remains compatible with Windows 10 x64, but Microsoft ended general Windows 10 support on 14 October 2025. Prefer Windows 11, or use Windows 10 only with Extended Security Updates or an edition that remains supported.

  1. Download the current installer and checksum

    Get Presspeech-Setup-0.1.11-x64.exe and its published SHA-256 file from the same GitHub release.

  2. Verify SHA-256 automatically

    In PowerShell, run the commands below from your Downloads folder. They validate the checksum file, compare it with the installer, and stop with an error instead of reporting success if anything differs.

    $installer = '.\Presspeech-Setup-0.1.11-x64.exe'
$parts = (Get-Content -LiteralPath "$installer.sha256" -Raw).Trim() -split '\s+'
if ($parts.Count -ne 2 -or
    $parts[0] -notmatch '^[0-9a-fA-F]{64}$' -or
    $parts[1] -ne (Split-Path $installer -Leaf)) {
  throw 'The published checksum file is not valid. Do not run the installer.'
}
$actual = (Get-FileHash -LiteralPath $installer -Algorithm SHA256).Hash
if ($actual -ne $parts[0]) {
  throw 'SHA-256 verification failed. Do not run the installer.'
}
"SHA-256 verified: $actual"
  3. Run the installer

    The current build is unsigned. If Microsoft Defender SmartScreen shows Windows protected your PC, confirm the filename and verified hash before choosing More info → Run anyway. If policy does not offer that choice, do not try to circumvent it.

  4. Launch Presspeech

    Open it from the Start Menu. It installs for your user account under %LOCALAPPDATA%\Programs\Presspeech and appears in Settings → Apps → Installed apps for normal removal.

Why the warning? Microsoft explains that unsigned downloads start without publisher reputation; each new unsigned version may prompt again.

First launch

Wait for the local model

  1. Check the selected path

    Presspeech tests the packaged CUDA runtime. A usable NVIDIA setup selects multilingual Parakeet TDT v3; otherwise a fresh installation selects English-only Whisper base.en for CPU use.

  2. Let preparation finish

    The first-run window shows the model, microphone, a selectable push-to-talk key and its global-listener status, and Start with Windows setting. Speak while its local microphone check runs. It briefly opens the selected input, discards those samples in memory, and distinguishes a detected input level from a connected-but-silent device or one that cannot be opened. If Right Alt enters @, €, or accented letters on your keyboard, Windows is using it as AltGr; choose F8 or another key so AltGr remains available for typing. Microphone and hotkey choices apply immediately, including in Try Dictation. Try Dictation becomes available when the model is ready; Finish Setup also requires the global hotkey to be ready. If preparation fails, choose Retry Speech Model; the window keeps tracking the new attempt. Choose Set Up Later to defer setup without marking it complete; the window opens again on the next launch with your microphone, hotkey, and Start with Windows choices kept. You can still finish after the model is ready if you plan to connect a microphone later. For keyboard-only use, press Left Alt with a command's underlined letter; Escape closes the current Presspeech window, and Ctrl+S saves Settings. If Windows refuses the requested startup registration, Setup stays open and reports the failure; choose Open Startup Settings to review Presspeech under Settings → Apps → Startup, then retry Finish Setup.

  3. Check microphone access

    If setup reports that the microphone is connected but no input level was detected, unmute it, speak, and choose Check Again. If the microphone cannot be opened, use the direct buttons to open Windows Microphone Privacy or Sound Input settings. Turn on Microphone access and Let desktop apps access your microphone, confirm the input, then check again. Unpackaged desktop apps share this privacy control, so Windows does not provide a separate Presspeech toggle. Outside the brief readiness check, Presspeech captures audio only while dictation is active; all captured audio is discarded after local processing.

  4. Try a private test

    Open the tray menu and choose Try Dictation…. Hold the push-to-talk key shown in Presspeech, speak, release, and check the result without pasting into another app. That key is Right Alt by default, or the alternative you selected during setup, such as F8 on an AltGr keyboard.

Everyday use

Hold. Speak. Release.

Dictate in your apps

Hold Right Alt by default. A high cue confirms recording; release to transcribe and paste. If a press is too short or the local recognizer detects no speech, the indicator briefly says No speech detected — try again and a Windows notification points back to Setup's microphone check instead of failing silently. The configured key is reserved for Presspeech while it is running, so it does not also open a Windows surface or invoke an F8–F12 command in the focused app; other keys and AltGr layout input continue to pass through normally. Press Escape during recording to cancel and discard the buffered audio without transcribing or changing the clipboard. The notification-area menu offers the same Cancel Dictation (Esc) action while recording. If Windows hides that icon in overflow, launch Presspeech again from the Start Menu to restore its window or open its controls; a second dictation process is not started. If the configured key stops responding while menu-based Dictate still works, choose Repair Global Hotkey in the notification-area menu, Setup, or Settings. Presspeech replaces the listener even when its thread still appears healthy and reports whether the replacement started. Hotkeys, hold/toggle behavior, microphone, text suffix, audio cues, models, and the 1/2/5/10-minute maximum recording length are configurable. Reaching that limit stops and transcribes rather than discarding the recording. Saving a different model starts preparing it immediately; Settings keeps its ready/error status visible and offers a retry, while dictation stays unavailable until preparation finishes.

Focus stays protected

A recording is bound to the window where it began. If focus changes during transcription, Presspeech leaves the text on the clipboard and notifies you instead of pasting into the wrong window. Windows blocks a standard app from sending the paste shortcut into a target running as administrator. Presspeech detects that boundary, leaves the transcript on the clipboard, and asks you to paste manually or reopen the target without Run as administrator; do not elevate Presspeech as a workaround.

Clipboard history and sync

Presspeech does not sync transcripts, but normal and recovery delivery writes finished text to the Windows clipboard. If clipboard history or cross-device sync is enabled under Settings → System → Clipboard, Windows can retain or sync that text. Third-party clipboard managers may also retain it. Review those settings before sensitive dictation.

Remote desktop aware

Moonlight and Microsoft Remote Desktop receive dedicated clipboard paste handling. Normal Windows apps retain the standard fast Ctrl+V path.

Updates require approval

Optional update checks read public GitHub release metadata. Presspeech downloads an installer only after approval, then verifies its published size and SHA-256 before offering to run it.

Privacy

Speech stays on the PC

No transcription service

Model files download from Hugging Face during setup, but dictation audio is not uploaded. There is no account, analytics SDK, crash reporter, or transcript history.

Diagnostics omit your words

Copy Diagnostics reports runtime state, timings, and character counts. It excludes transcript text, audio, and dictionary contents.

See the privacy page and machine-readable network-call inventory for the complete cross-platform network surface.

More detail

Support and source installation

Start with the Windows recovery guide. The Windows technical guide covers every setting, model choice, source setup, testing, packaging, and diagnostics. To report a problem, choose Copy Diagnostics and then Report a Problem… in the tray menu; paste the privacy-safe summary into the focused form. Also include the steps to reproduce, Windows version, CPU, GPU and driver when present, and whether the model reached Ready. For a recurring unmet need, choose Suggest an Improvement… and describe a measurable success condition.

GitHub issues are public. Do not add dictated text, audio, dictionary contents, credentials, or other private data. Use the private vulnerability-reporting form for a security issue.