Windows 11 Support

Windows 11 is a supported desktop target for Dictate. The normal path is the installed Dictate app entry, which starts the desktop capture app and tray process. The primary workflow is simple dictation and local dictation recovery; advanced configuration (model, shortcut, hotwords, update channel) lives in the dictate config CLI.

Supported Surface #

• Python 3.11 or 3.12.
• One-shot dictation: dictate --once.
• Clipboard output: dictate --once --copy through pyperclip.
• Push-to-talk tray app: dictate --type-backend pynput.
• Headless push-to-talk daemon: dictate --no-tray --type-backend pynput.
• Speech-to-text: Parakeet through ONNX Runtime on the CPU. GPU acceleration is not supported.

The Windows tray uses the native notification area. The Linux GTK/Ayatana tray remains a separate implementation.

Install #

The signed, stable Windows channel is Microsoft Store distribution. A stable Windows release means a Store package that we are happy to submit and support. Direct stable and unstable installers are also supported and update independently from GitHub release assets. Paid Authenticode signing for those assets is a future option only; it is not a current release blocker or standing task. Early testers can use staging MSI builds and accept the expected Windows untrusted-publisher warnings.

Direct MSI download:

1. Open the latest GitHub release.
2. Download Dictate_*_x64_en-US.msi when a staging MSI is attached.
3. Run the installer and open Dictate from the Start Menu.

Developer/source bootstrap from PowerShell:

powershell -ExecutionPolicy Bypass -Command "iwr -useb https://cdn.jsdelivr.net/npm/@arcforgelabs/dictate@latest/install.ps1 | iex"

Pre-stable developer/source bootstrap from the npm unstable channel:

powershell -ExecutionPolicy Bypass -Command "iwr -useb https://cdn.jsdelivr.net/npm/@arcforgelabs/dictate@unstable/install.ps1 | iex"

Direct installs can choose Normal or Beta updates inside Dictate. Microsoft Store
installs show `Microsoft Store · Stable`; checking for updates opens the Store's
Downloads and updates surface and never switches to the direct-install ecosystem.

Transcription is local-only: there is no cloud lane, account, or API key on
Windows or anywhere else. See `VISION.md`.

The npm package is an installer shim that publishes the PowerShell lifecycle scripts. The hosted bootstrap downloads the matching tagged Dictate source release and runs the platform installer. It is a developer/bootstrap path, not the public Windows install target. If Node.js is already installed, this is equivalent:

npx @arcforgelabs/dictate install

Use npx @arcforgelabs/dictate@unstable install or the @unstable CDN URL only for pre-stable feature testing. Stable public promotion still goes through CalVer release tags and the guarded Store/MSIX path.

From PowerShell in a repo root:

powershell -ExecutionPolicy Bypass -File .\install-windows.ps1

Windows setup wizard from a repo root:

powershell -ExecutionPolicy Bypass -File .\install-windows-wizard.ps1

That command creates .venv with Python 3.11 or 3.12, installs Dictate with the Windows dependencies, installs the Microsoft Visual C++ runtime if it is missing, seeds %APPDATA%\dictate\config.yaml, writes launcher scripts, prepares the default Parakeet local model, runs dictate doctor --quick, registers Dictate in Installed Apps, adds a Start Menu shortcut named Dictate, and enables launch on startup by default.

Skip model preparation or verification when needed:

.\install-windows.ps1 -NoPrepareTurbo -NoVerify

CI/smoke-test install without a Start Menu shortcut:

.\install-windows.ps1 -NoPrepareTurbo -NoVerify -NoShortcut

Host-driven QEMU/KVM VM smoke test from Linux:

scripts/windows-vm-smoke.sh --vm <your-windows-vm> --mode syntax
scripts/windows-vm-smoke.sh --vm <your-windows-vm> --mode install
scripts/windows-vm-smoke.sh --vm <your-windows-vm> --mode lifecycle

The VM smoke script uses libvirt virsh qemu-agent-command, so the Windows guest must have QEMU Guest Agent installed and running. It copies the source zip through QEMU Guest Agent file APIs, so guest-to-host networking is not required. syntax only parses the PowerShell scripts in Windows PowerShell. install resets Dictate app data in the guest, runs a no-shortcut install, compiles Python sources, runs focused tests, checks the version, verifies the fresh Parakeet default with dictate doctor --quick, and uninstalls. lifecycle adds update and post-update doctor/uninstall smoke checks. The maintained VM is win11-gpu; pass LIBVIRT_DEFAULT_URI=qemu:///system.

Release gate #

Before a Microsoft Store draft, test the published installer on the VM:

scripts/windows-vm-release-gate.sh --tag v2026.9.25-4

It downloads Dictate_<version>_x64-setup.exe from the GitHub release, copies it over SSH (the installer is too large for guest-agent file copies), removes any existing Dictate, installs silently, checks the Installed Apps version, launches the app in the signed-in console session through a one-shot scheduled task, waits for the engine's /api/health to report the same version, then uninstalls and checks the entry and executable are gone. The guest needs OpenSSH Server and a user signed in at the console. It uses -o IdentityAgent=none because the local SSH agent can stall on an approval prompt; override with DICTATE_WINDOWS_SSH_OPTS.

GitHub-hosted Windows user smoke test:

.\scripts\windows-user-smoke.ps1

This CI gate installs Dictate through the hosted developer bootstrap path from a deterministic source zip, verifies the Start Menu shortcut, default startup shortcut, Installed Apps registry entry, config seeding, dictate --version, dictate doctor --quick, dictate doctor --fix, hosted update, and uninstall cleanup. It does not test a real microphone, visible tray interaction, or Windows Search indexing.

Update or uninstall from a repo root:

.\update-windows.ps1
.\uninstall-windows.ps1

Use -RemoveUserData with the uninstaller only when config, logs, history, and downloaded model data should also be removed.

The setup wizard links to Dictate documentation, the dedicated Dictate privacy policy at https://arcforge.au/privacy/dictate, and Arc Forge terms at https://arcforge.au/terms. Dictate is built to be useful, but support and maintenance are best-effort. Verify important transcriptions, report issues, and consider paid support if Dictate saves you time and you have the means.

Manual install:

py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade pip
.\.venv\Scripts\python.exe -m pip install -e ".[windows]"
.\.venv\Scripts\dictate.exe doctor --quick --type-backend pynput
.\.venv\Scripts\dictate.exe doctor --quick --fix --type-backend pynput
.\.venv\Scripts\dictate.exe doctor --quick --update-paths --type-backend pynput
.\.venv\Scripts\dictate.exe doctor --quick --stt-backend parakeet --type-backend pynput

Run #

Tray app:

.\.venv\Scripts\dictate-tray.cmd

Headless push-to-talk:

.\.venv\Scripts\dictate-daemon.cmd

One-shot:

.\.venv\Scripts\dictate-once.cmd
.\.venv\Scripts\dictate-once.cmd --copy

Windows Paths #

• Config: %APPDATA%\dictate\config.yaml
• History: %LOCALAPPDATA%\dictate\recent-history.json
• Logs: %LOCALAPPDATA%\dictate\logs\
• Fallback logs: %TEMP%\dictate-logs\
• Launchers: .venv\Scripts\dictate-tray.cmd, .venv\Scripts\dictate-daemon.cmd, .venv\Scripts\dictate-once.cmd, and .venv\Scripts\dictate-controls.cmd
• Start Menu shortcut: %APPDATA%\Microsoft\Windows\Start Menu\Programs\Dictate.lnk
• Startup shortcut: %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\Dictate.lnk

Dictation timing in the log #

Each finished dictation adds one line to logs\latest.log, after Typed: [N characters, not logged]. It holds durations only, never the words:

Dictation timing: audio=5.60s release_to_text=931ms stop=45ms wait=12ms decode=820ms fixes=0ms history=3ms paste=51ms | clipboard lock=0ms save=4ms write=1ms keys=25ms read=120ms restore=15ms | capture_processing=30ms

• audio: length of the recording. Transcribing 5.6 s of audio... is this same length, not a wait.
• release_to_text: key release until the paste keystroke is sent; the wait a user sees. It is made of stop (microphone stopped), wait (queue to the decoder), decode (Parakeet), fixes (hotword and replacement fixes), history (saving to recent history) and paste.
• clipboard: the parts of paste (lock waits for the previous dictation's restore, save, write, keys), then the time until the target app read the text (read) and the restore of what was on the clipboard (restore). These last two happen after the text has appeared.
• capture_processing: resampling and gain control done while recording. It is not part of release_to_text.

Desktop App #

Open Dictate from the Start Menu. The desktop app includes:

• a capture home with a mic control
• a local dictations view for copy/paste recovery
• a command palette for daily actions
• a privacy pill stating that transcription stays on this device
• quiet update status when an app update is available

Advanced configuration remains available from PowerShell:

.\.venv\Scripts\dictate.exe config show
.\.venv\Scripts\dictate.exe config set-model parakeet-tdt-0.6b-v3
.\.venv\Scripts\dictate.exe config set-shortcut ctrl_r

Known Gaps #

• Microsoft Store Submission 2 is staged with refreshed listing assets and is held until the next version package is bundled and tested.
• Global hotkey reliability depends on pynput permissions and the active desktop/session.
• NeMo Canary on Windows is not part of the supported baseline.