Dictate

Desktop dictation that types into the focused app.

dictate runs as a small tray app. Press your configured push-to-talk shortcut, speak, and it transcribes into whatever app you are already using.

Local only. Transcription runs on your machine. No account, no subscription, no API key, no hosted model. Nothing you say leaves the device.

Current status: not a commercial product. Dictate was retired as one on 2026-08-25 and is now worked on when there is time. Linux installs, Windows 11 source installs, tray controls, startup integration, local dictation history, update, and uninstall paths are implemented.

The account, subscription, cloud sync and hosted-transcription code was removed on 2026-09-19 — 27,064 lines across 107 files. See docs/local-only-audit.md for what went and why, and VISION.md for the direction. The current transcription/model plan is docs/TRANSCRIPTION_PLAN.md.

Install

Windows 11 — Microsoft Store remains the intended public channel, with direct download from GitHub Releases available meanwhile.

The Store listing is overdue for an update and will get one once the first stable local-only build is done. Store packaging, submission automation and the listing copy all stay; only the in-app subscription part goes, since there is nothing to subscribe to. See docs/msstore-automation.md.

Windows direct download:

1. Open the latest GitHub release.
2. Download Dictate_*_x64_en-US.msi when a staging MSI is attached.
3. Run the MSI and open Dictate from the Start Menu. Windows may warn that the installer is from an untrusted publisher; that is expected for staging builds.

If a release does not yet have an MSI attached, use the developer/source bootstrap below or wait for the Microsoft Store package.

Windows developer/source install from the hosted bootstrap:

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

Unstable developer/source bootstrap for pre-stable feature testing:

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

Open Dictate from the Start Menu after install.

Linux (Ubuntu) — install the release .deb. It is built once in CI with the engine, model and desktop shell inside, so every install of a version is the same:

1. Download Dictate_<version>_amd64.deb from the latest GitHub release.

2. Install it with apt, giving the path to the file:

   sudo apt install ~/Downloads/Dictate_<version>_amd64.deb
   

3. Open Dictate from the app launcher. It starts on sign-in.

Dictate does not publish an apt repository, so apt upgrade does not see new versions. Update from the app, or install the next release's .deb the same way.

Linux latest code (per-user, builds on your machine, no sudo). This installs from source into ~/.local/share/dictate and compiles the desktop shell locally, so it needs cargo, npm and webkit2gtk-4.1-dev:

npx @arcforgelabs/dictate install

Or from a checkout: ./install-ubuntu.sh, or ./install.sh on other distributions.

Windows developer/source install, from the repo/source directory:

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

The local .ps1 installer scripts must be run from a checkout or extracted source directory. They will not work from C:\Windows\System32.

Node/npm users can also run the developer bootstrap:

npx @arcforgelabs/dictate install

Workflow

Open Dictate
Use the mic button or configured push-to-talk shortcut
Hold shortcut, speak, release
Review Dictations when needed

By default, Dictate installs a normal app launcher entry and starts on sign-in. Advanced configuration is available through the dictate config CLI.

Update And Uninstall

Windows developer/bootstrap install:

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

To test updates before they are promoted to stable:

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

Users on the app's npm-backed update path can opt into or out of unstable updates from the CLI:

dictate config set-update-channel unstable
dictate config set-update-channel stable

Windows from source:

powershell -ExecutionPolicy Bypass -File .\update-windows.ps1
powershell -ExecutionPolicy Bypass -File .\uninstall-windows.ps1

Linux .deb install: update from the app, or install the newer .deb with sudo apt install. Remove it with:

sudo apt purge dictate

Linux latest-code (per-user) install:

npx @arcforgelabs/dictate update --user
npx @arcforgelabs/dictate uninstall

From a checkout, ./update.sh and ./uninstall.sh do the same.

Use -RemoveUserData on Windows or --remove-user-data on Linux only when you also want to remove config, logs, history, and downloaded model data.

What It Does Today

• Starts from the Windows Start Menu or Linux app launcher
• Runs as a tray app
• Types dictated text into the focused app. A terminal, including the Grok CLI, receives that text with Shift+Insert so it is not filed as an attachment.
• Leaves your clipboard as it was: the text is pasted through the clipboard, then what you had copied is put back. On Windows the dictated text carries the markers Windows defines to leave content out of clipboard history (Win+V) and cloud clipboard. On Linux the most useful single format comes back (an image, else text).
• Supports configurable push-to-talk
• Presents a simple capture-first desktop UI
• Supports launch on startup
• Keeps a small local dictations history for copy/paste recovery
• Provides installer, updater, uninstaller, and doctor paths

Models

Dictation runs NVIDIA Parakeet TDT 0.6B through ONNX Runtime: v2 for English (the default, bundled with the desktop app) or v3 for 25 European languages. Every supported model runs on your machine. The Whisper-family backends (faster-whisper, WhisperX, whisper.cpp) were removed on 2026-09-30; a saved config that still names one falls back to Parakeet. The desktop UI keeps engine names out of the primary workflow; advanced users and tests can still select an explicit local engine through CLI options and dictate config.

Dictate runs on the CPU. GPU support (NVIDIA CUDA, AMD) was dropped on 2026-10-01 to focus on CPU latency. The installers no longer install a GPU runtime; their old GPU flags are accepted and ignored so existing update commands keep working. The engine has no device choice: a saved GPU device in config.yaml moves to CPU on upgrade, and --device is still accepted so the installers and old scripts run, and is ignored.

Meetings (speaker-labelled transcripts) are not part of Dictate for now: they were removed from every build on 2026-10-03 and are shelved as P3 (#140). On upgrade a saved Meeting model is removed from config.yaml with a one-line notice, and meeting transcripts you already saved stay in your notes with their speaker labels. The installers' old --meeting / -Meeting flags are accepted and ignored.

Commands

dictate
dictate --no-tray
dictate --once
dictate --once --copy
dictate doctor --quick
dictate doctor --quick --fix
dictate doctor --check-model-load

Advanced configuration remains available for automation and testing:

dictate --stt-backend parakeet --model parakeet-tdt-0.6b-v3
dictate config show
dictate --add-hotword AcmeWidget
dictate --list-hotwords

State

User state is local:

Linux:
  ~/.config/dictate/config.yaml
  ~/.local/share/dictate/

Windows:
  %APPDATA%\dictate\config.yaml
  %LOCALAPPDATA%\dictate\

Repo defaults intentionally ship with hotwords: []. Hotwords are user-specific and should not be packaged into the public repo default config.

Safety

• Audio and transcripts stay on the device. There is no hosted transcription path and no account.
• Dictation text can be sensitive; check logs and issue reports before sharing.
• Important transcriptions should be verified before relying on them.
• Support and maintenance are best-effort.

Desktop UI — the Quiet Console (preview)

A quiet desktop window sits alongside the tray — a capture home (the mic is the record button), the notes list, and the ⌘K palette; no settings menu (config via the dictate config CLI).

• ui/ — React/Vite front-end. For UI work, cd ui && npm run dev opens the whole app at http://localhost:5173 against a built-in mock (no engine, no Tauri, no STT) — the fast UI loop in a browser. Details, and when to use the Tauri shell instead, are in that README's Develop section.
• ui-shell/ — Tauri 2 shell that hosts it on Linux.
• src/dictate/ui_server.py — the loopback control server the UI talks to; the tray can launch the shell for the desktop capture surface.
• See design/PLAN.md for the cross-platform plan.

Install it like a normal app: on Linux that is the release .deb. It bundles the frozen Python engine (PyInstaller sidecar) and the prebuilt desktop shell, so there is no Python, pip or compiler step:

sudo apt install ./Dictate_*_amd64.deb

The app lives in the tray and desktop shell and does push-to-talk straight away. Build the package yourself in one step with scripts/build-linux-desktop.sh — see ui-shell/README.md. The per-user npm/install.sh route installs under ~/.local/share/dictate and builds on your machine; use it to try the latest code, not as the everyday install.

Docs

docs/ is published as the docs site at https://arcforgelabs.github.io/docs/dictate/ by arcforgelabs/docs. docs/docs.json is its navigation, and docs/releases/ holds one page per release.

• Windows 11 support
• Transcription deployment plan
• Release policy: channels, versions, validation
• Release/versioning (current procedure)
• Desktop packaging & CI runbook
• Microsoft Store automation
• Microsoft Store listing draft
• Deployment security
• Development streams
• Archived docs index
• Security policy

License

MIT. See LICENSE.