Dictate

Fully offline dictation for desktop VS Code and supported Windows text boxes on Windows x64. The extension includes multilingual Base speech recognition and its CPU runtime. No separately installed application, Python, online account, GPU or model download is required.

Install

This release supports Windows x64 only and requires VS Code 1.95 or later. The included model and runtime make the download approximately 240 MB. Speech recognition runs locally on the CPU.

Search for @id:ProjectAnomaly.dictate in the VS Code Extensions view. Select Dictate by Project Anomaly. For a local update package:

1. Copy the .vsix file to your PC. Do not unzip it.
2. In VS Code, press Ctrl+Shift+X, open …, and select Install from VSIX….
3. Select the file and reload VS Code if requested.
4. Open a local file or a new untitled text editor. Place the cursor where you want the text.
5. Press Ctrl+D, or click Dictate: Start in the status bar. Speak after the highlighted Dictate: Stop control and live wave appear.
6. Press the shortcut or status button again to stop and insert. Escape cancels. Ctrl+Z undoes the insertion.

The first recording loads the included model. Later recordings reuse it while VS Code remains open. Recordings stop automatically at 2 minutes. Changing the target editor/cursor, editing the document or leaving the VS Code window cancels the operation. Text is inserted at the original selection, and only if that target is still unchanged. A selected range is replaced. One cursor is supported.

Controls and settings

• Dictate: Choose Microphone selects your input. Windows must allow microphone access for desktop applications.
• Dictate: Dictation Settings opens language, Light/Verbatim cleanup, microphone and personal words.
• English, Spanish and Portuguese are available, including automatic selection between those three languages.
• Keyboard Shortcuts lets you change the binding or assign additional bindings to the same command. Existing desktop-app shortcut settings are not changed.
• Ctrl+D now starts/stops Dictate in supported VS Code contexts. In file editors, this replaces VS Code's normal Add Selection to Next Find Match action. You can assign that action another shortcut in Keyboard Shortcuts. Existing user-defined bindings can override the new default.
• Hover over the status button for a description of its current action.
• While recording, a floating translucent capsule shows animated orange, gold and teal curves driven by actual microphone loudness. Silence is flat. Loading/recognition use separate moving dots. The wave disappears on completion or cancellation and never takes keyboard focus. VS Code also retains its compact status-bar wave. Everything is local.

Dictating into chat

Click inside the chat message box, then press Ctrl+D to start. Speak, keep that same input and cursor unchanged, and press the shortcut again to insert. Review the text and send it yourself. Dictate never presses Enter or sends a message. Ctrl+Z uses the input's normal Undo.

The shortcut selects chat mode in VS Code chat inputs and when focus is outside the file editor (excluding terminals). Focus the intended chat box first: other accessible VS Code text fields can also receive text. You can also assign a simpler shortcut to Dictate: Chat Dictation: Start / Stop in Keyboard Shortcuts (Ctrl+K, Ctrl+S). A custom binding to the editor command does not automatically select chat mode. Use the shortcut to stop, so the chat input keeps focus.

Chat support uses Windows accessibility to verify the exact editable input, its content and selection. Standard text inputs, text areas and accessible rich text inputs are supported. Passwords, terminals and inputs whose content/selection cannot be safely inspected are rejected. Changing the input, cursor, content or VS Code window cancels insertion. If focus changes while text is being inserted, remaining characters are cancelled; some text may already be present. Line breaks and tabs become spaces so dictated text cannot submit a message. No clipboard contents are replaced.

Native text-area, single-line and rich-text webview fixtures were tested with the installed package, actual speech recognition, input events and Undo. Compatibility with each third-party chat extension and built-in chat remains unverified; inaccessible inputs show an error rather than receiving text.

Dictating in other Windows apps

Keep desktop VS Code running, then click the intended text box in another app. Press Ctrl+D, wait for Recording locally and the floating wave, speak, and press Ctrl+D again to insert. Escape cancels while recording or recognizing. Dictate never presses Enter, sends a message, executes a command or replaces clipboard contents. Line breaks and tabs become spaces. In standard Windows edit controls, place a cursor without selecting text; selected text is refused to protect it. Browser-style fields can replace a verified selection.

Outside VS Code, Windows reserves Ctrl+D for Dictate, replacing that app's usual action (such as adding a browser bookmark). Change Dictate: Desktop Shortcut in settings if preferred; editor/chat shortcuts inside VS Code are changed separately through Keyboard Shortcuts. If another program/window owns the shortcut, Dictate reports the conflict. Use one VS Code window for desktop dictation. Disable Dictate: Desktop Apps to release desktop shortcuts and turn off the floating wave. No startup task is installed.

The desktop component checks the original window/process, editable field, content and cursor before insertion, and checks for changes while recording. Changes cancel pending text. Password/read-only fields and terminals are rejected. If focus changes during character insertion, remaining text is cancelled; some text may already be present. Undo uses the receiving app's own behavior and cannot be guaranteed for every app.

Immediate checks cover actual Windows edit-control insertion/Undo, shortcut/cancellation replay and native wave rendering. Physical foreground shortcut routing and full end-to-end recording into other apps could not be validated because this Windows session denied access to the input desktop. Actual Word, Outlook, browser chat and company-PC compatibility remain unverified. These limits are not a background experiment or future review obligation.

Privacy and scope

The microphone is opened only during an explicit recording. Audio and recognition results are held in memory, and audio is discarded after processing or cancellation. Chat and desktop modes temporarily read the focused input's text and selection into local worker memory to verify the destination; these are discarded after insertion or cancellation and are not used for recognition. The extension saves no audio files, transcript history or document contents. Personal words/settings use VS Code's normal settings storage; VS Code Settings Sync, if separately enabled, may synchronize those settings.

The extension has no network requests, telemetry, accounts, downloads, web server, network ports or global keyboard hooks. Its bundled worker blocks Python sockets, DNS, external process launches and network file paths. Communication uses local process pipes. VS Code and other installed extensions have their own network behavior, which this extension does not control.

This version supports local files, untitled editors, accessible VS Code chat inputs, standard Windows edit text boxes and accessible Chromium-style inputs in other desktop apps. Full Word documents, terminals, notebook cells, remote files, elevated/security windows and inaccessible/custom controls are not supported. Support depends on the app exposing a verifiable editable field. No command is executed and no message is submitted. Other chat extensions may independently send your text when you submit it.

The included native speech worker and desktop shortcut/input/wave component run hidden and are managed by VS Code. Neither is separately installed. The desktop component starts with the extension to make its shortcut available, but does not open a microphone or load the speech model. Closing VS Code ends both components and releases their shortcuts. This package is locally distributed and unsigned; company extension policies or security software can still block it. Installation does not bypass company restrictions.

Licenses

Extension code is MIT licensed. Included speech model and third-party runtime notices are in THIRD-PARTY-NOTICES.txt and runtime/licenses/. This is a Windows x64 release; other platforms are rejected before the worker starts.

In Light mode, completed sentences inserted at the end of an editor, chat input or supported Windows text box leave one trailing space for the next recording. Verbatim mode, selected-text replacements, existing following text and technical literals retain their exact text. Each editor insertion keeps its own Undo step.