How to use voice typing in VS Code: built-in speech and system-wide dictation

VS Code has official voice support through the Microsoft VS Code Speech extension. The documentation says recognition is computed locally and recordings are never sent to an online service. Once installed, the extension adds editor dictation, voice chat, walky-talky shortcuts, optional keyword activation, and text-to-speech. It supports 26 languages through the accessibility voice language setting.

Editor dictation is not limited to source files. Microsoft explicitly documents rich editor surfaces such as the source-control commit input and pull-request review comments. That makes it a useful low-risk tool for developer prose. Voice chat can open inline or in the Chat view and can automatically submit after a pause; adjust the speech timeout or disable automatic submission if you need a review step.

After installing the Speech extension, open the Command Palette and run “Voice: Start Dictation in Editor.” On macOS, the documented default shortcut is Option-Command-V. Escape stops recognition. Holding the start shortcut enables walky-talky behavior: recognition remains active while the keys are held and stops on release. If the shortcut conflicts with your setup, remap it through Keyboard Shortcuts.

For chat, run “Chat: Start Voice Chat” or use Command-I on macOS according to the current documentation. Confirm where focus is before speaking because VS Code can open chat in different locations. Set `accessibility.voice.speechTimeout` to control how long a pause waits before submission, or set it to zero when you always want to review a prompt before sending.

VS Code Speech is intentionally editor-focused. VoiceGem adds a consistent global shortcut across VS Code, Cursor, Windsurf, browsers, issue trackers, email, and chat applications. It can select local speech models, maintain vocabulary and replacements, trigger modes by active application, and translate spoken symbol or casing commands through Developer Mode.

That broader integration requires more trust. A system-wide app needs microphone access and may need Accessibility permission to insert text into other applications. Apple warns that Accessibility access is powerful and should be granted only to familiar software. Keep VS Code Speech as a lower-permission option when it solves the task, and use system-wide input where cross-app consistency genuinely reduces work.

You can dictate source-like text, but literal speech recognition is optimized for language, not a programming grammar. Common punctuation may arrive, yet identifiers, operators, indentation, generics, regular expressions, and shell syntax remain difficult. A developer formatter can interpret explicit commands such as casing and brackets, but ambiguity remains. The safest units are names, comments, tests, repetitive declarations, and compact expressions.

For larger changes, dictate a detailed prompt into chat. State the objective, relevant files, constraints, expected behavior, tests, and non-goals. Review the transcript before submission, then review the generated diff. Format, type-check, lint, and test. Voice can increase the amount of useful context you provide; deterministic tooling still decides whether the code is acceptable.

If no microphone appears, confirm the VS Code Speech extension is installed and the command exists, then review macOS Microphone permission for VS Code. If the wrong language is recognized, set `accessibility.voice.speechLanguage` and allow the matching language extension to install. If chat submits too quickly, increase or disable the speech timeout. If a global app records but does not paste, review its Accessibility permission and target focus.

If technical terms fail, add vocabulary in the system-wide tool or simplify repeated names through deterministic replacements. VS Code Speech's documented feature set does not promise a project dictionary. If recognition works but code is malformed, separate speech errors from formatting errors: first capture literal text, then apply a code-aware mode only where needed. This makes debugging the pipeline possible.