VS Code's built-in dictation, or one Whisper key across your apps
The quick answer: press ⌥⌘V in the editor to use VS Code's own dictation, which runs on an on-device model on Apple silicon, or install a system-wide app like MetaWhisp, where Right Command dictates into VS Code and into any other app that accepts paste. Which one fits depends on where else you write.
Key takeaways
- VS Code's built-in dictation is still experimental. On Apple silicon Macs it runs on an on-device model and works offline after a one-time download; press ⌥⌘V in the editor or ⌘I in chat.
- Its transcript cleanup is on by default and sends the text, not the audio, to a Copilot language model. You can turn that setting off.
- MetaWhisp puts Whisper large-v3 on Right Command in any app that accepts paste. Local mode is free and keeps your audio on the Mac.
What Are Your Options for Dictating in VS Code?
On a Mac you have three routes. The first is VS Code's own built-in dictation, still marked experimental, which works in the editor, chat, the Agents window and the terminal and runs on an on-device model on Apple silicon. The second is the VS Code Speech extension, which VS Code's docs point you to on platforms without built-in dictation, including Intel-based Macs.
The docs add that the extension also brings voice chat, spoken chat replies and Hey Code keyword activation. The third route is a system-wide dictation app such as MetaWhisp. It isn't tied to VS Code: tap Right Command, talk, tap again, and the text is pasted into whatever app holds your cursor, from the editor to the terminal, a browser or Slack.
That's the real trade-off. The first two only work inside VS Code; the third works in any app that accepts paste.
macOS's own Dictation works in VS Code too, since Apple says it enters text anywhere you can type. In my own test (2026) its word error rate was 11-14%, so for anything long, compare it with the other routes on your own audio.
How Does VS Code's Built-in Dictation Work?
VS Code's built-in dictation, still marked experimental, turns speech into text in chat, the Agents window, editors and terminals. On desktop it uses an on-device model by default: on first use, VS Code downloads a model called nemotron-3.5-asr-streaming-0.6b, and after that, its docs say, recognition works locally and offline. On a Mac, that on-device model needs Apple silicon.
In an editor, run Voice: Start Dictation in Editor or press ⌥⌘V. In chat, press ⌘I or click the microphone. For the terminal, run Voice: Start Dictation in Terminal from the Command Palette. Hold a dictation shortcut for push-to-talk. In chat, Escape cancels and discards the text. A session stops by itself at 20 minutes and keeps what it heard.
It's available when AI features are enabled, through the dictation.enabled setting, which is on by default. VS Code asks for microphone access when you start dictating.
These details come from VS Code's Voice support docs as of September 2026, and the feature is experimental, so expect changes. Like any VS Code command, you can rebind the shortcuts in Keyboard Shortcuts. The agents.voice.language setting is a language hint; leave it on auto to use your system language.
Does VS Code Dictation Send Your Voice Anywhere?
On desktop, no. VS Code's docs say the default model processes microphone audio on your device and needs no internet connection after the initial model download. VS Code for the Web is different: on-device transcription isn't available there, so dictation streams audio to the Microsoft AI voice service and needs a network connection and a GitHub sign-in.
Text is a separate question. The dictation.experimental.llmCleanup setting is on by default, and while it's on, VS Code sends the transcript text to a Copilot language model to improve punctuation, capitalization and formatting. Turn it off to keep the text on your Mac as well. The docs say keeping both audio and transcript on the device takes the on-device model plus cleanup turned off.
Organizations can enforce both choices with enterprise policies, and on the web, a policy that requires the on-device model makes dictation unavailable.
MetaWhisp's local mode keeps audio on your Mac every time, with no account and no telemetry. In local mode with the default Raw mode, neither the audio nor the text leaves your Mac.
Structured mode and translation send only the transcript text, never audio: to your own OpenAI or Cerebras key, or to Pro's built-in cloud AI, and the optional local model keeps it on your Mac. Cloud transcription sends audio: Pro switches to it, and on the free tier you can add your own Groq or OpenAI key for it.
When Is a System-Wide App the Better Choice?
Pick a system-wide app when much of your writing happens outside VS Code, or when you want one key for everything. MetaWhisp's Right Command works the same in VS Code, Cursor, the terminal, GitHub in your browser and your chat app, because it pastes into any app that accepts paste. There's nothing to enable per editor or per workspace.
It also helps if you work in more than one language. Whisper large-v3 covers 99 languages and MetaWhisp detects which one you're speaking, while Right Option dictates and translates into 11 target languages. Stay with VS Code's built-in dictation if you only dictate inside VS Code on Apple silicon and you're fine with its Copilot cleanup, or you turn it off.
In my own test (2026), MetaWhisp's word error rate was 3.7% on the same audio as the other apps below. VS Code's dictation isn't in these results.
| App | Word error rate (WER) |
|---|---|
| MetaWhisp | 3.7% |
| SuperWhisper | ~3.5% |
| Wispr Flow | ~3.5% |
| MacWhisper | ~3.5% |
| Apple Dictation | 11-14% |
| ~5-7% |
Three apps edged out MetaWhisp in that test, and you should know that. What MetaWhisp offers is a local mode that's free and unlimited, one key in any app that accepts paste, and automatic detection across the 99 languages on the Whisper large-v3 model card.
Setting Up MetaWhisp for VS Code
Setup is four steps, and you only do it once:
- Install MetaWhisp. Get it from the download page; it's free and needs no account. It runs on macOS 14 or later on Apple Silicon (M1 or later). Allow microphone and Accessibility access when macOS asks (Accessibility lets it paste).
- Let the model download. A small starter model installs first, then Whisper large-v3 downloads once, about 3.2 GB. It runs through WhisperKit on your Mac's GPU and CPU.
- Learn the two keys. Right Command dictates: tap to start and stop, or hold it in push-to-talk mode. Right Option dictates and translates.
- Try it in VS Code. Put your cursor in a file or the chat box, tap Right Command, say a sentence, and tap again. The text is pasted where your cursor is.
The same key then works in Cursor, the terminal, GitHub in your browser and Slack. One habit to know: MetaWhisp pastes through the clipboard and doesn't put your previous clipboard back, so copy anything important again after you dictate.
How Do You Dictate Comments, Commit Messages and Prompts?
Dictate the prose and type the syntax. Comments, docstrings, commit messages, pull request descriptions and chat prompts are plain sentences, so they suit dictation. For a comment, type // or /* yourself, then dictate the sentence. VS Code's editor dictation also works in the Source Control commit box and in pull request comment fields, per its docs.
For Copilot Chat, press ⌘I in the chat input or click the microphone, speak, and press it again to keep the text; VS Code fills the input but doesn't send it. Voice Mode, also experimental, goes further and holds a spoken conversation with the agent, but it needs an eligible individual Copilot plan. With MetaWhisp, click into any of these fields and tap Right Command.
The terminal is the exception on the VS Code side: its dictation adapts speech for command lines and converts spoken symbol names.
Want the text tidied before it lands? MetaWhisp's default Raw mode pastes the transcript as Whisper wrote it. Clean mode removes Russian filler words offline, and Structured mode uses AI, free with your own OpenAI or Cerebras key or the optional local model, or built in with Pro.
Pro tip: Read every dictated terminal command before you press Return. One misheard word in a shell command can delete the wrong file or push the wrong branch.
Fixing Dictation That Doesn't Work in VS Code
Start with the setting, then the model, then the key you pressed.
VS Code's built-in dictation
- No dictation commands? The docs say built-in dictation needs AI features enabled and the
dictation.enabledsetting on. On an Intel-based Mac it isn't available, so install the VS Code Speech extension. - Model won't download? If your network blocks it, the docs describe installing the official model package with Chat: Install Dictation Model from Local Package.
- Wrong microphone? Run Chat: Dictate: Select Microphone. The choice applies to every dictation surface.
- Your chat text disappeared? In chat, Escape cancels the session and removes its text, so stop with the shortcut or the microphone button instead.
MetaWhisp
- Got a translation instead of your words? You pressed Right Option, which dictates and translates. Use Right Command for plain dictation.
- Your clipboard changed? That's expected: MetaWhisp pastes through the clipboard and doesn't restore the old contents.
- Nothing on an Intel Mac? MetaWhisp needs Apple Silicon (M1 or later) and macOS 14 or later, so use Apple's Dictation or the VS Code Speech extension there.
VS Code Dictation vs MetaWhisp, Side by Side
The VS Code column comes from its Voice support docs as of September 2026.
| What you get | VS Code built-in dictation | MetaWhisp |
|---|---|---|
| Where it works | VS Code editors, chat, the Agents window, the terminal, commit and PR comment inputs | Any app that accepts paste |
| Speech model | nemotron-3.5-asr-streaming-0.6b by default on desktop | OpenAI Whisper large-v3 through WhisperKit, on the GPU and CPU |
| Macs supported | Apple silicon; Intel Macs use the VS Code Speech extension | macOS 14 or later on Apple Silicon (M1 or later) |
| Offline | Yes on desktop, after the model download | Yes in local mode |
| Transcript cleanup | A Copilot language model, on by default (sends text) | Raw by default; Clean offline; Structured with your own key, the local model or Pro |
| How you start | ⌥⌘V in the editor, ⌘I in chat; hold for push-to-talk | Right Command: tap to start and stop, or hold for push-to-talk |
| Languages | A language hint setting; auto uses your system language | 99, detected automatically |
| Translation | Not in the docs | Right Option: 11 target languages |
| Cost | Part of VS Code; needs AI features enabled | Free local mode; Pro is $30/year or $7.77/month |
FAQ
Does VS Code have built-in dictation on Mac?
Yes. VS Code has built-in dictation, still marked experimental, in the editor, chat, the Agents window and the terminal. On a Mac with Apple silicon it runs on an on-device model after a one-time download. Intel-based Macs need the VS Code Speech extension instead.
What is the VS Code dictation shortcut on Mac?
In an editor, press ⌥⌘V or run Voice: Start Dictation in Editor. In chat, press ⌘I or click the microphone button. Hold the shortcut for push-to-talk. In chat, Escape cancels and discards the text.
Does VS Code dictation work offline?
On desktop, yes. After the model downloads, VS Code's docs say speech recognition works locally and offline. VS Code for the Web is the exception: it streams audio to Microsoft's cloud transcription service. Transcript cleanup by a Copilot model is on by default and sends text, so turn it off if you want everything local.
Can I dictate in the VS Code terminal?
Yes. Run Voice: Start Dictation in Terminal from the Command Palette. VS Code adapts the text for command-line input and converts spoken symbol names. MetaWhisp also pastes into the terminal. Either way, read the command before you press Return.
Can I dictate Copilot Chat prompts?
Yes. Press ⌘I in the chat input or click the microphone, speak, and press it again to keep the text; dictation fills the input but doesn't send it. With MetaWhisp, click into the chat box, tap Right Command, talk, and tap again.
Does MetaWhisp work in Cursor and other editors?
Yes, in any app that accepts paste, including Cursor, Xcode and the terminal, because it pastes the text where your cursor is. The keys are the same everywhere: Right Command to dictate, Right Option to dictate and translate.
Is MetaWhisp free?
Yes. Local mode is free and unlimited, needs no account, and your audio never leaves the Mac. Pro costs $30 a year or $7.77 a month and adds cloud transcription (5,400 minutes a month), day and week summaries and search by meaning. Pro's cloud features send data to servers.
What is the best way to dictate in VS Code on Mac?
If you only dictate inside VS Code on Apple silicon, try the built-in dictation first: it's already part of VS Code and runs on-device. If you also write in other apps, want 99 languages detected automatically, or want translation on a key, a system-wide app like MetaWhisp fits better.
