Claude Code

Speech to Text in Claude Code: Talk Through Your Prompts

Claude Code does its best work when you give it the whole picture, and most of us say far more than we bother to type. Here's how Claude Code's own /voice dictation works, where it stops, and how LumeVoice lets you talk to Claude Code in any terminal.

  • Hold fn or Ctrl + Win, talk, let go
  • Types into any terminal, editor or chat panel
  • Drops the ums, keeps your corrections

Also onMacWindowsiPhoneAndroid

Speech to text in Claude Code works, and you have two good ways to get it. Claude Code has its own dictation, /voice, for people signed in with a Claude.ai account on their own machine. LumeVoice works from the outside: hold one key, talk, and a cleaned-up prompt appears in whatever terminal Claude Code is running in.

Short answer

Does Claude Code have voice input?

Yes. As of October 2026, Claude Code has built-in voice dictation. Run /voice, then hold Space while you talk, or switch to tap mode with /voice tap. It needs a Claude.ai sign-in and a microphone on the machine running Claude Code, so it isn't available with an API key, Amazon Bedrock, Google Cloud or Microsoft Foundry, or over SSH.

How to use Claude Code's built-in /voice#

Anthropic built dictation into the Claude Code CLI, and it's well done. As of October 2026, Claude Code's voice dictation docs describe it like this:

  1. Sign in with a Claude.ai account. Run /login if you need to.
  2. Run /voice. Claude Code checks your microphone. On a Mac, your terminal app asks for microphone access the first time.
  3. Hold Space and talk. After a short warm-up, your words stream into the prompt. Let go, read what you said, and press Enter.
  4. Or switch to tap mode. With /voice tap, one tap of Space starts recording and a second tap sends the prompt, as long as it's at least three words.

Some details are genuinely thoughtful. Transcription is tuned for terms like OAuth, JSON and localhost, your project and git branch names are passed in as recognition hints, and dictation doesn't count toward your Claude usage limits.

Where /voice stops#

  • Your sign-in. It isn't offered when Claude Code runs on an API key, Amazon Bedrock, Google Cloud or Microsoft Foundry.
  • Remote work. The microphone has to be on the machine running Claude Code, so it doesn't work over SSH or in cloud sessions. The VS Code extension supports dictation too, but not in Remote sessions such as SSH, Dev Containers and Codespaces.
  • Languages. It covers 20 dictation languages and falls back to English for the rest.
  • Length. In tap mode, recording stops after 15 seconds of silence or two minutes in total.
  • Policy. An organization admin can turn it off.
  • Everything else. It only types into Claude Code. The commit message, the PR description and the Slack thread about the change are still on your keyboard.
  • Cleanup. The docs describe a live transcript of what you said. They don't mention removing the "um"s or acting on "no, wait" corrections.

How to dictate to Claude Code with LumeVoice#

LumeVoice doesn't connect to Claude Code at all. It types into whatever has your cursor, so it behaves the same in Terminal, iTerm2, Ghostty, Warp, Windows Terminal, or the terminal inside VS Code, Cursor or a JetBrains IDE.

  1. Download LumeVoice for Mac or Windows and sign in. On Windows it installs per user, with no admin rights needed.
  2. Click into the Claude Code prompt.
  3. Hold fn on a Mac or Ctrl + Win on Windows, say what you want, and let go. The cleaned-up text is pasted at your cursor.
  4. Read it and press Enter when you're happy. Lume pastes the text and stops there, so it never sends a prompt for you.
  5. For a long brief, double-press the key to go hands-free, talk as long as you need, and press it once to finish.

Because Lume pastes into the terminal window on your own computer, it doesn't care how Claude Code is signed in or where it runs. An API key, Bedrock, a tmux pane, a server you've SSH'd into: each one is just a prompt waiting for text.

Steering a cloud session from your phone? Switch to the Lume keyboard on iPhone, or tap the floating Lume bubble on Android, and talk.

Say the whole brief, then press Enter

LumeVoice turns a rambling explanation into a clean prompt for Claude Code, and into clean text everywhere else on your Mac or PC.

  • Filler words out, self-corrections applied, spoken lists formatted
  • A custom dictionary for your libraries, services and teammates
  • 3,000 free words, no card required
Download for MacMac (macOS 11.5+), Windows 10 & 11, iPhone and Android

Tips for spoken prompts Claude Code can act on#

Talking makes it easy to give Claude Code more context than you'd ever type. These habits make that context count.

Say the goal, the limits and the proof. What should change, what must not, and how Claude should check its work. An illustrative example:

You say: "Um, so in the billing service, add retries to the webhook handler, three tries with backoff, actually make it five, and don't touch the public API. Then run the billing tests."

Lume types: "In the billing service, add retries to the webhook handler, five tries with backoff, and don't touch the public API. Then run the billing tests."

Type the @, say the rest. File paths are awkward to say and quick to type. Type @, pick the file from Claude Code's autocomplete, then hold your key and explain what to do with it.

Plan big changes first. Press Shift+Tab to switch to plan mode before a long spoken brief. Claude proposes a plan you can correct before it edits anything.

Start long dictations with a typed line. Claude Code folds any paste over 800 characters or three lines into a placeholder such as [Pasted text #1 +6 lines], and still sends the full text. Its docs say Claude treats pasted text as material from somewhere else, and follows instructions inside it where your typed message asks it to. So type a few words first, such as "Do the following:", then dictate. For a full spec, dictate it into a Markdown file and ask Claude to read that file.

Teach Lume your words. Add library names, service names and teammates to the custom dictionary once. They're passed to speech recognition, so your stack comes out spelled the way you write it.

Write CLAUDE.md out loud. Explaining your conventions to Claude is like onboarding a new teammate, and that's easier said than typed. Dictate a first draft, then trim it.

Keep syntax on the keyboard. Shell commands with flags, regexes and exact identifiers are faster to type. Describe the behavior and let Claude write the code.

Claude Code /voice vs LumeVoice#

Claude Code /voiceLumeVoice
PriceIncluded with your Claude plan3,000 free words, then $14.99/mo, $79.99/yr, or $249 lifetime
Works withClaude.ai sign-in onlyAny sign-in: Claude.ai, API key, Bedrock, Google Cloud, Foundry
Over SSHNoYes, it types into your terminal window
Where it typesClaude Code CLI and VS Code extensionAny app on Mac and Windows, plus iPhone and Android
How you startHold Space, or tap modeHold fn or Ctrl + Win; double-press for hands-free
CleanupLive transcript tuned for coding termsRemoves fillers, applies corrections, formats lists, fixes grammar
Your vocabularyProject and branch names, automaticallyA custom dictionary you edit
Languages20100, detected automatically
Long promptsInserted as typed textPasted, so long ones show as a placeholder
AudioStreamed to Anthropic for transcriptionSent to the cloud, not stored after processing
LinuxYesNo Linux app

Swipe to see all columns

When Claude Code's /voice is all you need#

Try /voice first if you sign in with a Claude.ai account, work on your own machine, speak one of its 20 languages and mostly want to talk to Claude Code. It's included in your plan, needs no setup and knows your branch name. It's also your answer on Linux, where LumeVoice has no app.

LumeVoice earns its place when you want one voice tool for the whole job: Claude Code over SSH or through Bedrock, then the prompt in Cursor, the Linear issue, the PR description and the Slack update, all with the same key and the same dictionary.

Talk through the task, type the details#

Voice coding with Claude Code isn't about reading symbols aloud. It pays off on the parts that are plain language: the goal, the context, the constraints, the review notes. Speak those, type the syntax, and read before you press Enter. For more developer workflows, see LumeVoice for developers and our guide to voice dictation in VS Code, Xcode and IntelliJ.

Frequently Asked Questions

Can I dictate code to Claude Code?

You can, but it's rarely worth it. LumeVoice writes prose: it won't turn speech into brackets, operators or camelCase names. The habit that works is saying what the code should do, what must not change and how to test it, then letting Claude Code write the code. Type exact identifiers, or add the ones you use often to Lume's custom dictionary so they come out spelled right.

Does dictating with LumeVoice use my Claude tokens or plan limits?

No. LumeVoice turns your speech into text in the cloud, outside Claude, and types it into your terminal, so transcription never touches your Claude plan. Only the prompt you send counts, exactly as if you had typed it. LumeVoice is priced separately: every account starts with 3,000 free words, no card required, and plans after that cost $14.99/mo, $79.99/yr, or $249 lifetime.

Can I keep Claude Code's /voice on and use LumeVoice too?

Yes. They listen for different keys. Claude Code's dictation uses Space inside the Claude Code prompt, while LumeVoice listens for fn on a Mac or Ctrl + Win on Windows, anywhere on your computer. A sensible split is /voice for quick follow-ups in a local session and LumeVoice for long briefs, remote sessions and everything outside the terminal.

Can I turn a recorded walkthrough into a prompt?

Yes. The LumeVoice desktop app also transcribes audio and video files, including mp3, m4a, wav, mp4 and mov. Open one from the menu bar on a Mac or the system tray on Windows, or use Open With in Finder or the right-click menu in Explorer. Paste the transcript into a Markdown file in your repo, then ask Claude Code to read it.

Is it safe to dictate about private code?

LumeVoice sends your audio to the cloud for transcription and cleanup. The audio isn't stored after processing, nothing you say is used to train models, and your dictation history stays on your device. It needs an internet connection. If your employer limits which third-party services may see work content, check that policy first, the same as you would for Claude Code itself.

Updated