Claude Code /voice vs. local dictation: which should you use?

There are two different ways to talk to Claude Code today. The first is Anthropic's own /voice command, built directly into the CLI: hold a key, speak, release, and your words become the prompt. The second is a separate, local voice-to-text tool that types into Claude Code's terminal the same way it would type into any other window — a Slack message, a commit, a pull request description.

They look similar from the outside — you talk, text appears — but they solve different problems, and the difference that matters most is where your audio actually goes. Anthropic's own documentation is explicit about it: /voice "streams your recorded audio to Anthropic's servers for transcription. Audio is not processed locally." A local dictation tool works the opposite way — transcription happens on your machine, and nothing about your voice leaves it.

What does Claude Code's /voice command actually do?

Run /voice in a Claude Code session and it toggles voice dictation on. By default it runs in hold mode: hold Space, speak, release, and the transcript is inserted at your cursor and waits for Enter. A newer tap mode (/voice tap) lets you tap once to start recording and tap again to transcribe and send automatically. You can mix voice and typing in the same message — dictate the explanation, then type an exact file path or variable name.

Per the official Claude Code documentation, the feature is genuinely well built for coding specifically: common terms like regex, OAuth, JSON, and localhost are recognized correctly, and your current project name and git branch are added automatically as recognition hints. It supports 20 dictation languages (English, French, German, Spanish, Japanese, Korean, and 14 others), defaulting to English if your configured language isn't on that list. Transcription doesn't consume Claude messages or tokens and doesn't count against your /usage limits.

It also has real boundaries. It only runs inside the Claude Code CLI and its VS Code extension — not system-wide. It requires signing in with a Claude.ai account, so it's unavailable if your organization runs Claude Code through an Anthropic API key directly, Amazon Bedrock, Google Vertex AI, or Microsoft Foundry. And it needs a local microphone, so it doesn't work in Claude Code on the web, over SSH, or in VS Code Remote/Dev Containers/Codespaces sessions. If you're new to shaping prompts by voice rather than typing them, our guide to voice prompting and AI context covers the habits that carry over to any tool, /voice included.

Is Claude Code's /voice processed locally, or in the cloud?

In the cloud, by design. Anthropic states this plainly in the Requirements section of the voice dictation docs:

"Voice dictation streams your recorded audio to Anthropic's servers for transcription. Audio is not processed locally."

Two other requirements in the same section follow directly from that architecture. First, /voice needs a Claude.ai account — there's a speech-to-text service behind the scenes, and it's only reachable through that authentication path. Second, and more telling: Anthropic disables /voice entirely for organizations that have HIPAA compliance enabled, showing "Voice mode is disabled by your organization's policy" instead. That's Anthropic itself drawing the line — if your compliance posture treats audio as sensitive, the built-in cloud transcription path is switched off, full stop.

None of this makes /voice unsafe for typical use — it's a documented, first-party feature with no hidden behavior. But "not processed locally" is a factual claim worth taking at face value if you're dictating anything you'd rather not stream to a third-party server as audio: proprietary code, unreleased features, customer data, or anything else that shows up in your prompts.

What does local, system-wide dictation add that /voice can't?

A local dictation tool runs speech-to-text on your own hardware and types the result wherever your cursor is — no cloud round-trip, no destination-specific integration. That difference shows up in a few concrete ways:

• It works everywhere, not just Claude Code. The same tool that dictates your Claude Code prompt also dictates a git commit message, a GitHub PR description, a Slack reply, or a paragraph in your notes app — anywhere you can place a text cursor.
• Your audio never leaves your machine. There's no server in the loop to stream to, which matters directly for the HIPAA-restricted orgs and privacy-sensitive teams /voice explicitly locks out — see Inkvox's privacy approach for what "0 bytes uploaded" means in practice.
• It works where /voice explicitly doesn't. Anthropic's docs list Claude Code on the web, SSH sessions, and VS Code Remote/Dev Containers/Codespaces as environments where /voice "does not work" because the microphone lives on a different machine than the model. A local tool running on your own desktop has no such restriction for local sessions.
• No account gate. /voice is unavailable if you're on an API key, Bedrock, Vertex AI, or Foundry setup instead of a Claude.ai login. Local dictation doesn't care which backend Claude Code is pointed at — it just types.

Independent write-ups on /voice converge on the same limits. BuildMVPFast's comparison puts it plainly: "Claude Code's advantage is terminal-native integration with zero setup… the disadvantage is it only works inside Claude Code," noting that cross-app tools cover more ground if you switch between apps throughout the day. MindStudio's review reaches the same conclusion from the accuracy side, reporting that transcription is "less reliable for code snippets and exact syntax" and that many developers end up pairing /voice for quick in-session prompts with a dedicated dictation tool for anything longer or more technical.

If your day is split between Claude Code and other editors — say, VS Code with GitHub Copilot — the same local-vs-cloud tradeoff applies there too; see our notes on dictating in VS Code with Copilot.

How do you choose between /voice and local dictation?

They're not mutually exclusive, but one will usually fit your workflow better:

• Reach for /voice if you're logged in with a Claude.ai account, working on your local machine (not SSH or the web), only ever dictate inside Claude Code's terminal or its VS Code extension, and you're fine with your audio being processed on Anthropic's servers for a feature that costs nothing extra.
• Reach for local dictation if any of these apply: your org has HIPAA compliance enabled (/voice is disabled by policy); Claude Code is configured with an API key, Bedrock, Vertex AI, or Foundry (/voice is unavailable at all); you work over SSH, on a remote box, or in Claude Code on the web; you want dictation across every app you use, not just Claude Code; or you'd simply rather no audio — even a short one — leave your device.

Plenty of developers end up using both: /voice for a quick in-terminal prompt when it's convenient, and a local tool for everything else that isn't Claude Code — the commit message, the PR write-up, the Slack thread explaining what just shipped.

How do you set up fully local dictation for Claude Code prompts?

Two honest starting points, depending on how much setup you want to do:

• OS-native dictation — Windows Voice Access or macOS Dictation both run on-device once configured, and they're free. They're general-purpose accessibility tools first, though: not tuned for coding vocabulary the way /voice is, and not built as a dedicated long-form transcription tool. (We cover Windows' options in more depth in why Windows Voice Typing needs the internet, since not every built-in Windows dictation feature is actually offline.)
• A packaged local-Whisper app — a small category of tools run OpenAI's Whisper model on your own GPU and handle the hotkey, model management, and typing for you. That's the category Inkvox is built for.

Inkvox runs Whisper large-v3-turbo, quantized, directly on your GPU through Vulkan — NVIDIA, AMD, or Intel — with a CPU fallback if you don't have a compatible card. On a mid-range GPU like an RTX 3070, a sentence transcribes in roughly 0.3–0.4 seconds. Audio never leaves your machine: 0 bytes uploaded, no account required, no telemetry on what you say, and support for 100+ languages — versus the 20 languages /voice currently lists. The only network use is a one-time ~800 MB model download on install; after that, it works completely offline, in Claude Code's terminal and every other app you type into.

Inkvox is in free private beta for Windows 10/11. Join the waitlist to get access.

Frequently asked questions

Is Claude Code's /voice processed locally?
No. Anthropic's documentation states it directly: "Voice dictation streams your recorded audio to Anthropic's servers for transcription. Audio is not processed locally." Your microphone audio leaves your machine and is transcribed on Anthropic's infrastructure before the text lands in your prompt.

Does /voice work outside Claude Code, like in Slack or my browser?
No. It's built into the Claude Code CLI and its VS Code extension only, not as a system-wide dictation layer. It also doesn't work in Claude Code on the web, over SSH, or in VS Code Remote/Dev Containers/Codespaces, since it needs direct access to a local microphone.

Why is /voice disabled for my organization?
Per the docs, /voice requires a Claude.ai account and is unavailable when Claude Code runs on an API key, Bedrock, Vertex AI, or Foundry. It's also disabled by policy for organizations with HIPAA compliance enabled, shown as "Voice mode is disabled by your organization's policy."

Does /voice cost extra or use my usage limits?
No. Anthropic's documentation states transcription doesn't consume Claude messages or tokens and doesn't count toward the limits shown in /usage.

How many languages does Claude Code's /voice support?
20, per the official docs — including English, French, German, Spanish, Japanese, and Korean — falling back to English otherwise. Coding-term accuracy comes from context hints like your project name and git branch, not from broader multilingual tuning.

Whichever you pick, both beat typing out a 200-word prompt by hand. If you want the version where your voice never leaves your machine — in Claude Code or anywhere else — join the Inkvox waitlist.