1. Pick native Windows or WSL before setting up voice
Anthropic supports three Windows setups. Native Windows runs from PowerShell or CMD, needs no administrator rights, and uses Git Bash for its Bash tool when Git for Windows is installed, or a PowerShell tool when it is not. WSL 2 adds sandboxed command execution; WSL 1 is the fallback when WSL 2 is unavailable, without sandboxing.
For voice, that choice matters more than it looks. Native Windows records through Claude Code’s built-in audio module and Windows’ own microphone permissions. Inside WSL, Claude Code is a Linux program, so it needs a Linux audio path to the Windows microphone, and only WSLg provides one.
- Native, PowerShell: irm https://claude.ai/install.ps1 | iex
- Native, WinGet: winget install Anthropic.ClaudeCode (WinGet installs do not auto-update; run winget upgrade Anthropic.ClaudeCode).
- WSL: run curl -fsSL https://claude.ai/install.sh | bash inside the distribution and launch claude there.
- Check the result with claude --version, or claude doctor for a fuller report.
2. Turn on /voice in PowerShell, CMD or Windows Terminal
Sign in with a Claude.ai account. Anthropic says dictation is unavailable when Claude Code uses an Anthropic API key directly, Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry, and an API key in your environment takes precedence over a browser sign-in. Then run /voice: Claude Code checks the microphone and reports the mode and dictation language. Hold Space to talk, or run /voice tap to start and stop with single taps.
If you see “Microphone access is denied”, open Settings → Privacy & security → Microphone, turn on microphone access for desktop apps, and run /voice again. If recording starts but you get “No audio detected from microphone”, choose the right device under Settings → System → Sound → Input and check that it is not muted.
Transcription is not local on any platform. Anthropic says the recorded audio is streamed to its servers to be transcribed; it does not count against your Claude usage limits.
3. Make /voice record inside WSL2
Anthropic’s requirement for WSL is WSLg, which is included with WSL2 installed from the Microsoft Store on Windows 10 or 11. Microsoft lists Windows 10 build 19044 or later, or Windows 11, for WSLg and says Linux GUI support works only with WSL 2 distributions. On WSL1, Anthropic’s advice is to run Claude Code in native Windows instead.
WSLg routes audio through PulseAudio rather than an ALSA device, and WSL has no /dev/snd. Installing sox alone brings the ALSA backend, which cannot record there, and Claude Code reports “Voice mode could not find a working audio recorder in WSL”. The fix is to install SoX’s PulseAudio backend explicitly.
- In PowerShell, run wsl --update, then wsl --shutdown to restart WSL.
- In the distribution, run sudo apt install sox libsox-fmt-pulse.
- Start claude again and run /voice.
4. Windows voice typing (Win+H) as the built-in fallback
Windows 11 voice typing starts with the Windows logo key + H. Microsoft says it uses online speech recognition powered by Azure Speech services, so it needs an internet connection, a working microphone and a cursor in a text box. Win+Alt+H moves through the voice typing menu from the keyboard, and the gear icon holds the voice typing launcher, automatic punctuation, profanity filtering and default microphone settings. On Windows 10 the same shortcut opens the dictation toolbar, which also needs an internet connection.
Spoken commands work in a prompt. Microsoft’s English list includes “press Enter”, “backspace”, “delete that”, “undo that” and “stop listening”, so say “press Enter” only when you mean to submit. On Copilot+ PCs, Fluid dictation tidies grammar, punctuation and filler words with on-device small language models, but Microsoft still describes the voice typing recognition itself as online.
Voice typing decides for itself whether the focused field counts as a text box. A March 2026 user report in Anthropic’s issue tracker describes Win+H opening over the Claude Code prompt and asking the user to select a text box. If that happens to you, use /voice or a dictation app that pastes into the terminal.
5. Dictate on-device with IraVoice in Windows Terminal
IraVoice 1.2.2 runs on Windows 10 or 11, 64-bit (x64) and recognizes speech on-device with NVIDIA Parakeet, so after the one-time model download it needs no internet connection and no Claude.ai account. Hold Right Ctrl with the Claude Code prompt focused, speak, and release; IraVoice pastes the text with Ctrl+V and then puts your clipboard back. It does not press Enter by default, Esc cancels, and tapping Right Ctrl instead of holding it keeps listening hands-free until you tap again.
Windows Terminal, PowerShell, pwsh and cmd are on IraVoice’s list of recognized apps, along with VS Code, Cursor and the Claude app. In those, IraVoice finds the Git project the session is working in and types its names the way the code spells them: “get user by ID” becomes getUserById. It keeps only names, never source text, skips secrets and dependency folders, and nothing leaves the PC. If it mishears a word, press Ctrl+Alt+Shift+F, correct the text and press Enter, and it types that word your way from then on.
Two Windows limits are worth knowing. Inside a WSL tab, dictation still types into the window, but the Windows app cannot see the Linux processes behind it, so do not count on code names there. And if the terminal runs as administrator, Windows blocks simulated input from normal apps; IraVoice leaves the text on the clipboard and says so, so you can paste it yourself. If a terminal ignores Ctrl+V, switch IraVoice’s paste method to Shift+Insert, Ctrl+Shift+V or direct typing.