voice-input

name	voice-input
description	- User says they want to speak their prompt rather than type it

Voice Input

When to activate

• User says they want to speak their prompt rather than type it
• User asks how to use voice or dictation in Claude Code
• User asks about /voice, push-to-talk, or voice mode
• User wants to go hands-free while reviewing code on a second monitor
• User asks how to rebind the push-to-talk key or change the recording language

When NOT to use

• User is in an SSH session — microphone input is not forwarded over SSH; voice is unavailable
• User authenticated with a raw API key only (no claude.ai account) — voice requires a claude.ai account
• User is working in the Claude web interface — /voice is a CLI-only command
• User is on Linux and has not confirmed arecord or sox is installed
• User's question is about Claude API voice features — that is a separate system unrelated to this skill

Instructions

Enable voice

Run inside any Claude Code session:

/voice        # toggle on (defaults to hold mode)
/voice hold   # hold Space to record, release to send
/voice tap    # tap Space once to start, once to stop and send
/voice off    # disable

Choose the right mode

Hold mode — press and hold Space while speaking, release to send. Best for short-to-medium prompts. Less friction for quick questions.

Tap mode — tap Space once to start recording, tap again to stop and send. Best for longer dictation where holding a key is awkward.

Persist the setting

Add to ~/.claude/settings.json:

{
  "voice": {
    "enabled": true,
    "mode": "tap"
  }
}

This survives session restarts. Switch between "hold" and "tap" as needed.

Rebind the push-to-talk key

The default key is Space. To change it, edit ~/.claude/keybindings.json:

{
  "Chat": {
    "voice:pushToTalk": "v"
  }
}

Use any key that does not conflict with your normal typing. v, F9, or backtick are common choices. The binding is scoped to the Chat context.

Set the transcription language

Add a language key at the top level of ~/.claude/settings.json:

{
  "voice": { "enabled": true, "mode": "hold" },
  "language": "fr"
}

Supported languages: 20 total, including en, fr, de, es, nl, ja, zh, pt, ko, it. Use BCP 47 tags.

Linux / WSL setup

Linux (ALSA):

sudo apt install alsa-utils

Linux (sox alternative):

sudo apt install sox

WSL:

sudo apt install sox libsox-fmt-pulse
# WSLg must be active — update WSL from PowerShell: wsl --update

macOS works without any setup.

Example

Scenario: A developer is refactoring a large module and wants to dictate a detailed instruction without breaking their reading flow.

1. Switch to tap mode for longer dictation:

   /voice tap
   

2. Tap Space to start recording, then dictate:

   "Split the UserController class into three focused modules: user-auth.ts for login and token handling, user-profile.ts for CRUD on profile data, and user-preferences.ts for settings. Move the existing tests to match the new structure. Keep the existing public interface intact — nothing in routes/ should need to change."

3. Tap Space again to stop. Review the transcription in the input field, make any corrections, then press Enter.

Result: A precise, multi-sentence prompt delivered without typing — and without losing focus on the code being read.

---