Callout — Text-to-Speech for Claude Code

Speak Claude Code responses aloud using Kokoro TTS (local, 82M params, #1 on TTS Arena) with macOS say as fallback.

Features

/callout — speak the last response on demand
/callout Hello world — speak custom text
/callout --auto-on — automatically speak every response
54 voices (American, British, French, Italian, Japanese, Mandarin)
Configurable voice, speed, and engine
Audio ducking (lowers Apple Music while speaking)
Smart interruption — stops playback when you start typing
Silent operation — no output noise, just audio
Multi-session isolation — each Claude Code session gets its own cache
Markdown stripping for clean speech output
File reading — /callout path/to/file.txt reads a file aloud

Quick Start

git clone https://github.com/xicv/callout.git
cd callout
./install.sh

The installer handles everything: 1. Installs kokoro-tts via uv 2. Downloads model files (~335MB) to ~/.local/share/kokoro-tts/ 3. Registers the plugin hooks with Claude Code 4. Creates the /callout skill in ~/.claude/skills/ 5. Lets you pick your default voice

Then restart Claude Code and use /callout.

Usage

| Command | Description | |---|---| | /callout | Speak the last response | | /callout Hello world | Speak custom text | | /callout --voice=bf_emma | Use a specific voice | | /callout --speed=1.5 | Faster speech | | /callout --voice=am_adam --speed=0.8 Hello | Combine flags with text | | /callout --auto-on | Auto-speak every response | | /callout --auto-off | Disable auto-speak | | /callout --list-voices | Show all 54 voices | | /callout --config | Show current settings | | /callout --stop | Stop playback mid-speech | | /callout path/to/file.txt | Read a file aloud |

Stopping Playback

Two ways to cancel TTS mid-speech:

Start typing your next message — the interrupt hook kills playback automatically
/callout --stop — manual stop command

Configuration

Set environment variables in ~/.claude/settings.json under env:

| Variable | Default | Description | |---|---|---| | CALLOUT_VOICE | af_heart | Voice name (see --list-voices) | | CALLOUT_SPEED | 1.0 | Speed multiplier (0.5–2.0) | | CALLOUT_ENGINE | auto | kokoro, say, or auto | | CALLOUT_MAX_CHARS | 5000 | Max characters to speak | | CALLOUT_AUDIO_DUCK | true | Lower Apple Music during speech | | CALLOUT_DUCK_LEVEL | 5 | Duck volume (% of original) |

Available Voices

American English — Female: af_alloy, af_aoede, af_bella, af_heart, af_jessica, af_kore, af_nicole, af_nova, af_river, af_sarah, af_sky | Male: am_adam, am_echo, am_eric, am_fenrir, am_liam, am_michael, am_onyx, am_puck

British English — Female: bf_alice, bf_emma, bf_isabella, bf_lily | Male: bm_daniel, bm_fable, bm_george, bm_lewis

Other — French: ff_siwis | Italian: if_sara, im_nicola | Japanese: jf_alpha, jf_gongitsune, jf_nezumi, jf_tebukuro, jm_kumo | Mandarin: zf_xiaobei, zf_xiaoni, zf_xiaoxiao, zf_xiaoyi, zm_yunjian, zm_yunxi, zm_yunxia, zm_yunyang

Architecture

callout/
├── .claude-plugin/
│   ├── plugin.json           # Plugin manifest + hooks
│   └── marketplace.json      # Local marketplace config
├── scripts/
│   ├── speak.sh              # Main TTS engine (kokoro → say fallback)
│   ├── cache-response.sh     # Stop hook: caches last response
│   ├── interrupt.sh          # Kills playback on new prompt
│   ├── cleanup.sh            # SessionEnd cleanup
│   ├── strip_markdown.py     # Markdown → plain text (mistune)
│   └── audio-duck.sh         # macOS Apple Music ducking
├── settings.json             # Default env config
├── install.sh                # One-command setup
├── uninstall.sh              # Clean removal
├── pyproject.toml            # Python deps
└── LICENSE                   # Apache 2.0

How it works: 1. Stop hook caches every Claude response to /tmp/callout-{session_id}-response.txt 2. /callout reads the session-scoped cache, strips markdown, pipes to Kokoro TTS streaming 3. UserPromptSubmit hook kills only the current session's TTS process (via PID tracking) 4. SessionEnd hook cleans up only the current session's temp files

Each session is isolated — running multiple Claude Code instances won't cross-talk. The /callout skill is a minimal SKILL.md that delegates all logic to speak.sh via ${CLAUDE_SESSION_ID}, keeping Claude's processing fast.

Prerequisites

Claude Code
uv (for kokoro-tts and Python deps)
jq (usually pre-installed on macOS)
macOS (for say fallback and audio ducking)

Uninstall

./uninstall.sh

Interactively removes:

/callout skill from ~/.claude/skills/
Plugin registration and marketplace from Claude Code
Plugin cache from ~/.claude/plugins/cache/
Temp files and logs
(Optional) Model files from ~/.local/share/kokoro-tts/ (~335MB)
(Optional) kokoro-tts CLI tool

Credits

Kokoro TTS — 82M param open-source TTS model (Apache 2.0)
kokoro-tts CLI — CLI wrapper with streaming
kokoro-onnx — ONNX runtime + model files
ktaletsk/claude-code-tts — reference implementation

License

Apache 2.0

callout

Summary

Install to Claude Code