Give your code
a voice

The setup script checks your hardware, installs dependencies, walks you through cloning a voice from YouTube, and starts the server.

git clone https://github.com/adrianwedd/afterwords.git
cd afterwords
bash setup.shclick to copy

Setup installs the afterwords command to your PATH and starts the server. If Claude Code is installed, it also wires a Stop hook so every response is spoken aloud. Without it, you get a standalone TTS API at localhost:7860.

Each says “You are absolutely right. Your Claude Code session could sound like me.” — generated locally on a 32 GB M1.

Add your own:

afterwords clone "https://youtube.com/watch?v=..." myvoice 30click to copy

Three flagship voices, each synthesized by both Qwen3-TTS sizes. Click a tab to hear the difference between the 0.6B (default, fastest) and 1.7B (higher quality, slower) variants on the same 15-second reference.

Chatterbox and VoxCPM are also loaded as backends but their voice-cloning fidelity in this integration is under verification — samples are tracked in #14 and will be added once we’re confident they’re cloning the reference rather than producing default voices.

Claude Code integration

Standalone API

Programmatic cloning (--allow-clone)

/voice handles input. This project handles output. Together: voice conversations.

The server ships four MLX backends: Qwen3-TTS (0.6B and 1.7B, 8-bit), Chatterbox (fp16, multilingual), and VoxCPM 1.5 (44.1 kHz). Zero-shot voice cloning — no training. A 15-second reference + transcript = cloned voice on every backend.

The server runs on localhost:7860. No authentication. Use it from curl, scripts, other editors, web apps — anything that speaks HTTP. Endpoints marked --allow-clone require launching the server with that flag.

curl localhost:7860/health | jq .click to copy

{
  "status": "ok",
  "model": "mlx-community/Qwen3-TTS-12Hz-0.6B-Base-8bit",
  "backend": "mlx",
  "model_loaded": true,
  "ready": true,
  "voices": ["attenborough", "attenborough-chatterbox", "attenborough-qwen3-17b", "attenborough-voxcpm-15", "audrey", "..."],
  "default_voice": "galadriel",
  "loaded_backends": {
    "qwen3-0.6b":  {"loaded": true, "voice_count": 45, "sample_rate": 24000, "display_name": "Qwen3-TTS 0.6B", "supported_langs": ["en","zh","ja","ko","es","fr","de","it","pt","ru"]},
    "qwen3-1.7b":  {"loaded": true, "voice_count": 61, "sample_rate": 24000, "display_name": "Qwen3-TTS 1.7B", "supported_langs": ["en","zh","ja","ko","es","fr","de","it","pt","ru"]},
    "chatterbox":  {"loaded": true, "voice_count": 3,  "sample_rate": 24000, "display_name": "Chatterbox (fp16, multilingual)", "supported_langs": ["en","es","fr","de","it","pt","zh","ja","ko"]},
    "voxcpm-1.5":  {"loaded": true, "voice_count": 3,  "sample_rate": 44100, "display_name": "VoxCPM 1.5", "supported_langs": ["en","zh"]},
    "voxtral":     {"loaded": true, "voice_count": 0,  "sample_rate": 24000, "display_name": "Voxtral 4B", "supported_langs": ["en","fr","de","es","it","pt","nl","ru","zh","ja","ko","ar","hi"]},
    "soprotts":    {"loaded": true, "voice_count": 0,  "sample_rate": 24000, "display_name": "SoproTTS", "supported_langs": ["en"]}
  }
}click to copy

# Synthesize and play
curl "localhost:7860/synthesize?text=Hello+world&voice=snape" -o out.wav
afplay out.wav

# Non-English (the voice's backend must support the lang, or its family must)
curl "localhost:7860/synthesize?text=Ni+hao&voice=galadriel&lang=zh" -o hi.wav

# Pipe directly to speaker (macOS)
curl -s "localhost:7860/synthesize?text=Testing" | afplay -click to copy

# Interactive (prompts for URL and name)
afterwords clone

# Non-interactive (URL, name, start-second)
afterwords clone "https://youtube.com/watch?v=..." mycustomvoice 30

# Fully automated (skip transcript confirmation)
afterwords clone "https://youtube.com/watch?v=..." mycustomvoice 30 --yesclick to copy

curl -X POST localhost:7860/synthesize \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "voice": "my-session", "emotion": "cheerful"}' \
  -o out.wavclick to copy

# Clone a single voice
curl -X POST localhost:7860/clone \
  -F "audio=@sample.wav" \
  -F "session_id=my-voice" \
  -F "emotion=neutral"

# Build a palette with multiple emotions
curl -X POST localhost:7860/clone \
  -F "audio=@cheerful.wav" -F "session_id=my-voice" -F "emotion=cheerful"
curl -X POST localhost:7860/clone \
  -F "audio=@serious.wav" -F "session_id=my-voice" -F "emotion=serious"click to copy

# After cloning a new voice or editing voices/*.json:
curl -X POST localhost:7860/reload | jq .

# Or via the CLI wrapper:
afterwords reloadclick to copy

curl -X DELETE localhost:7860/session/my-voiceclick to copy

Drop a .afterwords file in any project root. The hook reads it before each synthesis — no server restart.

echo "galadriel" > ~/work/frontend/.afterwords
echo "snape"     > ~/work/backend/.afterwords
echo "loki"      > ~/fun/side-project/.afterwordsclick to copy

If your project uses multiple agents, map each one to its own voice:

# .afterwords — agent-to-voice mapping
default: data
clara-oswald: clara-oswald
donna-noble: donna-noble
k9: k9click to copy

When Claude Code spawns a subagent, the hook reads its agent_type and looks up the matching voice. Falls back to default: if no match.

Clone the same speaker multiple times with different emotional deliveries. The server groups them into a palette by session ID.

# Clone three emotions from the same speaker
curl -X POST localhost:7860/clone \
  -F "audio=@neutral.wav" -F "session_id=narrator" -F "emotion=neutral"
curl -X POST localhost:7860/clone \
  -F "audio=@cheerful.wav" -F "session_id=narrator" -F "emotion=cheerful"
curl -X POST localhost:7860/clone \
  -F "audio=@serious.wav" -F "session_id=narrator" -F "emotion=serious"

# Synthesize with a specific emotion
curl -X POST localhost:7860/synthesize \
  -H "Content-Type: application/json" \
  -d '{"text": "Great news!", "voice": "narrator", "emotion": "cheerful"}' \
  -o out.wavclick to copy

Each palette entry gets a quality rating based on clip duration: rough (<5s), developing (5–15s), good (15s+). If no emotion match is found, the server falls back to the best-quality entry for that session. Clean up with DELETE /session/{id}.

The afterwords CLI is added to your PATH during setup. It wraps launchd, health checks, and voice management into a single tool.

afterwords start       # start the server (auto-starts on login)
afterwords stop        # stop the server
afterwords restart     # restart after adding voices
afterwords status      # show health, model, loaded voices
afterwords logs        # tail the server log
afterwords voices      # list available voices
afterwords reload      # pick up new voices without restart
afterwords clone       # clone a new voice from YouTube
afterwords uninstall   # remove service and optionally hooksclick to copy

The server runs on localhost:7860 and auto-starts on login via macOS launchd. If you prefer to run it manually:

source .venv/bin/activate
python server.py                  # read-only (GET endpoints only)
python server.py --allow-clone    # enables POST /clone, POST /synthesize, DELETE /sessionclick to copy

The setup script installs everything else. Claude Code is optional — use bash setup.sh --server-only for the API without hooks. Either way, you get the afterwords CLI for managing the server.

Qwen3-TTS (Alibaba) · Chatterbox (mlx-community) · VoxCPM (mlx-community) · mlx-audio · MLX (Apple) · Claude Code (Anthropic)

Originally built for SPARK, a robot with an inner life. Full tutorial: Voice Cloning with Qwen3-TTS.