AI Voice Cloning Tool — Create custom notification sounds for Claude Code using any voice from YouTube videos. Powered by Qwen3-TTS, Whisper, and Demucs.

Claude Code plays a notification sound when it needs your attention — a permission prompt, a finished task, and so on. This tool lets you replace those sounds with a cloned voice of your choice, cloned from any YouTube clip.

Point it at a YouTube URL (an interview, a stream, a podcast), pick a clean few seconds of speech, and it generates a full set of Korean/English notification lines in that person's voice — ready to wire into Claude Code with one command. Everything runs locally on your machine (Apple Silicon or an NVIDIA GPU); no audio ever leaves your computer.

Prefer to just hear it first? Jump to the 🔊 Voice Samples at the bottom.

git clone https://github.com/t1seo/karina-voice-notification.git
cd karina-voice-notification

pixi install
pixi run install-deps-mac    # macOS (Apple Silicon)
pixi run install-deps-linux  # Linux (NVIDIA GPU)

The pipeline turns a raw YouTube link into clean voice notifications in six steps:

flowchart LR
    A([YouTube URL]) --> B[Download<br/>yt-dlp]
    B --> C[BGM Removal<br/>Demucs]
    C --> D[Split &amp; Select<br/>clean segment]
    D --> E[Transcribe<br/>Whisper large-v3]
    E --> F[Voice Clone<br/>Qwen3-TTS 1.7B]
    F --> G([Notification .wav])

Works with both Claude Code and Codex — same skills, same sounds.

Conversationally (recommended) — in Claude Code or Codex, run the skill and it walks you through it: paste a YouTube link, choose what each alert should say, done.

/generate-voice

Or via the interactive CLI:

pixi run pipeline          # menu-driven: URL → segment → generate
# or one-shot, non-interactive:
pixi run quickstart "https://youtu.be/VIDEO_ID" --line "idle_prompt:다 됐어요!"

Either way the notification set lands in output/notifications/.

Run the setup skill in either tool:

/setup-notifications

Or run the installer directly:

pixi run install-notifications          # both tools (auto-detects)
python scripts/install_notifications.py --tool codex   # just Codex
python scripts/install_notifications.py --dry-run      # preview changes

It copies the sounds and wires up the events — Claude Code: Stop + Notification hooks in ~/.claude/settings.json; Codex: a notify program in ~/.codex/config.toml (fires on turn completion) plus the skills into ~/.codex/skills/. Every edited file is backed up, and it's safe to re-run. Restart the tool afterward to load the hooks.

Good voice sources

• Interview clips, solo speaking, podcasts
• Enable BGM Removal for music videos

Avoid

• Noisy environments or multiple speakers
• Clips shorter than 5 seconds

Edit notification_lines.json to change the phrases:

{"text": "Your custom phrase here", "filename": "permission_prompt_1.wav"}

Three Korean notification lines, cloned in Karina's voice (interview source) with Qwen3-TTS. Press ▶ to play:

Task complete — 작업을 완료했습니다.

Permission required — 실행 허가가 필요합니다.

Authentication succeeded — 인증에 성공했습니다.

Players are waveform videos so they play inline on GitHub. Source .wav files are in assets/samples/; regenerate with pixi run samples.

MIT License