Voice Prompt Skill
jayu1023/voice-prompt-skill
Ditado por voz 100% local para Claude Code com aprendizado de vocabulário
Ver no GitHubSkill de ditado por voz para Claude Code que transcreve suas falas em tempo real usando faster-whisper, sem chamadas de API. Suporta comandos de voz para enviar prompts, pausar/retomar e aprende automaticamente seu vocabulário conforme você usa.
Instalação
npx convez-ai jayu1023/voice-prompt-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Digitar prompts sem usar teclado durante codificação
- Melhorar produtividade em sessões hands-free de programação
- Personalizar reconhecimento de fala com termos técnicos e nomes de funções
Veredito

Ferramenta madura e bem implementada para quem quer interação por voz com Claude Code. Destaque para processamento 100% local, cross-platform robusto e aprendizado adaptativo de vocabulário. Exige configuração de permissões no SO, mas a instalação é simples.
README
A tradução para o português deste README ainda está na fila.
Hands-free dictation for Claude Code. Speak your prompt, watch it get typed into the chat, and say "go" to submit. All transcription runs locally with faster-whisper — no API calls, no internet required after first model download.
✨ Features
- 🗣 Voice-to-text in real time via faster-whisper (5× faster than openai-whisper, no PyTorch)
- ⌨️ Auto-types into whichever window is focused — works with Claude Code in any terminal
- 🚀 Submit triggers:
go·send it·hit enter·submit·do it - ⏸️ Pause / resume on demand:
pause listening·mute·go to sleep↔resume listening·unmute·wake up - 🧠 Self-learning vocabulary: every prompt you submit is fed back into Whisper as bias context so accuracy improves with your jargon, function names, accent quirks, and project terms
- 🎚 Learner toggle at runtime:
stop learning↔start learning— full user control over what gets remembered - ✏️ User-editable vocab file at
~/.claude/skills/voice-prompt/state/vocab.txt— drop in domain words you use a lot - 🛑 Exit:
stop listening·quit listening·exit voice - 🔒 100% local — your speech never leaves the machine
- 🖥 Cross-platform: macOS, Linux (X11), Windows
- 🪶 Lightweight: ~150 MB model, no PyTorch
📦 Install
One-liner (recommended)
npx skills add jayu1023/voice-prompt-skill
Manual
git clone https://github.com/jayu1023/voice-prompt-skill.git
cp -r voice-prompt-skill/skill ~/.claude/skills/voice-prompt
bash ~/.claude/skills/voice-prompt/setup.sh
Platform-specific prerequisites
| OS | Extra install |
|---|---|
| macOS | brew install portaudio (the setup script handles this) |
| Linux (Debian/Ubuntu) | sudo apt install xdotool portaudio19-dev |
| Linux (Fedora) | sudo dnf install xdotool portaudio-devel |
| Linux (Arch) | sudo pacman -S xdotool portaudio |
| Windows | pip install pyautogui (the setup script handles this) |
macOS permissions (one-time)
After install, grant your terminal both of these in System Settings → Privacy & Security:
- Microphone — so audio can be captured
- Accessibility — so keystrokes can be sent
⚠️ Without Accessibility, typing silently does nothing. This is the #1 cause of "it doesn't work".
🚀 Usage
Start the listener in a separate terminal so your Claude Code window can keep focus:
python3 ~/.claude/skills/voice-prompt/voice_prompt.py
Or invoke through Claude Code itself — once installed, just ask:
"Start voice input"
Then click into your Claude Code window and start talking. As you speak, words appear in the input. When you say "go" or "send it", the trigger word is stripped and Enter is pressed.
Example session
You say: "explain the difference between use effect and use layout effect go"
Typed: "explain the difference between use effect and use layout effect"
Submitted: <Enter pressed> [learned ← phrase saved to history]
The learner explained
Every time you submit a prompt, the text you said (everything before go / send it) gets appended to:
~/.claude/skills/voice-prompt/state/history.txt
The next chunk of audio Whisper transcribes is biased by an initial_prompt built from:
- Your manually curated
state/vocab.txt(always loaded — edit it directly) - The last ~80 accepted prompts from
history.txt(only if learner is ON)
So if you keep saying "useLayoutEffect" and Whisper keeps hearing "use layout effect", once it transcribes correctly once and you submit, it learns. After a week of use, Whisper knows your codebase's vocabulary.
Privacy / control:
stop learning(voice) → flips a flag file, nothing new gets appendedstart learning(voice) → re-enables it--no-learn(CLI flag) → start a session with the learner off- Just delete
~/.claude/skills/voice-prompt/state/history.txtanytime to wipe what was learned
⚙️ Configuration
Edit ~/.claude/skills/voice-prompt/voice_prompt.py:
| Setting | Default | Notes |
|---|---|---|
MODEL_SIZE | base.en | tiny.en (fastest, less accurate), small.en, medium.en (best, slow on CPU) |
COMPUTE_TYPE | int8 | int8_float16 or float16 if you have Apple Silicon / GPU |
CHUNK_SECONDS | 2.5 | Lower = snappier feedback, higher = better context |
TRIGGER_WORDS | {"go", "send it", ...} | Phrases that submit. Use rarer words if you get false positives |
STOP_WORDS | {"stop listening", ...} | Phrases that exit the listener |
🛠 Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Words don't appear when typing | Missing Accessibility permission (macOS) | System Settings → Privacy & Security → Accessibility → add your terminal |
| Listener exits with "no audio device" | Missing Microphone permission | Same path → Microphone |
| Trigger word fires mid-sentence | Common word matched accidentally | Edit TRIGGER_WORDS — use phrases like "computer go" |
| Transcription is slow / garbled | CPU-bound on large model | Switch MODEL_SIZE to tiny.en or base.en |
| Linux: nothing types | Wayland session (xdotool is X11-only) | Install ydotool and replace the calls |
🤝 Contributing
PRs welcome — especially for:
- Wayland support (ydotool)
- A native Cocoa keystroke backend (faster than osascript)
- Wake-word activation ("hey claude…")
- Punctuation commands ("comma", "new paragraph", etc.)
📜 License
MIT — do whatever you want, just don't sue.
💬 Credits
Built by @jayu1023 with Claude Code.
Uses:
- faster-whisper — CTranslate2 reimplementation of Whisper
- sounddevice — PortAudio bindings
- pyautogui — cross-platform input on Windows
⭐ If this saved you from RSI or just feels magical, star the repo!