voiceinput

Voice-to-text input for Claude Code. Hold a push-to-talk key, speak, and the transcribed text is injected into your Claude Code session. Uses Groq's Whisper API by default for fast, accurate transcription.

Prerequisites

• Linux with Python 3.10+
• libportaudio2 for audio capture
• tmux for terminal multiplexing
• A Groq API key (free)

sudo apt install libportaudio2 tmux

Installation

cd voiceinput
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Input device access

Your user needs read access to /dev/input/event* devices:

sudo usermod -aG input $USER

Log out and back in for this to take effect. In the meantime, newgrp input works in the current shell.

Groq API key

export GROQ_API_KEY="your-key-here"

Add this to your ~/.bashrc to persist it.

Quick start

The easiest way to use voiceinput is the claude-voice wrapper:

# Optional: symlink for convenience
sudo ln -s "$(pwd)/claude-voice.sh" /usr/local/bin/claude-voice

# Start Claude Code with voice input
claude-voice

This starts Claude Code in a tmux session with a shared voiceinput daemon running in the background. Just hold the hotkey and speak.

Hotkeys

Key	Action
F4	Record and type text into Claude Code
F5	Record, type text, and auto-submit (sends Enter)

Multiple instances

Run claude-voice multiple times to open multiple Claude Code sessions. A single voiceinput daemon is shared across all of them. Text is injected into whichever session you most recently interacted with.

# Terminal 1
claude-voice

# Terminal 2
claude-voice

Focus the Claude session you want, hold F5, speak — text goes there.

Manual setup (without wrapper)

If you prefer to manage tmux yourself:

# In the Claude Code pane, get the pane target:
tmux display-message -p '#{session_name}:#{window_index}.#{pane_index}'

# In another pane, run voiceinput with that target:
source /path/to/voiceinput/.venv/bin/activate
voiceinput --target 0:0.0

Options

Flag	Env var	Default	Description
--hotkey	VOICEINPUT_HOTKEY	KEY_F4	evdev key name for type-only
--submit-hotkey	VOICEINPUT_SUBMIT_HOTKEY	KEY_F5	evdev key name for type+submit
--target	VOICEINPUT_TARGET	(auto-detect)	tmux pane target
--backend	VOICEINPUT_BACKEND	groq	groq or local
--groq-api-key	GROQ_API_KEY		Groq API key
--model-size	VOICEINPUT_MODEL_SIZE	base.en	whisper model (local backend)
--silence-threshold	VOICEINPUT_SILENCE_THRESHOLD	10	RMS silence threshold
--sample-rate	VOICEINPUT_SAMPLE_RATE	16000	audio sample rate
--device-path	VOICEINPUT_DEVICE_PATH	(auto-detect)	evdev input device path
--instance			named routing, e.g. --instance p210=session:0.0

Local transcription

To use local faster-whisper instead of Groq:

pip install -e ".[local]"
voiceinput --backend local

Troubleshooting

"No keyboard device found" — You're not in the input group or the device isn't detected. Specify manually with --device-path /dev/input/eventN. List devices:

python3 -c "import evdev; [print(evdev.InputDevice(d).path, evdev.InputDevice(d).name) for d in evdev.list_devices()]"

"(silence, skipped)" on every recording — Your mic may have very low levels. Lower the threshold with --silence-threshold 2 or check your audio device with pactl list sources short.

Wrong language detected — The Groq backend is set to English by default. For other languages, edit the language parameter in voiceinput/transcribe.py.

"No active tmux session found" — voiceinput can't find a tmux client to send text to. Make sure Claude Code is running inside tmux.