A low-latency, offline voice-trigger / hotphrase runner designed for NVIDIA Jetson Orin (JetPack 6 / CUDA 12.x), but also works on x86 Linux and CPU-only ARM boards.

• WebRTC VAD (20ms frames) → short speech segments for low latency
• faster-whisper (CTranslate2) CUDA on Jetson; auto fallback to CPU
• Fuzzy phrase matching → run shell commands
• UDP control + remote triggers:
  • CTRL:PAUSE, CTRL:RESUME
  • TRIGGER:<phrase> (must exist in triggers)
  • CMD:<shell command> (optional; can be disabled)
• Optional model prefetch and warmup to avoid a slow first trigger

git clone <YOUR_REPO_URL>
cd jetson-voice-trigger
chmod +x scripts/install_orin.sh
./scripts/install_orin.sh

source .venv/bin/activate
python self_test.py --model tiny.en

source .venv/bin/activate
python -m voice_trigger.app --asr_device auto --model medium.en --compute float16 --lang en

You can either edit defaults in src/voice_trigger/config.py, or provide an external JSON file:

triggers.json

{
  "open browser": "xdg-open https://www.wikipedia.org",
  "say hello": "bash -lc 'notify-send \"Trigger\" \"Hello\"'"
}

Run with:

python -m voice_trigger.app --triggers triggers.json

Enable UDP listener:

python -m voice_trigger.app --udp_in --udp_host 0.0.0.0 --udp_port 9999 --udp_token SECRET

Messages (prepend SECRET: when token is used):

• CTRL:PAUSE
• CTRL:RESUME
• TRIGGER:<phrase>
• CMD:<shell command> (disabled by default; enable with --allow_udp_cmd)

• For medium/large: --compute float16 and --asr_device cuda (or auto)
• Use --prefetch-model to download weights ahead of time
• Use --warmup-sec 1.0 to warm kernels/cache

Example:

python -m voice_trigger.app --asr_device auto --model medium.en --compute float16 --prefetch-model --warmup-sec 1.0

On Jetson/AArch64, you typically need to build CTranslate2 from source with CUDA:

./scripts/install_orin.sh

Install:

• libportaudio2 portaudio19-dev libasound2-dev

List devices:

python -c "import sounddevice as sd; print(sd.query_devices())"

• UDP CMD: executes arbitrary shell commands → disabled by default
• If enabling it, always use --udp_token

MIT