Jetson CUDA Voice Pipeline

Fully offline, GPU-accelerated local voice assistant for NVIDIA Jetson devices. No cloud for STT or TTS — only the LLM call uses the internet (OpenRouter or any OpenAI-compatible endpoint).

Architecture

ReSpeaker mic (hw:Array,0, S24_3LE, 16kHz)
    ↓ arecord raw stream — never restarted mid-conversation
openWakeWord — "Hey Jarvis" detection (~32ms chunks)
    ↓ wake word triggered → two-tone beep
_measure_ambient() — 480ms median RMS → dynamic VAD thresholds
    ↓
transcribe_stream() — VAD + whisper.cpp CUDA HTTP (~2-4s per utterance)
    ↓
ask_llm() — OpenRouter or local OpenAI-compatible API (~1-2s)
    ↓
Piper TTS — offline neural TTS, hot-loaded at startup → aplay
    ↓
ReSpeaker LEDs: 🔵 blue=listening  🩵 cyan=thinking  ⚫ off=done  🔴 red=error

Total latency: ~5-8 seconds from wake word to first spoken word.

Key Features

• Zero mic-restart gap — same arecord pipe feeds wake word detection and STT
• Dynamic ambient calibration — measures room noise floor on every wake word trigger (adapts to fans, AC, time of day)
• Conversation history — 20-turn rolling context for natural follow-ups
• Auto language detection — whisper -l auto, works multilingual
• ReSpeaker LED ring — visual state feedback (silent no-op if device not present)
• Fully configurable — all paths and thresholds via environment variables

Hardware Requirements

Quick Start

# 1. Install Python deps
pip install openwakeword piper-tts numpy requests pyusb

# 2. Build whisper.cpp with CUDA (see BUILD.md — ~45 min, one-time)
#    Then place binary at ~/.local/bin/whisper-server-gpu

# 3. Download Piper voice model
mkdir -p ~/.local/share/piper/voices && cd ~/.local/share/piper/voices
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json

# 4. Install and start services
export OPENROUTER_API_KEY=your-key-here
bash pipeline/setup.sh
bash pipeline/manage.sh start

# Say "Hey Jarvis" — blue LED = listening

Setup Details

Build whisper.cpp with CUDA

See BUILD.md for full instructions. Critical flag:

cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=72 -DCMAKE_BUILD_TYPE=Release
make -j4   # ~45 min — detach with nohup if needed

⚠️ CMAKE_CUDA_ARCHITECTURES=72 (sm_72 = Xavier NX) is critical. Default multi-arch compilation OOMs on 8GB Jetson.

Architecture map:

• Xavier NX / AGX Xavier → 72
• Orin → 87
• TX2 → 62
• Nano → 53

Piper Voice Models

mkdir -p ~/.local/share/piper/voices && cd "$_"

# English (required)
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json

# Greek (optional — any language from huggingface.co/rhasspy/piper-voices works)
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/el/el_GR/rapunzelina/medium/el_GR-rapunzelina-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/el/el_GR/rapunzelina/medium/el_GR-rapunzelina-medium.onnx.json

Service Install

setup.sh writes and enables the systemd user services automatically:

bash pipeline/setup.sh [/path/to/voice_pipeline.py] [API_KEY]

Or with env var:

OPENROUTER_API_KEY=sk-... bash pipeline/setup.sh

Re-run to update an existing install.

ReSpeaker Mic Gain & USB Autosuspend

# Optimal gain (no clipping, RMS ~180 ambient)
amixer -c 0 set Mic 90

# Prevent USB autosuspend (mic sleeps after 2s idle without this)
sudo tee /etc/udev/rules.d/99-usb-audio-nosuspend.rules \x3C\x3C 'EOF'
ACTION=="add", SUBSYSTEM=="usb", ATTR{idVendor}=="2886", ATTR{idProduct}=="0007", \
  ATTR{power/control}="on", ATTR{power/autosuspend}="-1"
EOF
sudo udevadm control --reload-rules

Management

bash pipeline/manage.sh start     # start both services
bash pipeline/manage.sh stop      # stop both services
bash pipeline/manage.sh restart   # restart both
bash pipeline/manage.sh status    # systemd status
bash pipeline/manage.sh logs      # tail live log
bash pipeline/manage.sh test-mic  # record 4s + play back
bash pipeline/manage.sh test-stt  # record 4s + transcribe
bash pipeline/manage.sh test-tts  # speak a test phrase

Environment Variables

Troubleshooting

Mic records silence

• Check gain: amixer -c 0 set Mic 90
• Use card name not number (hw:Array,0 not hw:0,0) — numbers shift on reboot
• ReSpeaker requires S24_3LE format, not S16_LE
• Disable USB autosuspend (see setup above)

Records full 6s timeout, never cuts off

• Room ambient noise > VOICE_SILENCE_RMS fallback. Dynamic calibration handles this automatically.
• If still an issue, set VOICE_SILENCE_RMS slightly above your measured ambient floor.

[BEEPING] or (bell dings) in transcript

• Speaker beep being picked up by mic. The 0.3s drain buffer after beep handles this.
• Check speaker/mic distance and speaker volume.

Whisper OOM during build

• Must use -DCMAKE_CUDA_ARCHITECTURES=72 — default multi-arch build exhausts 8GB RAM.
• Use -j4 not -j6.

LED not lighting up

• Install pyusb: pip install pyusb
• Only supported on ReSpeaker USB Mic Array v1.0 (2886:0007)
• All LED errors are silent — pipeline continues without it.

Wake word triggers constantly (false positives)

• Lower VOICE_WAKE_THRESHOLD to 0.7 or higher.
• Ensure no TV/radio playing phrases close to "Hey Jarvis".

File Structure

jetson-cuda-voice/
├── SKILL.md                  ← this file
├── BUILD.md                  ← whisper.cpp CUDA build guide
└── pipeline/
    ├── voice_pipeline.py     ← main pipeline
    ├── led.py                ← ReSpeaker LED control (optional)
    ├── setup.sh              ← one-command service installer
    └── manage.sh             ← start/stop/status/test