jetson-cuda-voice

name	jetson-cuda-voice
version	1.1.0
description	Fully offline, CUDA-accelerated local voice assistant pipeline for NVIDIA Jetson. Wake word (openWakeWord) → real-time VAD → whisper.cpp GPU STT → LLM → Piper TTS. Includes dynamic ambient noise calibration, conversation history, and ReSpeaker LED feedback. Tested on Jetson Xavier NX (sm_72, JetPack 5.1.4) with ReSpeaker USB Mic Array.
metadata	{"openclaw":{"emoji":"🎙️","os":["linux"],"requires":{"bins":"[Truncated]","env":"[Truncated]"},"notes":{"hardware":"Tested on NVIDIA Jetson Xavier NX (ARM64, sm_72, JetPack 5.1.4, 8GB). Mic: ReSpeaker USB Mic Array v1.0 (VID 2886:PID 0007) — requires S24_3LE format. Speaker: any ALSA device. LED feedback optional (requires pyusb). Other Jetson models: adjust CMAKE_CUDA_ARCHITECTURES (Orin=87, Nano=53, TX2=62).\n"}}}

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

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