LARS Voice Assistant - System Overview

Vision

An always-on, local-first voice assistant that: - Listens for "Hey LARS" wake word - Processes speech locally via Whisper STT - Routes to LARS (local Ollama model) for AI responses - Uses existing Nexus voice pipeline (InWorld AI) for TTS - Can delegate complex tasks to Claude Code via SSH

Key Requirements

1. Wake Word Detection

• Always listening, low power
• Custom "Hey LARS" trigger
• Can run on: Linux server, tablet app, phone app, Raspberry Pi
• Preferred: openWakeWord (open source, no expiration)

2. Speech-to-Text (STT)

• Use Whisper for transcription
• Local processing, no cloud
• Whisper is acceptable for STT quality

3. Text-to-Speech (TTS)

• DO NOT use Kokoro or Whisper TTS (poor quality)
• USE existing Nexus voice pipeline with InWorld AI
• Already working well, keep it

4. AI Processing

• LARS via Ollama (lars-trained model)
• 3D reasoning format (/)
• Local, private, unrestricted

5. Claude Code Delegation

• When LARS can't handle complex tasks
• SSH connection to Claude Code CLI
• Research N8N SSH node implementation
• Goal: Python/JSON equivalent (no N8N dependency)

Data Flow

[Microphone] → [Wake Word: "Hey LARS"]
     ↓
[Whisper STT] → Text
     ↓
[LARS/Ollama] → Response
     ↓ (if complex)
[Claude Code via SSH] → Enhanced Response
     ↓
[InWorld AI TTS] → [Speaker]

Deployment Options

1. Linux server (local-ai server)
2. Web app (tablet/phone)
3. Raspberry Pi dedicated device
4. Home Assistant integration