Ody — Voice Interface for Personal AI Workspace

Problem

Agentic AI workspaces assume the interaction layer is solved. It isn’t — latency, interruption, and hands-free control decide whether voice feels like a tool or a toy.

Solution

Personal research prototype (in development): OpenWakeWord → Faster-Whisper STT → Ollama LLM streaming → Kokoro TTS over asyncio queues with sentence-level streaming. Four interruption triggers cancel the LLM stream and TTS. Single animated orb with five visual states driven by Web Audio amplitude.

Results

• Full voice pipeline working end-to-end over 8 client and 15 server WebSocket events
• Local-first by design: no cloud dependency for the default loop
• Foundation for human–AI interaction experiments (RAG second brain next)