Voice control for Claude Code inspired by CAAL - full bidirectional loop working

[nayballs]

Hey! Just wanted to say thanks - CAAL inspired me to add voice control to my project Context Mirror.

What we built:

Context Mirror is a scaffolding tool that makes local LLMs (like Qwen-14B) more capable for coding tasks. Tonight we added Voice Mirror - a voice interface that lets you talk to Claude Code and hear responses via TTS.

The full loop works:

"Hey Claude" wake word (custom trained OpenWakeWord model)
Parakeet TDT 0.6B for STT (NVIDIA, ~6% WER, GPU accelerated)
Kokoro 82M for TTS
All local, no cloud STT/TTS costs
Demo we did tonight:

Said "Hey Claude, write a test file with a greeting"
Claude Code received it, created the file
Heard the confirmation spoken back
Also used it to check disk usage by voice and got a detailed spoken breakdown of what was using space.

Key difference from typical setups:

We added a claude_listen MCP tool that lets Claude Code wait for voice input asynchronously. The voice pipeline writes to a shared inbox, Claude picks it up, does the work, responds via the same inbox, and Kokoro speaks it. Full bidirectional.

Uses your existing Claude subscription - no additional API costs for voice since STT/TTS are local.

Repos:

Context Mirror: https://github.com/nayballs/Context-Mirror
Voice Mirror: https://github.com/nayballs/Voice-Mirror
Anyway, just wanted to share and say thanks for the inspiration. Your architecture ideas around VRAM management and the wake word + STT + TTS pipeline were really helpful in thinking through how to approach this. Albeit These Repo's of mine are private ( for now ) They are not ready yet I just wanted to say thank you again.

[cmac86]

Cool! Glad CAAL helped inspire it. Drop a link when you make them public. Cheers!