My voice coding stack: talk to Claude Code, hear it talk back

I rebuilt my dev setup so I can talk to Claude Code and it talks back.

The stack

• Speech to text: Handy with a Parakeet model, running locally.
• Cleanup: a small LLM pass that fixes identifiers and punctuation.
• Text to speech: VibeVoice on Apple Silicon.

Gotchas

• The OpenAI compatible endpoint in Ollama forces temperature to 1.0 when the client does not send it. Setting top_k 1 in a Modelfile makes output deterministic anyway.
• Number examples in a cleanup prompt can make a small model invent amounts that were never spoken.
• Loading a TTS model in float32 on a Mac doubled its memory. Switching to bfloat16 took it from 12 GB to under 7 GB.

This is placeholder content. Replace it with the real post.