Home Assistant Voice turns reliably local

This is the most convincing case yet that open, self-hosted voice can feel good in daily use, not just in demos. It also underlines how far the ecosystem still is from true plug-and-play.

• –The post is strongest where it gets specific: tested GPUs, model sizes, response times, and which STT/TTS components actually held up in real-world use.
• –Speech-to-text looks like the real make-or-break layer here; the author got much better results from Parakeet-based Wyoming ONNX ASR than from weaker default local setups.
• –llama.cpp plus better-quantized local models noticeably improved tool calling, context handling, and recovery from misheard commands, which is exactly where local voice assistants usually fall apart.
• –Home Assistant’s extensibility shines through integrations like LLM Conversation and llm-intents, but the need for custom prompts and sentence automations shows the product still expects hobbyist-level effort.
• –For AI developers, the interesting part is less “smart home” and more the stack design lesson: local UX quality comes from pipeline engineering, not just dropping in an LLM.