my local, AI Voice Assistant (I replaced Alexa!!)

A fully local voice assistant is now practical for home automation: Home Assistant can run an offline wake word, speech-to-text, intent handling, and text-to-speech on your own hardware—then scale those pieces across multiple devices using the Wyoming protocol. The payoff is control without cloud dependence, plus the ability to swap in a local “brain” like Llama 3 for more capable, context-aware conversations that can drive real actions in the house.

The build starts with Home Assistant on a local device (a Raspberry Pi in the demo) and adds a “voice pipeline” through add-ons. Whisper provides offline speech-to-text, Piper handles offline text-to-speech, and Open Wake Word listens for a chosen wake phrase. Home Assistant’s Assist layer then connects those audio components to home automation actions—turning phrases into commands like switching lights—while keeping everything on the local network. Early tests show the system working but requiring careful phrasing and reacting with some latency, which sets up later improvements.

Next comes scaling beyond a single box. The Wyoming protocol turns extra hardware into “satellites” that can listen and speak while delegating the heavy lifting back to Home Assistant. A second Raspberry Pi is configured with a ReSpeaker 2-Mic Pi Hat, then the Wyoming satellite software is installed and run as a service so it stays online. Home Assistant connects to it over the network, and the assistant can control devices from anywhere in the house. The demo also adds an LED status behavior using the ReSpeaker’s pixel ring so the user can tell when the assistant is actively listening.

The real leap in capability arrives when the “conversation agent” is replaced with a local large language model. Instead of relying on Home Assistant’s default conversational behavior, the system points to Ollama running Llama 3.2 (downloaded and served locally). With the LLM in place, the assistant can answer factual questions and—crucially—maintain context across turns, enabling follow-up commands like turning a light back on after a prior interaction.

To reduce bottlenecks, the pipeline is offloaded to more powerful machines using Wyoming containers. On a Windows laptop, Docker runs Wyoming Whisper (speech-to-text) and Wyoming Piper (text-to-speech), and Home Assistant switches its voice pipeline endpoints to those remote services. The demo then adds a second LLM server (“Terry,” an AI server) so multiple models can be used together. The result is a fast, fully self-hosted assistant that can control smart lighting and handle more natural conversation—while still staying offline from the cloud.

The build ends with remaining gaps compared with Alexa: custom wake word training (e.g., “Terry”) and custom voice generation aren’t fully solved yet. The next step is training a new wake word using Open Wake Word’s Google Colab workflow and uploading the resulting model files into Home Assistant via the Samba add-on. Custom voice cloning is flagged as the next frontier for a future video, after hours of troubleshooting. Overall, the message is clear: with Home Assistant, Wyoming, and local LLM tooling, a cloud-free home assistant can be assembled piece by piece—and then upgraded as hardware and models improve.