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

About this video

The creator replaces a cloud voice assistant with a fully local system built on Home Assistant and the open-source Rhasspy/Wyoming stack. Level one runs entirely on a Raspberry Pi 4: Whisper handles speech-to-text locally, Piper handles text-to-speech, OpenWakeWord provides the wake word, and Home Assistant’s Assist handles intent so commands like turning off a named lamp work without any cloud service. He then builds a remote Wyoming satellite on a second Raspberry Pi using a two-microphone hat, registers it with Home Assistant on port 10700, and turns it into an always-on service with status LEDs. In the later levels he offloads the heavy work to a laptop and then to his AI server: Ollama running Llama 3.2 becomes the conversation agent via port 11434, while Wyoming Whisper (port 10300) and Wyoming Piper (port 10200) run in Docker containers, giving the assistant conversational memory and faster responses — it correctly answers that The Matrix was released in 1999. He also trains a custom ‘Terry’ wake word and shows voice-sentence automations for canned replies. He closes by admitting the stock experience is still slower and less polished than Alexa, that a custom Terry voice eluded him despite a full day of troubleshooting, and that a screen-equipped ESP32-S3 box is an option for a satellite with a display.

Key takeaways

• A local voice pipeline can be assembled from Home Assistant add-ons: Whisper for speech-to-text, Piper for text-to-speech, and OpenWakeWord for wake-word detection
• Wyoming satellites let cheap Raspberry Pi remotes reuse one central Home Assistant voice pipeline, so multiple rooms can share the same assistant
• Swapping the built-in conversation agent for a local LLM such as Ollama with Llama 3.2 adds general question answering and conversational context the basic intent system lacks
• Offloading Whisper and Piper to a more powerful machine in Docker noticeably speeds up response times compared with running everything on a Raspberry Pi 4
• Custom wake words require training a small model rather than just typing a name, and custom voices were the one goal he did not achieve in this video

Episode details: published November 4, 2024 · 39:41 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: your home automation SUCKS!! · Next: Cloning my Voice Into an AI Assistant →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube