Local Voice Assistant

An Alexa-equivalent running entirely on the Raspberry Pi cluster — no cloud, no audio leaving the house.

I wanted an Alexa without Amazon. This is one: the whole assistant (wake word, speech recognition, a 4-billion-parameter language model, and text-to-speech) runs on three Raspberry Pi 4s, with a fourth handling storage. No cloud, no audio leaving the house.

The cluster the assistant runs on: four Raspberry Pi 4 (8GB) boards, two cooling fans, and a gigabit switch.

What it does

Answers questions, gives live weather, sets timers, looks things up on Wikipedia, and describes what a camera sees — all locally. One node runs the wake word, speech recognition, and vision; one runs the language model; one does text-to-speech; the fourth is the cluster control plane and NAS.

The interesting part

A custom “ok computer” wake word that kept firing on “ok” alone, and what fixing it taught me about training data and shortcut learning.

Read the full writeup →