Aurora and Wake Word Detection

Aurora is my open source Rasperry Pi voice assistant, based on the OpenAI Realtime API. I just shipped native wake word detection, including a trained model (Aurora), a training pipeline and on-device support for gathering training data from your long suffering household ("Aurora, learn how I say your name").

When I first built this project I used Porcupine from Picovoice. They had a free one seat version that worked great for this, but sadly they're pulling the plug at the end of this month. I can't afford their commercial license for an open source assistant so I urgently needed a replacement.

OpenWakeWord is the obvious replacement and seems quite popular with the Home Assistant community. The training notebook does not work though, and some alternatives suggested in the comments wouldn't run for me either. Maybe I'll go back to this if it comes back to life but it didn't look like a good bet for now.

I turned of course to Claude Code. To start with it agreed with the OpenWakeWord intuition but after some persuasion it put together a small CNN (looking at a spectrogram of the current audio capture) and trained and tuned it based on synthetic data. This worked surprisingly well. I'm in the process of fine tuning the model with samples captured from the Raspberry Pi from various family members (kicking off a data capture session is now a feature of the assistant). The detector is a drop in replacement for Porcupine so if you're looking to replace Picovoice that would be a good starting point.

If you use Aurora upgrade to the latest code and then follow the instructions in .env.example to migrate from Porcupine. You can use aurora.onyx to get up and running immediately, or see the wake_word folder readme to train your own from scratch. If you use the Raspberry Pi sample collection copy those files into the data folder and retrain to improve the model.