the dream was one wake word model that runs everywhere, on the phone, on the desktop, and on little esp32 boards stuck in every room. the research killed it. the openwakeword stack needs a speech embedding network that is way too heavy for a microcontroller. the author's own faq says processing a single frame takes several seconds on an esp32-s3, about a hundred times too slow for real time. the weights are only a few megabytes, flash is fine, the problem is activation memory for an embedding network that size.
so the minimum is two builds. on the phone and desktop, openwakeword runs trivially through the onnx runtime, and the custom word trains locally, it is a tiny model, minutes on cpu. on the esp32 it is microwakeword instead, a different framework with a completely different audio front end, and its training is the hard part. it wants twenty to fifty thousand positive clips, the sample generation pipeline has no windows build, and the project's own notes say the output will most likely not be usable at first. everyone who has done it says budget a weekend of retrain and reflash cycles and treat the first model as a benchmark.
the one thing that carries over is the audio. one pool of raw recordings feeds both frameworks, the features just have to be regenerated per side since they do not share a front end. so it is double the training runs but not double the data collection. the alternative is a hosted training service that ships a tiny integer model an esp32 can run at real time, which is tempting until you count the lock in.
takeaway: anyone promising one wake word artifact for everything is wrong, the embedded world and the host world do not share a runtime, and the data is the asset worth building.