mycroft-precise

> It reliably responds to the wakeword ("hey Mycroft") from men, and only responds about 50% of the time to women. They have instructions on how to train your own version of the wakeword listener. https://github.com/MycroftAI/mycroft-precise#train-your-own-...

Ok, for those that want proof, its pretty simple to do. 1) we know that sending voice data to "HQ" costs power 2) we know that live transcription costs a huge wedge of power 3) we know that wakeword matching is quite power efficient. (see https://rhasspy.readthedocs.io/en/latest/wake-word/, https://github.com/MycroftAI/mycroft-precise) So, in a quite room we know that to save power and data, devices won't be streaming data/listening. We can use that as a baseline for power and network usage. Then we can start talking, measure that then start saying watchwords and see what happens. That aside, we know that its really expensive to listen & transcribe 24/7. Its far easier and cheaper to monitor your web activity and surmise your intentions from that. There are quite a few searches/website visits that strongly correlate to life events. Humans

Precise [1], Snowboy [2], and Porcupine [3] are all designed to work offline. [1] https://github.com/MycroftAI/mycroft-precise [2] https://github.com/kitt-ai/snowboy [3] https://github.com/Picovoice/porcupine

This sounds very close to how I started out with a similar project, named Rhasspy: https://github.com/synesthesiam/rhasspy-hassio-addon I ended up using a set of JSGF grammars (one per intent) to generate a statistical language model for use with pocketsphinx. Rhasspy also features a web-based interface for creating custom words -- I have a mapping from Sphinx phonemes to eSpeak phonemes so you can iterate over a pronunciation until it sounds right. As you mentioned, the wake/hotword stuff with Sphinx isn't terribly robust. I've been Docker-izing Mycroft Precise (https://github.com/MycroftAI/mycroft-precise) to address this.