snowboy

Great question. This is technically referred to as "Wake Word Detection". You run a really small model locally that is just processing 500ms (for example) of audio at a time through a light weight CNN or RNN. The idea here is that it's just binary classification (vs actual speech recognition). There are some open source libraries that make this relatively easy: - https://github.com/Kitt-AI/snowboy (looks to be shutdown now) - https://github.com/cmusphinx/pocketsphinx This avoids having to stream audio 24x7 to a cloud model which would be super expensive. This being said, I'm pretty sure what the Alexa does, for example, is send any positive wake word to a cloud model (that is bigger and more accurate) to verify the prediction of the local wake word detection model AFAIK. Once you are positive you have a positive wake word detected - that'

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

As far as I know there are no open source solutions for wake word detection. Most 'open source Alexa' projects require you to press a button to make it listen. There is a free offline wake word detector here: https://github.com/Kitt-AI/snowboy ... ok that was closed source when I looked last! Looks like you still need to use their website for training though.

If you make your own Alexa device you can do so. It's simple enough with a Raspberry Pi and a microphone. Unfortunately you lose the far-field mics is you do this. https://github.com/alexa/alexa-avs-sample-app https://github.com/Kitt-AI/snowboy

Yes, activation is local. You can see this in the documentation for making your own Alexa device. [0] To my knowledge all the cloud-based systems use something like this for both privacy and technical reasons. Sensory includes a chart showing the CPU requirements for wake word activation and accuracy. [1] The main reason custom names aren't allowed is accuracy and training time. Even with a well-trained model like Sensory's using 90 MIPS the wakeword isn't recognized 10.2% of the time and has false acceptance. Using a set wakeword could also allow for a FPGA or custom chip to have the wakeword model hard-coded for a much lower power hit for mobile applications. One of the other wakeword solutions, Kitt-AI Snowboy [2] allows for personal hotwords, but it takes some effort to make a personal training set and this isn't nearly as robust as w