add third_party hotwording example

[proppy]

It would be nice to have an example showing integration between https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/micro/examples/micro_speech and google-assistant-grpc

This sample could show:

• how to run the micro_speech in same process as the assistant
• how to communicate with micro_speech running on a MCU (ex: GPIO interrupt)
• how to customize the model with different keywords

[proppy]

in https://github.com/googlesamples/assistant-sdk-python/issues/356#issuecomment-595132272 @BlackLight suggested looking at https://github.com/mozilla/DeepSpeech.

Generalizing the scope of this issue.

[TheSpaceGod]

@proppy do you know if there is any timeline for putting together this example by chance? I am working on a pet project where this would be very helpful. Thanks!

[proppy]

No works has started around this in this particular repository, but @BlackLight has been recently blogging about how he integrated multiple hotwording model in his https://github.com/BlackLight/platypush/ project.

[Shulyaka]

The Mycroft-precise project has a decent hotword detection. With about 500 samples of me saying 'ok google' and about 5k samples of other sounds, I've been able to train the model to a very usable level. It also responds to my only voice.