​

What do you suppose I do with those feature / data sets? Since "OK google" responds to my voice independently of the rate of speech, methinks they are using a combination of regression analysis and discrete time warping. 

But, it's seemingly both speaker and pitch independent too, so there must be something else going on. There's no way they implemented a full Hidden Markov Model inside the phone's DSP, (it wouldn't make sense for just one hotword).

Thoughts?