You could test your local drain hypothesis pretty easily. Turn on voice activation of e.g. Google on an Android. This means your device is now listening to and

You could test your local drain hypothesis pretty easily. Turn on voice activation of e.g. Google on an Android. This means your device is now listening to and analyzing every word to see if it ever recognizes 'Okay Google'. If your theory is correct you'd notice a substantial battery drain compared to having this functionality disabled. I'm going to guess you will notice basically nothing. The nice thing about neural network driven systems is that while they are very resource intensive to build, once built they are (generally) extremely resource friendly. And training a network to recognize one of a wide array of terms would be no more difficult than training one to recognize Okay Google. And the best thing is that false hits don't really matter that much, if you assume the only motivation is advertising. Though, for what it's worth, my bet would be on exfiltration of the data instead of local analysis. Put the shoe on the other foot. Imagine your goal is to exfiltrate data from a device in a covert fashion. You can come up with countless clever ways starting at the most low level with extreme compression. All that matters is that what is said can be recovered - it doesn't need to sound pleasant, or even intelligible as long as another system can reconstruct the speech (even if at great computational expense) from what you send. And from there you can come up with all sorts of clever ideas such as only sending the data during other normal transmissions.