You could easily process the audio and filter in on the device and just send the transcribed parts that match your filters for further processing. No need to send a constant audio stream. Eventually it would show up in the battery usage but things like the google song recognition also heavily use the mic and process audio and it doesn’t show up as heavy as you would expect.

Yes exactly and some providers also accept crypto.

As I read, they used timing analysis which should be preventable by using an anonymous VPN to connect to tor and streaming something over the VPN connection at the same time. Some of them support multi-hop, like mullvad, which will further complicate the timing analysis because of the aggregated traffic.