Set up a network gateway once and wiresharked it. Echos don’t transmit anything that is said when the blue light isn’t on. You’re right that the voice prompts are being data mined to hell and back, but your day to day conversations aren’t.
Where I find particular concern isn’t the data molining of conversations, but rather the fact that any cloud connected device can serve to update whoever they decide what your home’s current WAN IP address is, which can be used along with cookies or apps on your phone to tell when it has joined a particular wifi network, thereby knowing who is in what house when. Oh, Jill spends a lot of time in the same house as Bob, but only stays over night half the week? Oh, Jill is doing a lot of searches for stuff pregnant women search for? Better start flooding Bob with ads for engagement rings!
I’m legit surprised more isn’t being done with ultrasonics/infrasonics in the exact areas you describe.
There was one project I worked on to help spec-out and design a device location system at a hospital and one of the solutions we vetted actually used low power ultrasonic exciters producing known and unique patterns paired with tiny little microphone equipped pucks with just enough smarts to record the relative volumes of multiple exciters and send that via radio to a central server along with its own unique ID to be used for triangulation. Ended up being too expensive for us and we went with something based on a combination of BLE and LoRA, but holy shit was the sound based stuff cool (and wouldn’t crowd the radio environment nearly so much, which can be a big deal in ultra dense environments like a hospital.)