You need real white noise - not very loud, but it has to be random as in not predictable.
https://www.npr.org/templates/story/story.php?storyId=956894...
"And to prevent sophisticated laser devices from capturing the telltale vibration of his voice on the glass, music played between the panes."
Then you'd have to account for stuff like building sway.
With researchers like Mordechai Guri out there, I'd say the GP comment is solid general advice.
Take small microphone element that captures human speaking range vibrators and connect it to equally small retroreflector. Then you can read the sound from any angle using laser/sensor next to each other. There would be no electrical or radio frequency transmissions at all.