I used to run forensic network capture and analysis tools.

First thing, traffic is encrypted. All you will see is a blob of traffic passing through. You used to see hostnames with TLS, but now with quic, you see nothing. This makes it hard.

You could root the phone and install a root ca certificate for a decrypting proxy, you might see more, but the data itself (not just the transport protocol) could be encoded or even encrypted within the network encapsulation.

Next, you'd have to reverse engineer the protocol if they're using something nonstandard. Also, malware can often be set up to "behave" when it can detect analysis. I'm all but certain Google would do this.

Maybe you could do statistical analysis of the traffic and attempt to baseline normal vs when it's transmitting audio. It would be a bit of a blind guess at best.

If I had more time, I'd love to try it. I have an old pixel7 pro. Maybe I can sort something out.

If its real time monitoring you, but not if its logging data to send later when it would be expected to be doing so.

Audio doesnt take up much space.

The consensus so far seems to be that they don't collect as much data as people think, partly because they can't process all of it, and partly because educated guesses are good enough to target ads often enough.