I believe him, although his argument is rather weak. It's true that it would be prohibitive to run speech recognition at all times, but what if Google would choose to do so when detecting a particular pattern, like you are meeting a specific person? Even simpler, we could imagine it tries to detect speech for 5 seconds every 10 minutes; that way it could bypass a lot of the arguments listed here. We also know that the NSA can activate your microphone remotely, so the technical capability is definitely there and Google has the means to do it.
One thing that can happen is you talk about something and the other people start looking it up. And then because you’re on the same network and use the same external IP address, you get ads related to their activity, which is related to what you were talking about.
Ofc they are not spying on my conversations, they are just collecting thousands of data points on me so they can sell ads.