He's asking the opposite - remove the bits with no speech.

I used to do more or less this same thing manually, years ago, for a corporate vlog in which people drove around all day fixing pipe leaks and occasionally commented on what they'd done - they wanted the clips where they commented, edited together.

I basically just looked at the audio waveform and figured out where I should probably cut, and then listened to it to narrow it down.

If someone hasn't already trained an AI for this, they should.