That sounds awkward.

The tech for isolating a speaker at conversational distances exists. You use half a dozen microphone transducers (minimum; Crappy microphone transducers are cheap and quality is expensive, so just use a bunch of them), and through a combination of phase and intensity they decode relative location, and amplify that phase expectation while suppressing everything that isn't phased like that. Sound is slow, and readily susceptible to real-time triangulation. The math/processing is much easier if the parallaxes are fixed (eg the microphones are arranged in a line array on the top band of a rigid pair of smart glasses), but with a little latency it's not prohibitive for a deformable array to solve for its own relative position as well.

Yes. This is what my hearing aids do. Three microphones on each side, an accelerometer, and they communicate with each other. I have a directional cone in front of me. The sounds behind me are significantly muted. In lecture mode, the most prominent voice is localized in space and everything else muted. I can have people talking fairly loudly behind me and I'm unaware of it, in that mode.

But it's not magic. It provides maybe 5 to 10 dB of SNR improvement? Just like the machine-learning based noise suppression. I'd estimate another 5 at most 10 dB there. (Manufacturer claims 12 dB. They're optimistic.)

That's a meaningfully helpful improvement. But I'm still mostly lost in noisy environments.