Thanks for sharing this, @OldMusicGuy . Your observation in bold above is probably the difference between using beamforming to suppress the background noise and using the DNN to suppress background noise.
There is no control on how much background noise can be suppressed in beamforming. The suppression by beamforming is just fixed and aggressive by the nature of the beamforming technology itself. But with the DNN, the suppression can be more flexible and selective between sound components. The DNN doesn’t “block out” the background noise like the beamforming does. It just reduces the volume level of the background noise relative to the volume level of the speeches (the term that they use is “rebalancing”). That way, they can still allow in background noise, albeit not at an intrusive level anymore, but at a subdued level that nevertheless still has a presence.
I don’t know this for a fact, but I see a possibility they can probably also play game to elevate the volume of the speech more than what it really is at, just so that they don’t have to subdue the background noise so much as to lose its presence altogether (like with beamforming), in a situation where the contrast level might have to subdue the noise to almost nothing if the speech volume were to be kept fixed as is.
With a DNN implementation, there’s a lot of flexibility in how a sound scene can be rebalanced, yet still retain the openness of the open paradigm.