Very impressive, no question.

Interesting and impressive; particularly that it also seems to be able to equalise the voice sensibly, removing some really quite severe microphone peaks. Some of this process is often already done in mobile 'phone speech compression, where the audio is analysed/deconstructed using fast Fourier transform and reconstructed at the far end, but the primary goal of that is data compression, not audio quality improvement.

The human voice comprises a fundamental frequency and overtones, all of which are modulated by the vocal tract, so all of the components of the voice, apart from plosives and sibilance, are harmonically related to that fundamental. If it isn't related, it isn't the voice. The fundamental also generally identifies whether the voice is male or female, which must help with the EQ part of the process. Extraneous noise removal, reverb removal, voice characterisation and equalisation, possibly plus a few other tricks.

They seem to have made those various bits work together rather well, at least for a white American male.

I imagine it will also manage female voice and voices from other regions with different sound characteristics. I wonder how well and what languages.