p.s. Before somebody better comes along and blasts me as incompetent, I should note that I have not been actively designing data converters for the past few years though I have followed them somewhat, but there may be breakthroughs I missed. The fundamentals have not changed AFAIK. I have in the past designed a number of data converters, ADCs and DACs, using a variety of architectures including segmented R2R and binary converters, slope and dual-slope, flash and folded-flash, successive approximation, delta and delta-sigma, phase ADCs, ADCs using Hadamard sequences and other interesting tricks, etc. Most of these were at much higher rates than audio, however, so my experience is limited to ~16-18 bits (and if you think 16 bits at 20 kHz is hard, try doing it at 100 MHz, or several GHz).
p.p.s. A signed-magnitude converter implemented the way they describe, using separate sections for (+) and (-) sides, means the two sides have to be very precisely aligned in level and time to keep glitches from occuring at the crossover. That may account for the higher distortion, although other factors like the output buffers and layout are likely also major players.