If I understand this right, 16-bit data will be converted to float, mix, dither, and then converted back to integers. This happens even when bit-depth and sample rates match, as long as the mixer is not bypassed. Since dither is imperfect, 16-bit data will have audible noise, while 24-bit noise...