A Pro I Know
One pro I've worked with is K.K. Proffitt, Chief Engineer of Nashville's JamSync recording studios and contributing editor to Surround Professional magazine:

"There's been a big change in the last four or five years," she explains. "In the old days, we would always try to get the signal-to-noise ration right up there. Because in the old days what you were using was really PCM-encoded (pulse code modulation) audio and people would be downloading the files themselves - it really wouldn't be streaming. So the trick was with limited bandwidth get the signal-to-noise as high as you could so you could pull things down to, say, 8 bits. You would use dither to get as much as you could out of that last little bit and there would be different kinds of dithering algorithms depending on whether you expected the resultant audio to be used in another form. I don't know who would use 8 bits again, but there are people who did it. Basically, with the compression algorithms for streaming technology that use masking and perceptually coded program material, you're dealing with an entirely different issue. Because they use adjacent bands that mask each other, with important material in certain bands being preserved while those containing less important perceptual material being thrown away, the problem becomes entirely different. You really need dynamic range in order for the perceptual encoding algorithm to work properly. So we've had to make an about-face here and start looking at what is our peak-to-average and how much should we have, and how well will that stream. We also have to look at if you're going to go with an encoding algorithm for streaming or you want to downstream a 16-bit recording to someone who's going to make a CD from it and have it be "almost CD quality." You then have to really think about how much noise-shaping, what kind of noise-shaping, what kind of dither. I have several different approaches to dither, and on any different kind of program material I will simply do the old-fashioned thing of listening - I'm not a prognosticator of what will work with certain different types of program material at this point for stream. I really actually listen to it."