Facing the Codec Challenge Page 3

Lessons Learned The score for each trial shown in our graphs is the score each listener gave to the codec-processed version minus the score he gave to the "hidden" reference, because what matters is the difference between them. Theoretically, the scores could run from +4 to -4. But a tabulated score would be positive only if the listener gave a perfect score to the codec version - that is, confused it with the reference copy, usually because he couldn't tell them apart! As you can see, only a single score is more than +1, and the vast majority are negative. How far below the zero point each colored bar falls indicates how much sonic degradation the listener perceived in the compressed version. What do these graphs tell us about codec performance?

Codec abilities vary with different kinds of signals. At the higher bit rates, all three codecs did well with at least one musical excerpt, with the results for one track clustered close to zero. But for each codec there was also at least one selection on which it didn't do so well. Applause (Track 1 in the graphs) posed problems for all the codecs. If you listen to a compressed recording of a live concert, of any type of music, you are likely to encounter unnatural-sounding applause at some point. Based on these results, I now consider applause a more relevant test of codec quality than solo harpsichord (Track 8), much less solo castanets (Track 4), two kinds of sounds codecs have had difficulty handling in the past. Olé.

Listeners vary in their responses to codec degradation. The same comparison could elicit varying responses even though the scoring was similar. Although all three listeners disliked the way Real/64 reproduced the applause track, Red gave it a score of -2.5 and commented that it sounded "like a swimming-under-water effect, [with] wavering, muffled highs, [image] width closes in." It wasn't "as bad as a digital answering machine," he continued, "but nothing I'd want to listen to by choice." Blue, who gave it a poorer score of -3, characterized it only as "swooshy/ fizzy, dull," while Orange, who gave it a worst-case score of -4, called attention to its "mono-ish sound" quality.

At other times, one or two of our listeners seemed to have caught something the others didn't. Orange, for example, thought that Real/132 and Real/64, as well as MP3/128, did a pretty bad job reproducing the segment from Pearl Jam's "Daughter" (Track 5), a certified codec killer in our round of tests two years ago. His comments included "muffled" for Real/64 and "smeared cymbals" for the others. In contrast, both Red and Blue thought these three codecs came much closer to the original in the Pearl Jam comparison.

By looking at one color at a time, you could say that listeners Red and Orange were both more critical than Blue, giving scores consistently lower than Blue's for each track and having fewer positive scores. Orange had only two positive scores. I'm sure that most untrained "man off the street" listeners would do much worse than our three with most of these tracks.


ARTICLE CONTENTS

X