What the ear cannot hear#
Week 2 ended with a proposed explanation. This hour is about the difference between proposing one and testing it.
Loss designed to be inaudible#
Perceptual coding does not try to keep the signal. It keeps the parts a model of the listener says will be noticed, and discards the rest.
From Jonathan Sterne, MP3: The Meaning of a Format (2012), chapter 1.
A codec is an argument about hearing, built into a file format.
A comparison that cannot answer it#
One encoding cycle at 320 kbps: deviation score 0.3. Ten encoding cycles at 96 kbps: 9.8. Conclusion: ten encoding cycles cause a large loss.
The numbers are fine. The conclusion does not follow.
Two things changed#
Cycle count: 1 → 10.
Bitrate: 320 → 96.
There is no way to divide the 9.5-point difference between them. This is not a subtle error — it is everywhere, because the two conditions were probably never designed as a comparison at all.
The repair#
Change one thing. Hold everything else fixed.
One cycle at 128 kbps versus ten cycles at 128 kbps.
Bitrate constant; only the cycle count differs. Wednesday’s data: 3.2 points — roughly a third of what the confounded comparison implied.
Four things a protocol must state#
- The independent variable
- The outcome measure
- The control
- What was held fixed
Leave any one out and a reader cannot tell what your result is about.
The number that matters most#
You are reporting a difference between two conditions, so what governs you is how much that difference moves when the whole procedure repeats.
For us: 0.5 points.
A non-detection is a finding#
Run the corrected comparison at 320 kbps: 0.4 points.
That is at or below repeatability — a non-detection, not a small difference. Reporting it as degradation would be reporting your own noise.
“We did not detect a difference” and “there is no difference” are different claims. Only the first one is yours.
Where that leaves the question#
At 128 kbps: detectable. At 320 kbps: not.
A worse headline than “generation loss ruins everything,” and a better answer — conditional on a stated source, score, setting, scenario. Not a claim about MP3 in general.
The control, again#
Ten copy operations, no encoder: scores zero.
A clean control does not prove the procedure is sound. It rules out one specific failure — that the procedure manufactures differences regardless of treatment. Run it first.
Before Wednesday#
Read the comparison you’ve been handed and the measure definition. Compute nothing yet.
Workshop: What the ear cannot hear
Text walkthrough#
For a reader who cannot use the deck view: perceptual audio coding discards whatever a model of the listener predicts will go unnoticed, so a codec encodes a theory about hearing. A supplied comparison claims ten re-encoding cycles at 96 kbps cause a 9.5-point-worse score than one cycle at 320 kbps, but the comparison changes two variables at once — the cycle count and the bitrate — so the difference cannot be attributed to either one. Correcting it by holding the bitrate fixed at 128 kbps and varying only the cycle count gives a smaller, attributable difference of 3.2 points. The repeatability of that kind of difference, re-measured, is 0.5 points, which matters because running the same corrected comparison at 320 kbps gives 0.4 points — a non-detection, meaning this procedure cannot tell whether the two conditions differ at all, not evidence that they are identical. A protocol has to state its independent variable, outcome measure, control and what was held fixed, or a reader cannot tell what the result is actually about; the exact-digital-copy control, run through the same number of operations and scoring zero, checks the apparatus rather than proving the procedure sound.