Artifact Training Page - analyzing the results

2011-02-10 23:35:57

I did some ABX testing using foobar2000 and the samples from ff123's Artifact Training Page. I managed to get 0.1% probability that I'm guessing on all of them except Sarah McLachlan's Wait (there were 3 of us who did the test and only one heard the difference here). Now I'm trying to analyze the artifacts that I have heard. I view the files in Audacity and Sonic Visualizer and look what differences correspond to the artifacts. But I can't see what caused the distortions in 41_30sec_scmpx128.mp3 and left-dist_q0.ogg - the waveforms look quite similar. On most of the other samples the differences where clearly visible or became clearly visible after applying a high-pass filter, but I don't know how to see the difference in these two. Any hints?

Artifact Training Page - analyzing the results

Reply #1 – 2011-07-27 12:13:11

Depending on the specific type of artifcact you're talking about, there may be simple example waveforms that illustrate the effect, but I'm not versed enough to tell you what's what in that respect.

It's good that you try to understand how lossy compression affects the waveform, but as a (maybe premature) warning: don't conflate how it looks with how it sounds, and try not to jump to conclusions based on a picture of a waveform or spectrogram.

The audibility of an artifact does not correspond 1:1 with it visibility. Some artifacts are much more visible than audible, and some easily-heard artifacts are hard to see. Visual inspection of a waveform does not result in an accurate idea of how it sounds. For example, you can chop a tiny bit off  the tops of a few select peaks. This is clearly visible, but not even remotely audible.

Artifact Training Page - analyzing the results

Reply #2 – 2011-07-27 14:42:34

+1. Trying to hear with your eyes is the source of much audio hokum.