What makes you think every human hears the same thing, if i had a hearing loss between 3khz and 5khz and the Psychoacoustic model used in the codec thinks that a tone of 4khz would overshadow the 2khz tone in a specific music file the psychoacoustic model is flawed in this case. So if i heard the source i would hear the 2khz better then the 4khz one, in the compressed file the codec filtered out the 2khz tone making my experience different from the source.
|But with the pictures you can't see wich 128kbps encoding of different codecs (without parametric stereo and Spectral band replication) is better sounding to you, you only see how accuratly it reproduces the original sound. But with the 256kbps WMA 10 Pro graph you can clearly see that it is WAY better then the MP3 at 256kbps, if someone could hear the difference is something else, and thats why spectrographs are more accurate then listening tests in my opinion.|
Your paradigm of analysation doesn't allow for this sort of analysis, though, despite it being a very realistic and sensible approach when taking into account psychoacoustics, because your analysis works off of a superficial impression of the data with a particular systemic bias towards how it will be intepreted (which essentially amounts to it looking good to you because it superficially appears more similar to the source sonogram). That, sir, is not a scientific approach; it's pseudo-science and now we've several other posters who are committing the same fallacy in their own evaluation. It is entirely scientific to claim that you need ABX listening tests to evaluate codec performance, because the ultimate basis for evaluation of a codec is qualitative and whether the qualitative loss is minimised; that is the objective of psychoacoustics and lossy audio compression. The object of lossy compression is minimum qualitative loss with maximum quantitative loss. Psy models are tuned with listening tests. After listening tests, you may have data which you can correlate with some measured data for making scientific models and predictions, but you still need the listening test data in order to competently evaluate the efficacy of your psy model. Otherwise, again, you cannot evaluate the qualitative performance, which is a problem since that is precisely what you need to maximise. Thus, your sonograms and their superficial appearance do not host the totality of relevant information or background for making competent evaluations of codecs. Insisting that they do is what is the pseudo-science and conjecture here, and it's unfortunate that this sort of thing is taken prima facie as the basis upon which to choose a codec.
hydrocity - One sonogram with a superficial appearance closer to the other doesn't necessarily ensure maximisation of qualitative performance, particularly when dealing with higher compression ratios. If you're having trouble ABX'ing a compressed file against the source file, that means it's likely you are unable to perceptually distinguish them, and as such the codec is doing its job of minimising qualitative loss.
Note: I'm not saying the graphs aren't interesting; they are. I simply disagree with your methodology for analysing the performance of a codec. The information can still be useful, however, when placed against a backdrop of other data and compression theory.