I'm ABXing different file types to determine what audio best suits my needs, but I'm confused about some things. I've read the ABX FAQ, as well as the Wiki page, but I still have a couple questions. I've experimented with WinABX as well as fb2k's ABX plugin.

How do I determine what sections of the song I should listen to and how long the clips should be? Should I listen to the same part of the song multiple times for the same test, or "randomly" select different parts myself?

Should I listen to the entire clip, or switch back and forth between the two (WinABX refers to this as a "quick switch")?

When should I stop trying? I know 13/16 correct tests is a general consensus for giving a good margin of error, but if I'm at 12/16, how far should I go before I just give up?

Is normalizing the audio good/necessary? I know the fb2k plugin requires RG. Does WinABX automatically do this for me?

Thanks for your answers. I appreciate your help. I'm sorry if these questions have been answered previously.

There's no fixed rule for this, just use what works (or you think that works) best for you. Usually, best results are achieved focusing in a short part of the song that could seem to sound different. Sometimes it can be difficult or impossible to find it. But, as I said, there's no fixed rule, and I guess many people will have a different preferred technique.

But, if you feel you do well at the beginning and bad at the end, then you must rest and let your ears "freshen" when they get tired.


Getting a probability of being guessing (p) < 1% is quite enough, if as you say the the number of trials is nof fixed, that is, you keep trying until you get a good enough result. That would be true up to nearly 100 trials.



It's not neccesary if the clips are already level-matched and there is no clipping in any othe wav test files. If there's clipping, RG is a way to avoid it. As to level matching, RG can do a good job here, but I think there may be some extreme cases where it may not do a good enough job. However, at lossy codecs testing, it will work just fine.



No it doesn't. WinABX doesn't touch the test files, they must be already "prepared" for the test. And, AFAIK RG is not required at the fb2k plugin, it can be disabled at its preferences page.

Actually, I don't think volume correction is all that important. The differences between encoded files and the original are usually below 1dB, which I find quite hard to ABX (although others might be more sensitive for this). Still, it's not a bad idea to match the volume of the samples if you have the option, because if one sample happens to be significantly louder it is often (unfairly) perceived as being of higher quality.
In my opinion, offset correction is more important than volume correction, because offsets of about 30ms are not uncommon and can be blatantly obvious if playback begins with a sharp impulse.

I think the abc/hr approach would be better suited for your demands.
With abx you only know whether you hear a difference or not, but you can not determine what sounds better. With abc/hr you get to rank the samples you are testing in a blind fashion. In case you are having trouble ranking the samples, you can still do individual abxing inside of the abc/hr application.

I just listen to the file until i hear a spot that seems like it might be worth testing for some reason. Then I mark it and loop a 2 or 3 second patch a few times to see if theres anything there.

Basically its trial and error; you prospect along trying whatever looks promising and moving on if theres nothing there.

Usually, codecs don't change volume of the compressed files at all, so correction would not be needed. However, with loud music and high compression, clipping can be an issue to take into account.

Apart from that, 1 dB level difference is easily ABXable. It seems that level differences as low as 0.3 dB have been ABXed.

Also, keep in mind that the ABR/CBR --alt-presets include --scale (93 for 128kbps AFAIK), which could cause an additional effect on ABXing.

dev0

93% is 0.6 dB softer. This is most likely audible and therefore RG should be used in this case and similar ones.