I got twigged to this new app via one of rec.audio.high-end's endless debates on ABX. It purports to automate the process of level-matching, aligning, and subtracting two recordings, allowing relatively easy 'difference signal' extraction (what Cool Edit/Audition users have long called 'invert/mixpaste'). Haven't tried it yet, but thought people here might be interested.

http://libinst.com/Audio%20DiffMaker.htm

some of the tests the author suggests are suitable for the Diffmaker :

Mentioned at The Audio Critic as an alternate to ABX in finding the differences between two 'all else being equal' setups:

A New, and in Some Ways Preferable, A/B Comparison Technique

Of course it won't be accepted by subjectivists. cf this comment

Yes, I saw the AC article too, but figured it requires subscription to access.

Thanks for the interesting link.
I've always liked the difference signal test method as one of many tools to analyse audio. Almost a year ago the validity of the test has been discussed in this HA thread.
It is especially useful to quantify very small differences. Most people e.g. agree that there is no point in ABX'ing lossless codecs, since the audible difference is assumed to be non-existent (the difference signal will most likely be the noise of the playback system).
If one would compare a 24/192 and a 24/96 (apart from sample rate identical) signal, and the difference signal is inaudible, I would feel safe to say that both signals sound identical.

I can hardly wait to have this tool available for OSX

It's hard to believe why so many people have tremendous comprehensive problems with understanding the meaning of the word "lossless". The "audible difference is assumed to be non-existant"? THERE IS NO DIFFERENCE AT ALL! It's a mathematical fact.

Not anymore. =) I was subscribed and I also bought the back issues when before they were available online in PDF form. Then I go to the site one day and everything is free.



Imagine explaining to someone that and MD5 sum is just as good as a diff check between two binaries.

I don't really have an interest in checking the differences between codecs, but the idea of recording the output of various equipment then checking the differences with this tool is intriguing. But I'm a Mac user, so I'm with Kees de Visser here .

You're absolutely right that there is no data difference. What you seem to forget however is that we don't listen to the data directly but to the output of a DA converter. The analog output of the DAC suffers from additional noise and distortion (be it at a very low level). So even with identical data the analog signal will be (hopefully inaudibly) different on every playback. These differences will show up in an analog difference test, not in a bit-compare test. From a psycho-acoustical point of view it is very interesting to compare acoustical stimuli instead of data.

Example: there have been numerous reports of presumed audible differences between data-identical cd's. Waving this claim by proving that the data are identical and the cd's therefore can not sound different is the commonly accepted answer (which makes sense IMO). With the difference signal technique it becomes possible to verify the DAC's output and check if the difference signal is large enough to make an audible difference plausible. This test should be equally suited to check other subtle changes like jitter effects on DAC performance.

Making an audio diff still doesn't answer if said differences are audible when played in their original context.

Then again, getting a certain population to accept masking will never happen, thus even this technique won't hold sway.

...At which point you accept that the input data has nothing to do with the difference, so it doesn't matter even mentioning it.

More specifically: Don't mix what you're comparing.


Now, about the "software", it can only prove bit-perfect copies, it doesn't solve anything else:
If there isn't a difference, it isn't ABXable. If it is ABXable, there is a difference. If there is a difference and is not ABXable, it is inaudible. You can only use the difference signal to train you to hear it on the real (not diff'ed) ABX session.

Agreed.
The proof of zero difference is however limited to digital signals, so the variable under test should be in the digital domain.
Analog signals can't be perfectly identical. To test audibility I'm afraid the only place is in the analog domain (e.g. for digital sources: after a DAC). Unfortunately the "monitoring variables" like noise and distortion of the DAC, amps, speakers, quality of listener(s) etc. will be included in the test.
Absolutely. I can't think of any exception.
Sure, "under the given test conditions". Perhaps the results would have been different with better equipment, better (trained) ears, different audio sources etc. ABX doesn't show differences below the level of audiblilty. There are many situations where one might want to quantify the inaudibility. A typical audio chain consists of many cascaded devices from microphone to speaker. Cascading ABX-transparent devices doesn't guarantee transparency of the whole chain. If the total chain is not known it can make sense (e.g. for a mic pre-amp designer) to keep undesired elements below the audibility threshold as far as possible. The difference test might give some insight. It can not replace double blind listening tests to prove audible differences.

So, has anyone tried the app yet?

I've tried it for about 10 min. and found an unexpected difference when comparing a 24-bit file and its 16-bit dithered version. The difference should be noise only, but for some reason, DiffMaker produces some audible music with it. Perhaps the synching algorithm isn't perfect, or I've missed an important setting. More time required to investigate what goes wrong (but after my vacation)

Correction: I've heard the problem when playing the difference in the application directly. The difference signal is also saved to disc and this file seems OK.

One issue to bear in mind is that this technique will make phase distortions very apparent. For example if you reverse the absolute phase of signal B, A-B -> A+B and the resulting "difference" signal will be large if A and B were similar or identical to begin with. This is relevant for people trying to use this tool to compare lossy encoders, or any kind of analogue filter, which may induce phase distortions (which are usually inaudible).

Of course this doesn't invalidate the statement that if A-B is inaudible, A and B are indistinguishable, but it's worth remembering.