Audio Blind Test (ABX)
Can you really hear the difference? Compare an original clip with a lower bit depth, a treble cut above 16 to 20 kHz, or a tiny volume change, without knowing which is which. Scored with an exact p-value. Runs in your browser, no files to download.
A is the original. B is the same clip reduced to 10-bit with TPDF dither. X is secretly A or B, picked at random each trial. To pass 10 trials you need 9 correct (p < 0.05).
Starts quiet on purpose. Raise it slowly to a comfortable level, then leave it alone for the whole test.
How the ABX blind test works
Why sighted listening fools everyone
When you know which version is playing, your brain fills in the difference you expect to hear. This expectation bias is why expensive cables and "hi-res" badges sound better in a shop. The other trap is loudness: a version that is even 0.5 dB louder tends to sound clearer and fuller, so an unmatched comparison is rigged before you start. Try the loudness challenge to hear how small that nudge is.
ABX removes both. A and B are labelled, X is one of them at random, and you only score by naming X correctly. All three versions are the same clip at the same position, synthesised in your browser, so switching is instant and nothing gives the answer away except the sound itself.
What the science says about hi-res audio
16-bit audio already gives about 96 dB of dynamic range, more than a quiet room and a safe listening level need, and 44.1 kHz captures everything up to 22 kHz. Most adults cannot hear above 16 to 18 kHz anyway; check yours with the hearing test. Controlled blind studies mostly find small or no audible differences for playback. A 2007 study by Meyer and Moran found listeners could not detect a CD-quality bottleneck, while a 2016 meta-analysis by Reiss found a small effect, mostly in trained listeners. 24-bit and high sample rates are genuinely useful when recording and mixing. As a playback upgrade, the claims usually outrun the evidence. Lossy codecs often cut content above 16 kHz too, so the low-pass challenge is a rough guide to that part of the lossless vs lossy debate.
Reading the p-value
The p-value is the exact chance of scoring at least as well by guessing. Below 0.05 means that would happen less than 1 time in 20, so you probably heard something real. With 10 trials you need 9 right (p = 0.011), as 8 of 10 still gives p = 0.055. With 16 trials, 12 right passes (p = 0.038). A fail does not prove there is no difference, only that this run did not show one. Retrying until you pass cheats the maths.
Tips for a fair test
Use decent wired headphones; laptop speakers roll off long before 16 kHz, which the frequency response test will show. Set the volume once and keep it. Loop a short passage, such as the cymbal tail at the end, and switch quickly between A, B and X, since auditory memory for fine detail fades within seconds. Take a break if you start to guess.
Related tools
Check your gear with the headphone test and the frequency response test, find your own limits with the hearing test, and check whether a voice clip is synthetic with the AI voice detector.
Frequently asked questions
What is an ABX test?
An ABX test checks whether you can tell two versions of a sound apart. A is the original clip, made in your browser. B is that same clip with one change you picked. X plays either A or B. You can listen to all three, then say whether X was A or B. Only those calls are scored.
How is X chosen?
Each trial picks X at random from A or B, with an equal chance of either. The X button only says Mystery. The pick is not written into a label or anywhere else on the page before you answer. "X is A" and "X is B" are checked against that pick, not against a fixed file. The next trial picks again.
What does the score mean?
The score is how many trials you named correctly, out of 10 or 16. Guessing is right half the time, so chance is 50%. The result also shows an exact one-sided p-value: the chance a fair coin would get at least that many correct in that many flips. You pass only when that p-value is under 0.05, which takes 9 of 10 (p = 0.011) or 12 of 16 (p = 0.038). Eight of 10 is p = 0.055 and does not pass. A miss does not prove the two versions sound the same. Accuracy uses the 10 or 16 trials, so the finished score is never divided by zero.
Which clip does each button play?
A always plays the original. B always plays the modified copy, and the buttons are not swapped. Bit depth reduces B to 8, 10, or 12 bits, or quantizes it to 16-bit from the 32-bit clip, with TPDF dither on or off. The treble challenge removes everything above 16, 18, or 20 kHz. Loudness plays B at 1, 0.5, or 0.2 dB quieter. X plays whichever of A or B was picked for this trial, at the same position in the clip.
Can I change the challenge while a trial is running?
No. During the test, challenge, difficulty, dither, and the trial count are ignored, so A and B stay as they were when you pressed start. Volume and loop still apply to every button the same way. Cancel drops the score and returns you to setup.
Does the audio stop when I leave?
Yes. Leaving the page stops the clip that is playing and closes the audio context, so the sound does not keep going in the background.