ABX Blind Test: Can You Hear the Difference?

A free online ABX blind test - you hear A, you hear B, then X is one of them at random and you say which. Sample-accurate and double-blind, in your browser, with the playback position carrying across every switch - the closest thing to "is this thing actually different" you can do without a measurement rig. Try a built-in synthetic test in two clicks, or drag in your own pair: the tool checks what the two files really differ in before the first round, so the score can be read against it. Your files can be DSD as well as PCM, and a blind preference mode asks which of the two you like better.

My tests

Completed tests save here automatically. Click any saved test to re-open its result and certificate.

Pick what to compare

Source

Classic questions

Check the test itself

Score - Take the test to see your number.
Could it be luck? - The p-value: probability your hits came from random guessing.
Confidence No result yet Four levels: not significant · significant · strong · very strong.

How ABX testing works

The protocol

ABX is the gold standard for auditory discrimination. You get two known references (A and B) and an unknown X, randomly assigned to be A or B each round. Audition each as many times as you like, then commit an answer.

Sample-accurate switching matters. The position carries across A→B→X in this tool, so you swap mid-phrase. Your auditory echoic memory is roughly 3-4 seconds; if the switch resets the playhead, the gap erases the trace you were comparing to.

A hard cut between two different waveforms clicks, and a cut from a clip into itself does not, so a plain switch from A to X would tell you what X is by the click alone. Here every switch fades the outgoing clip out over 5 ms and the incoming one in over the next 5 ms, the same way whichever clip comes next.

The statistics

The p-value is the probability of getting at least your hit count from pure random guessing on a binomial distribution. The conventional threshold for ruling out guessing is p < 0.05, and it holds for a round count fixed before the first answer: stopping when a run looks good, or repeating runs until one passes, lets chance results through. That is why the tool shows the p-value only once every round you chose is answered.

Longer tests need more hits but a lower hit rate to clear that bar. The table below lists the scores that do and the p-value each one carries.

What differs between your files

What the check measures

Before the first round, the tool compares the two files themselves. It finds an offset between them (a render that starts a few milliseconds late) and lines them up, measures each file's level over the stretch both cover, and subtracts one from the other after matching. What is left of that difference, the null, says how much of the waveform the two share.

It also compares their long-term spectra in third-octave bands, because a null alone misleads: two strips of the same noise share no sample at all, so their null is as large as a null gets, yet they sound alike. The check reports what differs, by how much and where. Whether you can hear it is what the rounds are for.

Two cases it settles

Identical as played. Two renders that differ only in gain, or only in when they start, are the same audio once matched and lined up. There is nothing to hear, so a score near half is the right result, and the result and the certificate say so. Running a pair like that on purpose is a control: a high score on identical clips points to a cue outside the sound.

Same spectrum, different waveform. Two takes of the same noise, or a reverb tail with a different random seed, differ in every sample and match in spectrum and level. On a long loop the ear has little to hold on to. Loop half a second instead: heard on repeat, a short stretch of noise takes on a pattern you can compare.

Two masters, hi-res and DSD files

Two masters of one mix

Of two versions of a song, the louder one tends to sound better on first hearing. Level matching takes that away: the tool measures each file's integrated loudness in LUFS, the ITU-R BS.1770 measure Spotify and YouTube normalize tracks to, and turns the louder one down to match. What is left is what the mastering did to the music.

The file check also gives each version's peak-to-loudness ratio, how far its peaks rise above its loudness. Compression and limiting shrink it, so the gap between two masters' figures is the plainest number for how much harder one was squeezed. Run an ABX test to see whether you can tell them apart, then the blind preference test to see which one you choose without knowing which is which.

What reaches your DAC

A browser plays at the rate your operating system has set for the output, often 48 kHz, and resamples every file to it. Both files take that path, so it favors neither, but a 192 kHz file on a 48 kHz output loses everything above 24 kHz before the DAC sees it. The file check names the rate and where to change it: set the output to the files' rate and reload the page.

DSD files, DSF or DFF from DSD64 to DSD512, are converted to PCM in your browser, as a player does when it sends DSD to a PCM input: a lowpass filter flat to 30 kHz (40 kHz when the output runs above 100 kHz) removes the noise DSD carries above the audio band. A browser cannot send native DSD to a DAC, so the test compares the DSD recording with its PCM version, both played as PCM.

What your hit rate actually proves

The score a run needs before it counts as evidence, for each round count the tool offers. Each p is the one-sided binomial probability of scoring that well or better by guessing, for a round count chosen before the test starts. The conventional line for ruling out guessing is p < 0.05.

RoundsSignificant from (p < 0.05)Strong from (p < 0.01)One right fewer: not significant
55 of 5, p = 0.031Not reachable on this many rounds4 of 5, p = 0.188
109 of 10, p = 0.01110 of 10, p < 0.0018 of 10, p = 0.055
1612 of 16, p = 0.03814 of 16, p = 0.00211 of 16, p = 0.105
2015 of 20, p = 0.02116 of 20, p = 0.00614 of 20, p = 0.058

A score below its line is not evidence of a difference, however close it lands: 8 of 10 (p = 0.055) and 11 of 16 (p = 0.105) are scores guessing reaches often. Nor is a near miss a reason to keep running tests until one passes, which is how chance results get reported as findings. In practice borderline scores are uncommon: a listener who hears the difference tends to score well above the line, and one who does not lands near half. A pass says A and B were told apart under these conditions. It does not say how large or how important the difference is.

What a blind test can and cannot prove

An ABX test gives you two known samples, A and B, then an unknown X that is one of them, and asks which. Repeat it enough times and the proportion you get right separates genuine discrimination from guessing - because guessing converges on 50%.

The statistics are simple binomial: every score has a known chance of turning up by guessing, which the hit-rate table on this page lists, so a high score over a round count fixed in advance is evidence of a real audible difference. The reasoning only runs one way, though: a null result means this difference was not audible to you under these conditions, not that no difference exists.

The value of blinding is that it removes everything except the sound. Knowing which is the expensive one, or which you just bought, reliably changes what people report - not because they are dishonest, but because expectation is part of perception.

The ideas behind the controls

Level matching
The single most important control. A 0.5 dB difference is audible as "better", so any unmatched comparison tests level rather than the thing you meant.
Null result
Not hearing a difference under test. Evidence about this comparison and this listener, never proof that two things are identical.
p-value
The probability of scoring this well by chance. The same hit rate is weaker evidence over fewer rounds, which is why trial count matters as much as hit rate.
Expectation bias
The measurable effect of knowing which is which. It is what blinding exists to remove, and it affects everyone.

Explained in depth Burn-in Harmonic Distortion Jitter

ABX blind test FAQ

  1. What is an ABX blind test?

    An ABX test is a statistical method for showing whether you can tell two audio sources apart. You hear sample A, sample B, then a randomized sample X (either A or B) and decide which one X matches. Get enough trials right and the result is statistically significant; flip a coin and you will not.

  2. How many ABX rounds do I need to be statistically significant?

    Set the count before the run starts, then read the score once, against its line. On the default 10 rounds, 9 correct is significant (p = 0.011) and 10 of 10 reaches p = 0.001; 8 of 10 (p = 0.055) is not significant, and the tool says so, with no borderline verdict in between. On 16 rounds the line is 12 correct (p = 0.038), and on 20 rounds it is 15 (p = 0.021). More rounds let a lower share of right answers count. Running test after test until one passes is the one thing to avoid: it lets chance results through.

  3. Why does ABX testing matter for audiophile gear?

    Because most "I can hear it" claims fall apart under blind, level-matched conditions. ABX cuts through expectation bias by hiding which sample is which. Pass an ABX test on two cables, two DACs or two file formats and you told them apart under those conditions. Fail it and the difference was not audible to you in that test, which bounds it without proving it absent.

  4. Are my audio files uploaded anywhere?

    No. The ABX test runs entirely in your browser using Web Audio AudioBuffers, and so does the check of what the two files differ in. Your files never leave your device, never hit our servers, never get logged. The certificate you can share is just an image. It contains the result, not the audio.

  5. How do I level-match the two samples for a fair test?

    The tool does it for you: it measures each file's integrated loudness in LUFS (ITU-R BS.1770, the measure streaming services normalize tracks to) over the stretch both cover, and turns the louder one down to match. If one render starts a few milliseconds later than the other, it finds the offset and lines the two up first. Two files that are the same audio at different gains are matched by their exact gain ratio, so they null. The file check above the Start button shows the gap it removed. Untick level matching only when the level difference is the thing you want to test.

  6. What if my two files differ only in level?

    Then level matching makes them identical, and the file check says so before the first round: there is nothing left to hear. A score near half is the right result on a pair like that, and the result and the certificate call it a control run rather than a fail. Running one on purpose is worth it: on identical clips, a high score points to a cue outside the sound, such as a noise from the computer, rather than to hearing.

  7. Can I compare a DSD file with a PCM version?

    Yes. Drop a DSF or DFF file, DSD64 to DSD512, into either slot. The tool converts it to PCM in your browser, the way a player does when it plays DSD through a PCM input, with a lowpass filter flat to 30 kHz that removes the noise DSD carries above the audio band. Then compare it with the WAV or FLAC version as usual. A browser plays PCM only, so your DAC's own DSD input is not part of the test. A DSD256 file takes a few seconds per minute of music to convert on a recent computer, and nothing is uploaded.

  8. Does the browser change the sound?

    It plays both files the same way, so it cannot favor one. What it can change is what reaches your DAC. A browser plays at the rate your operating system has set for the output, often 48 kHz, and resamples every file to it, while a dedicated player in exclusive mode (ASIO, or WASAPI exclusive on Windows) switches the DAC to each file's own rate. That is why a 192 kHz file can sound different there. The file check names the rate your browser plays at: set the output to the files' rate and reload the page, and a file at that rate plays without resampling. On Windows, set the bit depth there to 24-bit as well.

  9. How do I compare two masters of the same song?

    Load both and keep level matching on. The louder master is turned down to the same integrated loudness, as Spotify and YouTube do when they normalize tracks, so what is left is what the mastering did. The file check gives each version's loudness in LUFS and its peak-to-loudness ratio, which compression and limiting shrink. Run an ABX test to see whether you can tell them apart, then the blind preference test to see which one you choose without knowing which is which.

  10. What is the blind preference test?

    A second question for two files: which one do you like better? Each round plays them as clip 1 and clip 2 in a new random order, and you pick one. The tool keeps the key, so nobody at the controls has to remember which file is which, and it hides the split until the last answer. At the end it shows how often you chose each file and how likely a coin would be to split the rounds as unevenly.

  11. Why are two strips of white noise so hard to tell apart?

    Because they differ in every sample and in nothing the ear follows over a long stretch. Two takes of the same noise share their spectrum and level, so the file check reports no shared waveform and matching spectra, and on a long loop a score near half is common. Loop half a second instead: a short stretch of noise heard on repeat takes on a pattern of its own, which gives the ear something to compare. The same goes for a reverb tail rendered twice with a different random seed.