Timing ABX Test — Latency Drill

This blind timing test measures the smallest right-channel delay you can hear in a stereo click, using randomized double-blind ABX trials. Every click is synthesized in the browser with the Web Audio API, and nothing is uploaded.

How to use it

1. Set the delay rung. The presets are 40 ms, 20 ms, 10 ms, 5 ms and 2 ms, and the session opens at 20 ms. Changing the rung resets the trial count so the statistics never mix two difficulty levels.
2. Learn A and B. Listen A plays the click in both channels at once (synchronized stereo, 0 ms). Listen B delays the right channel by the selected number of milliseconds. Both are open references you can switch between as often as you like.
3. Listen to the unknown. Listen Unknown X plays a copy secretly drawn from A or B. X is re-drawn for every trial, so no pattern carries over from the previous answer.
4. Commit with the guess buttons. Choose X is Signal A or X is Signal B. The design is forced choice, so there is no “same” answer to hide behind.
5. Watch the statistics, not your gut. Trials, score, accuracy, streak, p-value and a significance badge update after every answer, and the last ten trials stay listed for review.
6. Walk the ladder. Finish one rung, reset, then move to the next. Your threshold is the smallest delay that reaches significance on two separate days.

What the click actually is

Both references share one stimulus: a 1 kHz tone burst shaped by an exponential decay, e^(-800t), inside an 8 ms window, repeating once every 2 seconds. The decay time constant is 1.25 ms, which makes the burst a sharp transient rich enough to expose both timing and timbre cues. Signal B delays only the right channel; signal A leaves the two channels aligned.

The delay is quantized to whole samples, and that sets the real precision floor. At 44.1 kHz one sample is 22.7 microseconds, so the 2 ms setting renders as 88 samples, or 1.995 ms, while the 40 ms setting renders as 1,764 samples, or 40.000 ms. At 48 kHz the 2 ms setting is exactly 96 samples. Nothing else differs between A and B: no level change, no EQ, and no added noise.

Haas fusion, echo thresholds and comb filtering

A delay shorter than roughly 35 ms fuses with the direct sound into one wider auditory event, which is the precedence effect and the reason a 20 ms offset reads as spatial width rather than a second hit. The echo threshold depends on the signal: transient clicks break into two separate events at far shorter delays than sustained music or speech, so the 40 ms rung is deliberately past that point for a click and reads as an obvious double hit.

Between 2 ms and 10 ms you are mostly hearing comb filtering. A delayed copy summed with its original cancels at odd multiples of 1/(2τ), where τ is the delay time, so the pattern’s null spacing is 1/τ:

Delay	Comb spacing	First null	Delay as a distance at 343 m/s
2 ms	500 Hz	250 Hz	0.69 m
5 ms	200 Hz	100 Hz	1.72 m
10 ms	100 Hz	50 Hz	3.43 m
20 ms	50 Hz	25 Hz	6.86 m
40 ms	25 Hz	12.5 Hz	13.72 m

The distance column is the physical path difference that produces the same delay at 20 degrees Celsius, where sound travels at 343 m/s. At 2 ms the offset is about the width of a close speaker pair and the effect is a mild timbre shift; at 10 ms the notch spacing drops to 100 Hz, which is the hollow, flanging coloration the preset marks as comb filtering.

How the scoring decides what you heard

This is a two-alternative forced choice. X is drawn at random on every trial with a 50% chance of being A, so pure guessing converges on 50% over many trials. The p-value is the one-tailed binomial probability of scoring at least this well by guessing: P(K >= k) = sum of C(n, i) * 0.5^n for i running from k to n. At the alpha = 0.05 convention, the decision boundary looks like this:

Trials	Minimum correct	p-value	Verdict
5	5	0.0313	significant
8	7	0.0352	significant
10	9	0.0107	significant
10	8	0.0547	not significant
20	15	0.0207	significant

There is no fixed number of rounds. One worked example from the same math: at 10 trials, 9 correct gives P = 11 / 1024 = 0.0107, while 8 correct gives 56 / 1024 = 0.0547. You can run as many trials as you want, and the badge reports a minimum of 5 trials until the earliest possible significant result (5 out of 5, p = 0.03125) can be reached. Seven out of ten does not pass, since that pattern arises by luck about one session in six. The honest framing matters: a blind test measures whether you can hear a difference under controlled conditions. It says nothing about which signal is better, and nothing about the quality of your gear.

Listening conditions that make the result valid

• Use a quiet room and headphones. The test needs two genuinely separate channels. A mono button, a phone with a single speaker, or a speaker pair too close together collapses the delay cue and makes the result meaningless.
• Hold the playback level fixed. Delay does not change level, so any loudness difference you notice is coming from your system. Disable loudness normalization, spatial audio and crossfeed for the duration.
• Do not peek. The A and B buttons toggle playback, and the interface never marks X. Looking away from the history row between trials keeps expectation out of the answer.
• Take breaks. Expectation and fatigue build within minutes on a 2-second click loop. Ten to fifteen trials per rung per sitting is enough.
• Repeat before you believe it. One significant session is evidence; the same rung significant on two days is a stronger claim about your hearing rather than about that afternoon.

Accuracy, limits and troubleshooting

• Mono playback defeats the test. Check that your device is not in mono mode and that a Bluetooth speaker pair is not summed to one channel.
• Bluetooth does not invalidate the comparison. Codec transport latency is shared by both channels, so it cancels out, though heavy SBC or AAC processing can soften the 2 ms transient.
• Speakers work but add their own delays. Room reflections arrive after the direct sound and blur the very threshold you are measuring; headphones isolate the effect under test.
• Crossfeed and 3D audio deliberately blend channels. Turn them off.
• Keep the click level moderate. Loud transients can distort small drivers and make the result about distortion rather than timing.
• Educational measurement only. This is not a hearing test, and it cannot diagnose a hearing problem. If you are concerned about your hearing, see an audiologist.

Related tools

• Audio Latency Test — measure your own round-trip latency and buffer behavior.
• Stereo Perception Test — verify channel separation and phantom-center imaging.
• Blind Frequency ABX Test — the same scoring discipline applied to Hz discrimination.