Level ABX Test — Loudness Drill

This blind level test measures the smallest loudness step you can hear on a 1 kHz tone, using randomized double-blind ABX trials with strict binomial scoring. Everything is synthesized locally in the browser with the Web Audio API, and nothing is uploaded.

How to use it

1. Choose the rung. The presets are 3.0 dB, 1.0 dB, 0.5 dB, 0.2 dB and 0.1 dB, and the session opens at 1.0 dB. Changing the rung resets the trial count.
2. Compare the references. Signal A is a 1 kHz sine peaking at 0.5 of full scale, about -6 dBFS. Signal B is the identical sine multiplied by 10^(-ΔdB/20), so its level is exactly the rung you selected. Both are two-second loops with short raised-cosine fades.
3. Listen to the unknown. Listen Unknown X plays a copy secretly drawn from A or B on every trial, with nothing in the interface marking the draw.
4. Make the forced choice. Answer X is Signal A or X is Signal B. There is no “same” option, because the two references always differ by the selected attenuation.
5. Watch the statistics. Trials, score, accuracy, streak, p-value and the significance badge update after each answer, and the last ten trials stay visible.
6. Repeat across rungs and days. Your just-noticeable difference is the hardest rung that reaches significance twice on separate days, not the best single run.

What the decibels mean

The level difference between the references is defined by the gain ratio, ΔL = 20 * log10(V_A / V_B). Inverting it gives the amplitude multiplier used for B, which is 10^(-ΔL/20):

Rung	Amplitude of B versus A	How much quieter B is	Typical context
0.1 dB	0.98855	1.14% lower	Extreme rung, near the perceptual limit
0.2 dB	0.97724	2.28% lower	Mastering-room territory
0.5 dB	0.94406	5.59% lower	Trained engineer under headphones
1.0 dB	0.89125	10.87% lower	Standard JND benchmark
3.0 dB	0.70795	29.2% lower	Half power, obvious to almost everyone

Three decibels is a doubling or halving of electrical power, which is why the 3.0 dB rung is the warm-up. Loudness and level are not the same quantity, though. The equal-loudness contours in ISO 226:2023 show that the same dB change produces a different loudness sensation depending on frequency, which is why this drill holds frequency constant at 1 kHz and why the rungs cannot be read as universal loudness constants.

Weber’s law and the 1 kHz tone

The just-noticeable difference for level follows Weber’s law: what matters is the ratio of the two stimuli, not their absolute difference, and decibels express that ratio logarithmically. A 0.5 dB step is a ratio of 1.059, and a 3 dB step is a ratio of about 1.412, regardless of whether the quiet reference is at 60 dB SPL or 80 dB SPL.

Human loudness JND for a steady 1 kHz tone in quiet conditions is commonly placed between 0.5 dB and 1.0 dB. Listeners with critical-listening training and controlled headphone monitoring can reach 0.2 dB, and 0.1 dB is the practical floor for a single listener running a browser test on consumer gear. The tone here is two seconds long, which is long enough for a reliable loudness judgment and short enough to limit adaptation. Shorter tones, lower frequencies and higher frequencies all widen the JND, which is why one rung from this test does not describe your hearing in general.

What the scoring proves, and what it does not

This is a two-alternative forced choice with a hidden X drawn at 50/50 on every trial, scored against the one-tailed binomial probability of guessing that well: P(K >= k) = sum of C(n, i) * 0.5^n from i = k to n. At alpha = 0.05, the boundary is:

Trials	Minimum correct	p-value	Verdict
4	4	0.0625	not significant
5	5	0.0313	significant
10	9	0.0107	significant
10	8	0.0547	not significant
20	15	0.0207	significant

The trial count is open-ended and the badge reports a minimum of 5 trials before significance is possible. One worked example: 5 correct out of 5 gives P = 1 / 32 = 0.0313, and 9 out of 10 gives 11 / 1024 = 0.0107. Three correct out of four has a p-value of 0.3125, which is fully consistent with coin tossing, and 8 out of 10 misses the threshold at 0.0547. That strictness is the point: a blind test measures whether you can hear a difference under controlled conditions. It does not measure the quality of your gear, and it will not tell you which of two masters is better.

Listening conditions that make the result valid

• Quiet room, fixed level. Ambient noise forces you to raise the playback level, and loudness adaptation sets in faster at high levels. Set the monitor level once and leave the slider alone.
• Disable normalization and dynamics processing. ReplayGain, Sound Check, loudness equalization, night mode and Bluetooth AGC all change gain behind your back, and every one of them can swamp a 0.1 dB difference.
• Keep levels matched to within ±0.1 dB in any comparison where level is not the variable. At the 0.1 dB rung the effect under test is 1.14% of amplitude, smaller than a typical system volume step, so riding the slider mid-session destroys the measurement.
• Do not watch a meter. The dashboard deliberately shows score rather than which reference is playing, and a visible meter or spectrum display would leak that information. If you have a decibel meter running, turn it away from you.
• Break every 10 to 15 trials. Loudness judgments drift with fatigue, and long sessions also invite guess-clicking.
• Stay on one playback chain per session. Switching headphones, speakers or output devices mid-session changes the noise floor and the apparent JND.

Accuracy, limits and troubleshooting

• System volume steps vary. Some operating systems and Bluetooth devices move the volume slider in 0.5 dB to 1 dB increments while others are finer, but all of them are coarser than the tool’s sample-exact attenuation. Set the system volume once and let the web audio do the fine work.
• Bluetooth can change gain between connections. Codec gain and absolute-volume synchronization differ per link, so reconnect once, set the level, and do not re-pair mid-test.
• Browsers mix in floating point, which preserves a 0.1 dB difference through the signal path; resampling a steady tone does not alter its level either. The weakest link is almost always the monitoring chain and the room, not the browser.
• Small drivers compress. At high levels, headphone or speaker nonlinearity flattens the very level difference you are trying to detect, making the result a test of the transducer rather than your hearing.
• A 1 kHz result does not transfer to music. Broadband, transient program material gives the ear more cues and more masking at the same time. A 0.5 dB result here is a floor for that tone and those conditions, not a promise about mixes.
• This is educational measurement, not a hearing test. Level discrimination is not a diagnostic threshold. If you are concerned about your hearing, see an audiologist.

Related tools

• Decibel Meter — check playback levels before and during a session.
• Loudness Meter — read integrated LUFS and true peak per ITU-R BS.1770.
• Blind Pitch ABX Test — cents discrimination with the same scoring engine.