Bit Depth ABX Test — 16 vs 8-bit

A blind bit depth test presents audio clips rendered at different quantization levels without revealing which sample is playing. Listeners compare the recordings in randomized trials to identify whether quantization noise, truncation artifacts, or dynamic range differences between sixteen-bit and lower bit depths are audible on their playback system. This drill runs the comparison from 8-bit (where nobody is fooled) down to 14-bit (where almost nobody is), and finds the rung where your ear stops filing reports.

How to use it

1. Begin at the 8-bit rung. Signal A is the 16-bit reference; Signal B is the same material truncated to 8-bit integers. The difference is a coarse, granular buzz, deliberately easy.
2. Work down the ladder. The rungs are 10-bit, 12-bit, and 14-bit; each step shrinks the quantization error by a factor of four in voltage (6 dB of dynamic range per added bit).
3. Study quiet passages on purpose. Quantization artifacts live at low signal levels; listen to reverb tails and fade-outs during the open A/B phase, where the stair-step granularity is most exposed.
4. Answer X trials blindly. The engine tracks trials, correct picks, streak, p-value, and confidence exactly as in the rest of the Blind ABX series.
5. Record your last significant rung. That is your audibility threshold for truncation on that system, in bits.

PCM resolution: the 6.02N rule

An N-bit linear PCM word divides full scale into 2^N discrete amplitude codes. The rounding error of quantization is bounded by ±½ of one least-significant bit (LSB), and treating that error as noise uniformly spread over one step gives a noise power of Δ²/12 for step size Δ. The signal-plus-error geometry produces the classic formula: the available dynamic range grows 6.02 dB per bit (20 × log10 2), and a full-scale sine wave adds a further 1.76 dB because its RMS is 1/√2 below its peak.

Bit depth	Amplitude codes	Dynamic range (6.02 × N)
8-bit	256	48.2 dB
10-bit	1,024	60.2 dB
12-bit	4,096	72.2 dB
14-bit	16,384	84.3 dB
16-bit (CD, IEC 60908)	65,536	96.3 dB
20-bit	1,048,576	120.4 dB
24-bit	16,777,216	144.5 dB

Why is 16 bits enough for a consumer format? Because the bottleneck sits downstream of the disc. A quiet listening room runs around 25–35 dBA of ambient noise, and at comfortable playback levels the useful span between that floor and safe peak levels is on the order of 60 dB. With the CD’s ~96 dB of coded range converted to sound, the quantization floor lands 30 dB or more below what the room already masks, which is why the Red Book format (IEC 60908, 16-bit / 44.1 kHz) survived three decades of listening. The 24-bit argument is not about hearing a lower noise floor either; it is about headroom for capture and processing, since every gain change, fade, and plugin before printing the master re-quantizes the signal and compounds whatever error is introduced.

The character of the error depends on how it is introduced. Raw truncation ties the error to the signal: as a decaying tone sinks by one step at a time, the rounding repeats at the tone’s own period, producing correlated harmonic grit rather than smooth hiss, and the decay audibly “crunches” through a handful of coarse steps before vanishing. Dither breaks that correlation by adding noise around ±½ LSB before rounding, which raises the measured noise floor (a TPDF stack of two uniform dithers carries twice the power of a single rounding error, +3 dB) but converts the artifact from distortion into benign, decorrelated hiss and preserves low-level linearity. The Dithering and Noise Shaping demo plays that trade-off directly; this drill tests the undithered, honest-to-worse case.

How the scoring decides what you actually heard

Perceptual claims about formats are notoriously easy to fake, so the engine applies the same discipline as the rest of the ABX series: fixed references, a randomly assigned hidden X per trial, forced choice, and cumulative binomial scoring against a 50% guessing null. What differs here is the staircase logic. Because each rung halves and halves again the error voltage, your results across rungs sketch a threshold curve: significant at 8 and 10 bits, chance at 14 bits, for example, says your system resolves truncation somewhere between 60 and 84 dB of quantized dynamic range. The engine’s needs more trials badge matters especially on the middle rungs, where real perception starts at barely-above-chance rates and only accumulated trials separate signal from noise. A useful protocol: three short sessions on the same rung on different days, then trust the verdict only when it repeats.

Accuracy, limits and troubleshooting

• Your playback path is not 16-bit anyway. Browsers mix in floating point and most OS audio stacks dither or resample before the DAC, so what you are hearing is the injected coarse quantization riding on top of an effectively very-low-noise chain, not your sound card’s native word length.
• Volume is a confound. 8-bit material measures slightly louder at peak because of added error energy; if one clip clearly wins on level alone, the drill’s loudness normalization is doing its job, so do not override it with the volume knob.
• High monitoring gain flatters you. Turning up to hunt for 14-bit grit also raises room noise and headphone hiss, and the trial collapses into hearing noise, not quantization. Keep one level.
• Closed backs with high sensitivity reveal their own hiss floor around the quiet passages this test targets; if the rings feel muddier than expected, try a lower-sensitivity headphone.

Related tools

• Dithering and Noise Shaping — hear truncation, TPDF dither, and shaped dither side by side.
• Dynamic Range Test — measure your device’s noise floor and effective resolution.
• Blind ABX hub — jump between the seven threshold drills sharing one scoring engine.
• Spectrum Analyzer — watch the harmonic sidebands of undithered quantization appear live.