Quantization noise is the small error introduced every time an analog signal is converted into a digital value. Because digital systems can only represent a finite number of distinct levels, the original smooth signal gets rounded to the nearest available step. That rounding error, spread across thousands or millions of samples, behaves like low-level random noise layered on top of your signal.
How Rounding Creates Noise
An analog signal is continuous: it can take on any voltage at any moment. A digital converter, though, has a fixed number of steps it can assign. Think of it like measuring someone’s height but only being allowed to round to the nearest inch. A person who is 5 feet 7.3 inches gets recorded as 5 feet 7 inches. That 0.3-inch gap is the quantization error for that measurement.
In an analog-to-digital converter (ADC), the gap between two adjacent digital steps is called the step size. Any input voltage that falls between two steps gets pushed to the nearest one. The error for any single sample lands somewhere between plus and minus half a step. Across many samples with varying input levels, these errors are essentially random and uniformly distributed, which is why engineers treat the cumulative effect as noise rather than a predictable distortion. The average error works out to about 0.3 of one step size.
Mathematically, the digital output equals the true input multiplied by a gain factor, plus an error term. That error term is the quantization noise. Its power (the RMS value) for a full-scale sine wave input is the step size divided by the square root of 12.
The 6 dB Per Bit Rule
The most useful thing to know about quantization noise is its relationship to bit depth: every additional bit of resolution in a digital system improves the signal-to-quantization-noise ratio (SQNR) by approximately 6 decibels. This is often written as SQNR ≈ 6.02 × q, where q is the number of bits.
That 6 dB gain per bit matters because decibels are logarithmic. Each extra bit doubles the number of available steps, cutting the step size in half and pushing the noise floor further below the signal. In practical terms:
- 8-bit audio gives roughly 48 dB of dynamic range, enough for basic speech but noticeably noisy for music.
- 16-bit audio (CD quality) reaches about 89 dB of signal-to-noise ratio, which comfortably covers the range of most recorded music.
- 24-bit audio pushes the theoretical limit to around 137 dB, well beyond what any microphone or speaker can reproduce, giving engineers headroom during recording and mixing.
These figures represent best-case scenarios for properly scaled signals. If a signal only uses a fraction of the converter’s full range, the effective noise ratio drops because fewer bits are doing useful work.
What It Sounds and Looks Like
In audio, quantization noise is most audible at low bit depths or when a signal is very quiet. At 8 bits, you can hear a gritty, buzzing texture underneath the music. At 16 bits the noise floor sits low enough that it’s inaudible under normal listening conditions. At 24 bits it’s irrelevant in any real-world playback scenario.
In digital photography and video, quantization noise can show up as banding or posterization, where smooth gradients (like a sunset sky) break into visible stair-step bands of color instead of blending seamlessly. When the analog voltage from an image sensor gets digitized into a raw pixel value, it’s rounded to the nearest integer, and that rounding is the quantization error. In practice, though, other noise sources in a camera sensor (thermal noise, read noise) are typically much larger than the quantization error itself, so banding from pure quantization is rarely the dominant problem in modern cameras.
Dithering: Trading Noise for Accuracy
One of the most effective ways to deal with quantization noise is, counterintuitively, to add more noise. This technique is called dithering. A tiny amount of random noise (about half a step’s worth) is deliberately mixed into the analog signal before it reaches the converter. This causes samples near a step boundary to randomly land on either side of it, rather than always rounding the same direction.
By itself, a single dithered sample isn’t more accurate. But when you average multiple samples together, the proportions that landed above and below the boundary reveal where the true signal actually sat. The result is that you can extract resolution finer than the converter’s native bit depth. A 12-bit converter with dithering enabled can effectively perform at 14-bit resolution when averaging is applied. In audio mastering, dithering is standard practice whenever reducing bit depth (for example, converting a 24-bit mix down to 16-bit for CD release) because it replaces the harsh, signal-correlated distortion of raw quantization with a smooth, low-level hiss that sounds far more natural.
Noise Shaping in Modern Converters
Modern high-resolution converters, particularly the sigma-delta type found in most audio interfaces and measurement instruments, use a technique called noise shaping to push quantization noise out of the frequency range you care about.
A sigma-delta converter samples at a rate many times higher than the minimum required (oversampling), then uses a feedback loop with an integrator to reshape the frequency distribution of the quantization noise. The integrator acts like a high-pass filter for the noise, forcing most of the quantization error energy up into very high frequencies while leaving the lower frequencies (where your actual signal lives) much cleaner. A digital filter then strips away those high-frequency noise components entirely.
The combination of oversampling, noise shaping, and digital filtering dramatically increases the effective resolution. A converter with a simple 1-bit quantizer at its core can achieve 24-bit performance by running at a high enough oversampling rate and applying aggressive noise shaping. This is why sigma-delta converters dominate audio and precision measurement applications: they trade raw speed for extraordinary noise performance in the frequency band that matters.
Why Bit Depth Choices Matter
Understanding quantization noise helps explain why different applications use different bit depths. Telephone audio uses 8 bits because intelligibility is the only goal and bandwidth is limited. Music streaming and playback use 16 bits because 89 dB of dynamic range exceeds what most listeners can perceive in a normal environment. Professional recording and mixing use 24 bits not because anyone can hear the difference during playback, but because the extra headroom means small signals stay well above the noise floor even before processing, and gain adjustments during mixing don’t amplify quantization artifacts into the audible range.
In digital signal processing systems beyond audio, the same principle applies. Every multiplication, filter, or transform performed on digital data introduces its own rounding errors, and those errors accumulate through a processing chain. Using wider bit depths at each stage keeps the compounding quantization noise from degrading the final output. Each additional bit of internal processing resolution reduces the quantization noise power by 6 dB, giving engineers a straightforward way to budget precision against computational cost.