Bit Depth and Quantization Noise

Bit depth decides whether the quietest detail survives, and quantization noise is the price.

Related

Bit DepthDitherNoise Floor

The problem

At the same 44.1 kHz, why does an 8-bit recording hiss like sand while a 24-bit one is nearly silent?

How it works

Every sample is an integer, and bit depth sets how many steps are available. Sixteen bits give 65,536 steps; 24 bits give about 16.7 million. Coarser steps flatten the quietest detail, and the error becomes quantization noise. That noise is spread fairly evenly across the band, and each extra bit lowers the floor by roughly 6 dB. Sixteen bits are enough at normal listening levels; the extra room mainly prevents clipping and absorbs accumulated error across processing. Floating point replaces steps with an exponent, so processing adds almost no noise.

Practice and settings

Record at 24 bits with 12 to 18 dB of headroom. Mix internally at 32-bit float. Deliver 24-bit files, and add dither only when making the 16-bit version. Judge depth by whether the quietest passage is buried, not by file size. Tracks with heavy attenuation, such as long reverb tails, expose the floor first.

What to listen for

Hear a soft piano note's tail: low depth cuts it short or adds grain. Hear the breath at the end of a vocal phrase, which turns into a thin veil when the floor is high. Hear a fade-out, where quantization noise often becomes obvious in the last seconds. In quiet chamber music, listen between the instruments, not the notes.

Common misconceptions

That bit depth sets the frequency limit; that 24 bits always sounds better; that dither can rescue clipping.

Sources

• Ken C. Pohlmann, Principles of Digital Audio — 量化噪声与位深的换算核对，未抄原文
• Stanley Lipshitz, John Vanderkooy, Quantization and Dither — 抖动与噪声整形的结论核对，未抄原文

Used for fact-checking only; all prose is original.