Skip to content

Bit depth sets the noise floor

More bits do not make audio warmer or more detailed. Each one lowers the noise floor by about 6 dB, and dither turns rounding error on quiet sounds into steady hiss.

Updated · 6 min read Lesson 2 of 14 in Audio Science

In short

  1. Bit depth sets how far the noise floor sits below full scale, about 6 dB per bit: 96 dB at 16-bit and 144 dB at 24-bit.
  2. Undithered rounding turns into distortion on quiet sounds. Dither swaps it for a steady, low hiss.
  3. Record at 24-bit with headroom, mix in float, and dither once when you export to 16-bit.
In this lesson (8)

The smoother staircase

You set up a session at 48 kHz and 24-bit and feel good about it. 24-bit must sound warmer than 16-bit, you think, with finer detail in the highs. More bits, smaller steps, a smoother wave.

That picture is wrong in a useful way. has nothing to do with frequency response or warmth. It sets one thing: how far the sits below the loudest possible sample.

Why it matters

Quiet details live near the floor

Each sample is stored as a whole number on a grid of levels. Whatever the true value is, it gets rounded to the nearest level, and that rounding error is quantization noise. With enough bits the error is far below anything you can hear. With too few, or with a very quiet signal, it becomes audible.

On a finished master, 16-bit gives about 96 dB between full scale and the floor, which is plenty for playback. Recording is different. You track with , peaks around -18 to -12 , and the quiet end of a performance, a fade or a reverb tail, can sit 60 dB below that. At 16-bit that leaves the tail only a few steps of the grid. At 24-bit the grid is 256 times finer, and the analog noise of the microphone and preamp is far louder than the rounding error. That is the case for recording at 24-bit.

Figure 1. The distance from full scale down to the size of one step, 6.02 dB per bit. A 24-bit file has about 48 dB more room under the music than a 16-bit file.

The science

6 dB per bit

An NN-bit fixed-point format has 2N2^N levels: 65,536 at 16-bit and 16,777,216 at 24-bit. Each added bit doubles the number of levels and halves the size of a step, which lowers the noise floor by 6.02 dB. The range from full scale down to one step is

20log⁡10(2N)≈6.02 N dB20 \log_{10}\left(2^N\right) \approx 6.02\,N \ \text{dB}

which gives 96.3 dB at 16-bit and 144.5 dB at 24-bit. Measured as a full-scale sine against the rounding noise, the figure is slightly higher, 6.02 N+1.766.02\,N + 1.76 dB, or about 98 dB at 16-bit.

Rounding error only behaves like noise while the signal is large compared with one step. On a very quiet signal it follows the waveform, and you hear distortion: a gritty, buzzy edge on the tail of a note that comes and goes with the note. fixes this. Adding a tiny amount of random noise, on the order of one step, before rounding makes the error independent of the signal (Lipshitz, Wannamaker and Vanderkooy, 1992). You trade the distortion for a steady hiss at a slightly higher level, and the quiet note stays audible inside it.

Figure 2. A decaying note rounded to a deliberately coarse grid so the steps show. Without dither, once the note shrinks toward half a step it turns into square-edged steps and then silence: distortion that follows the note. With dither the error becomes noise that no longer follows the note, and on average the fading note is still there inside it.

Listen

Lower the bit depth

A quiet, decaying note, or a real mix played quietly, stored at fewer and fewer bits. Listen to the tails turn gritty, then turn on dither and hear the grit become hiss.

Loading demo…
100%

Start with your speakers or headphones low.

Inside the DAW, mixing usually happens in 32-bit , which stores a scale factor with every sample. That keeps levels above 0 dBFS from between plugins and holds rounding error at about 24-bit precision at any level. It is a working format, and the lesson on 32-bit float headroom covers its limits.

Try it

Make quantization audible

0 of 7 steps done

The same quiet signal is grainy at 16-bit without dither, hissy with dither and clean at 24-bit. None of the versions sounds warmer. Only the noise changes.

Common mistake

Exporting to 16-bit without dither

You reduce a 24-bit or float mix to a 16-bit file without dither, on the idea that a clean export should not add noise. Without dither the rounding error becomes low-level distortion on fades and reverb tails. With dither it becomes a hiss around -96 dBFS that nobody hears at normal listening levels. Dither once, as the last step, whenever the bit depth goes down.

The opposite mistake is expecting more bits to change the tone. Between a 24-bit master and a properly dithered 16-bit copy of it, the only difference is noise around -96 dBFS.

Takeaway

Bits are margin

Record at 24-bit with peaks around -18 to -12 dBFS, so loud moments have headroom and quiet ones stay far above the grid. Mix in the DAW's float engine. When you export a 16-bit file, apply dither once at the end. If someone says a format sounds warmer because of its bit depth, ask what happened to the noise floor.

3 questions

Check yourself

  1. How much lower is the quantization noise floor of a 24-bit file than that of a 16-bit file?
  2. What does dither do when you reduce bit depth?
  3. Why record at 24-bit when a 16-bit master already has about 96 dB of range?

Sources

Show the source
  • Lipshitz, S. P., Wannamaker, R. A., & Vanderkooy, J. (1992). Quantization and dither: A theoretical survey. Journal of the Audio Engineering Society, 40(5), 355-375.

Next lesson · 3 of 14

Where 32-bit float headroom endsIn a 32-bit float mix, levels over 0 dBFS survive between plugins. They clip at the exits: a fixed-point export, the interface and plugins built to clip.

Previous: Sampling is taking photos of air

Portrait of Virzy Guns

Written by Virzy Guns

Songwriter and producer, founder of Virzy Guns Production. Now building HealingWave.