Bit Crushing — Breaking the Digital Grid
Waveshaping bent the relationship between input and output. Bit crushing does something different: it deliberately makes the digital grid coarser. We can reduce amplitude resolution, reduce time resolution, or do both.
HEAR IT → SEE IT → MANIPULATE IT → NAME IT → SHOW THE MATHS
1 · The experiment
Start with the tone, then load a real recording. Pull the bit depth down slowly. After that, switch on sample-rate reduction. They are often bundled together in “bit crusher” plugins, but they are not the same operation.
2 · Bit depth = the Y axis
Bit depth controls how many amplitude values are available. Fewer bits means larger jumps between allowed values.
Input is shown lightly; quantised output is shown strongly. At low bit depths the smooth waveform becomes a staircase.
3 · A term you keep seeing…
Transfer function — and transfer curve
A transfer function describes how a system transforms an input into an output: the relationship between what goes in and what comes out.
For the nonlinear graphs we have been drawing, transfer curve or input–output characteristic is the more precise term.
The horizontal axis asks: “What value came in?” The vertical axis answers: “What value comes out?”
4 · Bit depth and sample rate are two different axes
Amplitude resolution
Reduce bit depth and amplitudes are rounded onto fewer permitted levels.
Time resolution
Sample-rate crushing holds a value for longer before allowing another update.
A typical “bit crusher” may do both. That does not make them the same process.
5 · Quantisation error is distortion
Quantisation replaces the original value with the nearest available level:
where L is the number of available levels. The error is simply:
The error is related to the signal, so at coarse resolution it can sound distinctly tonal, gritty or buzzy.
6 · Dither: change the error
Switch Dither ON and try 4-bit or 3-bit audio. A tiny random signal is added before quantisation.
Dither does not give the missing bits back. It trades correlated quantisation distortion for a more noise-like error and can preserve very low-level information statistically.
7 · Remember the mirror?
Now try sample-rate reduction. Holding samples introduces spectral images and can create aliases. The Nyquist lesson from earlier returns: frequencies that cannot be represented correctly fold into the available band.
OUTPUT LPF is deliberately available as an experiment. Filtering can remove some high-frequency content before/around rate conversion, but a normal filter cannot magically identify aliases once they have already folded into the wanted band.
8 · What does oversampling fix?
Try changing Process Rate 1× → 8× while using a very low bit depth.
The important discovery is:
4-bit @ 384 kHz is still 4-bit
Oversampling gives us more samples in time. It does not create more allowed amplitude levels.
9 · But oversampling can still matter
If a process creates high-frequency components — nonlinear processing, hard transitions, sample-rate conversion — processing at a higher internal rate can give filtering more room before returning to the host rate.
So oversampling is a tool for a particular problem, not a universal “quality” switch.
10 · The section closes where it began
We started with tiny operations and ended up explaining filters, dynamics, distortion, aliasing, oversampling, modelling and quantisation. Nothing magical appeared.
Bit crushing is simply another arrangement of familiar ideas: measure/represent a sample, force it onto a smaller set of values, optionally hold it through time, and listen to what information was changed.
And now the earlier digital-audio vocabulary joins the DSP toolbox: