What the phrase actually refers to
Audio is stored as a stream of amplitude numbers. In the ordinary representation those numbers are linear: 0.5 is exactly half of 1.0, and doubling the number doubles the signal. A logarithmic representation stores the size of a number instead of the number itself, so halving the stored value doubles the amplitude. That single change is the whole idea.
The important thing is that decibels already do this. The decibel is defined as 20 times the base-10 logarithm of an amplitude ratio, which means every dB readout in your session is a logarithmic quantity describing a linear one. When someone says mixing in the log domain, they usually mean one of three specific things, and they are worth separating.
- The control taper: how much amplitude change one unit of fader travel produces. This affects your hands, not the file.
- The internal arithmetic: whether gain changes and buses sum as amplitude values or as logarithmic values. This does affect the file.
- The meter and the unit: whether the level is reported in dBFS. It already is, on every meter you own.
Most of the popular argument is about the first item, and the first item does not change the output. If a fader reads 0 dB it multiplies by 1.0 whatever curve the pointer travelled along, so two engineers setting the same dB value by different routes end up with bit-identical audio.
Where linear arithmetic actually shows up
There is a real difference, and it lives in gain changes and bus summing. Applying a gain of 0.5 to a linear signal halves every sample. Applying a gain of minus 6 dB means multiplying by 10 raised to the power of minus 6 over 20, which is 0.5012, not exactly 0.5. Multiply a large number of samples by an inexact factor and the rounding errors accumulate in one direction instead of cancelling.
Summing is where this becomes audible. If twenty-four channels arrive at a bus already near full scale, adding them in phase gives a peak of about 24 times the individual signal, roughly 27.6 dB above where each one sat. A fixed-point integer bus cannot represent that and clips hard.
| Operation | Linear amplitude | In decibels |
|---|---|---|
| Fader at unity | 1.0 | 0.0 dB |
| Fader at half travel | 0.5 | -6.0 dB |
| Fader at one tenth | 0.1 | -20.0 dB |
| Two sources, in phase | 2.0 | +6.0 dB |
| Twenty-four sources, in phase | 24.0 | +27.6 dB |
This is the argument worth making, and it points at arithmetic precision rather than at mysticism. A logarithmic bus also does not sum the way you would expect, which is why processors that genuinely work in the log domain specify a minimum input level and a maximum signal-to-noise ratio: below roughly minus 90 dBFS a logarithmic representation runs out of usable digits. The trade-off is that such a bus is not free. It introduces its own quantisation floor and complicates every plugin that passes audio through it, and for most material the artefact it prevents is one you would never have heard anyway.
What floating point actually bought
The second substantive point concerns headroom, and it cuts against received wisdom. Traditional advice says you must keep peaks below minus 6 dBFS or minus 12 dBFS so that summing has room. On a fixed-point 24-bit bus that was sound: 24 bits gives about 144 dB of range, and clipping there is destructive. In 32-bit floating point the exponent does the scaling, and the practical usable range sits somewhere around 1500 dB.
- Clipping a float mix bus requires you to push the input past what any realistic arrangement produces.
- Headroom discipline now matters for editing clarity, not for preventing an audible failure.
- A modest fader reduction still helps, because it leaves more precision for plugins downstream that are fixed point internally.
That is not a licence to work carelessly. Lower input into a plugin means more of its own internal headroom. One genuine problem survives in a float session: sample peaks are not the same as true peaks. A signal reading minus 0.1 dBFS at the samples can exceed 0 dBFS after the reconstruction filter a DAC applies, and those inter-sample peaks are why a true-peak meter exists and why minus 1 dBTP is the safe ceiling before lossy encoding.
The last step that never became optional
Reducing bit depth is quantisation: rounding each sample to a smaller set of levels. Where the level falls between two representable values the quantisation error is not random, it correlates with the signal, and the result is distortion rather than noise. Dither adds a small amount of noise, deliberately and at a known level, so the total error becomes random instead of correlated, and the distortion energy is spread thinly enough that it sits below the noise floor of the format. That is the whole mechanism. It is not a tonal colouring, and it does not make anything sound better in the flattering sense that plugins claim. Its job is to make truncation inaudible, which is a narrower and more defensible claim than the marketing usually makes.
The point that catches people out is that dither belongs at the very end, and a streaming master never really has an end you control. You export at 32-bit float, the service transcodes to AAC or an equivalent at a fixed bit depth, and that transcode is the quantisation step. If you dither at export you are paying for noise twice, which is harmless. If you skip it, the transcoder handles it as well as it can, which is not something you chose.
The bottom line
The popular version of this idea is wrong: nearly every fader in a DAW is already log-tapered in its user interface, so switching tapers does not change a single sample. What is genuinely true is narrower and more useful — gain changes and summing behave differently under a linear representation, floating-point summing means gain staging matters less than folklore insists, decibels were logarithmic before the topic was fashionable, and dither is still mandatory at export.
Set a reference level without touching pitch
432Hz MASTER shifts one ratio across the whole spectrum, so any change you hear after matching level is gain, balance or dynamics — never tuning.
Frequently asked questions
Are DAW faders already logarithmic?
The user interface is. Almost every channel fader is scaled in decibels, so equal distances of mouse travel are equal decibel changes. The samples underneath remain linear amplitude values.
Does a log fader sound better than a linear one?
No, because the taper only decides where the control sits, not what the audio does. Two faders set to the same dB value produce identical output, whatever curve the pointer follows.
Do I still need gain staging in a floating-point session?
You need it far less. With 32-bit float and roughly 1500 dB of headroom, headroom problems have to be created deliberately before they can be heard. Keep levels tidy for editing reasons, not to prevent clipping.
Why can I not skip dither if I export at 32-bit float?
Because streaming services re-encode your float file to AAC or a lossy format, which is a fixed-bit-depth step. That conversion quantises, so dither has to happen before it, not at your original export.