Clipping vs limiting

From the broadcast.audio wiki · Chapter 05 of 9

Part of a series on
Broadcast audio processing

Chapters

Reference

Contents

1Turning down versus shaving off
2Distortion, spent deliberately
3Knee: how hard is hard
4Oversampling, and what it does not do
5Inter-sample peaks
6Why clipper first, limiter last
7In the product, briefly
8See also

Loudness is a ratio problem. However you get there, being louder means reducing the distance between the peaks of your signal and its average — because the peaks are what hits the ceiling, and the average is what people hear.

There are only two ways to close that gap. Bring the average up, which is compression and levelling: chapters 2 and 3. Or bring the peaks down, which is this chapter. Every real chain does both, but the last few decibels are always won on the peaks, because compression runs out of road long before peaks do.

Turning down versus shaving off

Say you have one 3 dB peak sticking out of an otherwise well-behaved signal.

A limiter handles it by reducing the gain. It sees the peak coming (or reacts to it), pulls the gain down by 3 dB, holds it there, and releases back to unity over some recovery time. Everything around the peak — the tail of the previous note, the start of the next one, the room, the reverb — comes down too, and comes back up as the limiter recovers. The peak is controlled and the waveform is otherwise undistorted.

A clipper handles it by removing the part above the ceiling. It touches only the samples that exceed the threshold. The rest of the signal — including the sample immediately before and after — is bit-for-bit unchanged. There is no gain movement at all, so nothing pumps, nothing breathes, nothing ducks.

Both cost something, but they cost different things. The limiter's cost is modulation: a dip in level around every peak. Do it occasionally and nobody notices. Do it continuously, which is what happens when you ask a limiter for real loudness, and the dips merge into a permanent, audible pumping — the "sucking" sound of a limiter working too hard, where the mix breathes against every kick drum and transients lose the space behind them.

The clipper's cost is distortion. The material you removed doesn't vanish into thin air; the resulting waveform contains harmonics and intermodulation products that were not there before.

Distortion, spent deliberately

The reason clipping is not a disaster is that its cost is highly dependent on the material, in a way that turns out to be very convenient.

Clipping a transient — a snare, a pick attack, a consonant — puts a millisecond or two of harmonics underneath something that is itself a broadband burst of energy. The distortion is masked by the very event that caused it. You can take several dB off a drum transient this way and a listener will report that the drum sounds punchier, which it does, because you removed the peak without touching the body.

Clipping a sustained tone — a bass note, an exposed vocal, a solo horn, a sine-like synth — is a different story. The harmonics are continuous, they are harmonically related to a pitch you can hear, and there is nothing to mask them. It sounds like a rasp, a buzz, or a hardness that follows the note.

So the practical rule is: clipping is cheap on dense, transient, broadband material and expensive on sparse, tonal, exposed material. Which is exactly why an aggressive setting that sounds enormous on a rock track sounds wrong on a solo piano, and why a station that plays a mixed schedule has to make a compromise there.

It is also cumulative. Every stage that clips the same peaks adds its own contribution, and they do not politely overlap, so part of designing a chain is making sure each peak is only paid for once.

On FM there was never a choice about whether to clip. The deviation limit (chapter 8) is a hard wall, the audience is in a noisy car, and the competitive pressure is real. The only question was ever where to spend the distortion — a decision broadcast engineers have been making very carefully for a long time, which is why good broadcast clipping sounds far better than the word suggests.

Knee: how hard is hard

A pure hard clipper has a corner in its transfer curve, and a corner generates a lot of high-order harmonics. A soft knee starts bending the curve some distance below the ceiling, so the transition is gradual.

Soft is generally less harsh — the harmonic series decays faster and there is no sudden onset — but it is not free, because it touches more of the signal. Below the knee the waveform is being gently compressed rather than left alone, which takes the edge off transients in a way a hard clipper does not. Hard is more audible when it acts and more transparent when it doesn't; soft is the reverse.

Oversampling, and what it does not do

A clipper creates harmonics. Some of those harmonics are above half the sample rate, and in a sampled system anything above half the sample rate does not simply disappear — it folds back into the audio band as aliasing.

That matters because aliased products are not harmonically related to anything. Ordinary clipping distortion sits at multiples of the notes being played, so it is partly masked and partly sounds like "drive". Aliased distortion sits at arbitrary frequencies, moves in the wrong direction when the pitch moves, and sounds like grit and intermodulation. It is the least musical distortion you can make.

Running the clipper at a higher internal sample rate pushes the fold-back point further up, so more of the harmonics stay where they belong and get filtered out on the way back down. Be clear about what this achieves, though: oversampling does not reduce the distortion, it relocates it. The harmonics are still generated; they simply end up above the audio band instead of scattered through it.

A different approach attacks the same problem directly. Take the difference between the clipper's input and its output — the error signal, which is exactly the material you removed — low-pass it so it contains nothing above the audio band, and subtract that band-limited error from the original instead. You get peak reduction with no out-of-band products at all. The trade is that a band-limited correction cannot perfectly cancel an instantaneous peak, so a little overshoot survives and something downstream has to catch it.

Inter-sample peaks

Samples are not the signal. They are points on a band-limited curve, and the curve between them can — and regularly does — go higher than any of the points.

A file whose highest sample is exactly 0 dBFS can reconstruct to +1.5 dB or more when a converter, a resampler or a lossy decoder builds the waveform back. This is not hypothetical: it is the everyday consequence of clipping a signal flat to full scale and then encoding it, and it shows up as clipping in the listener's playback chain that you cannot hear on your own meters.

Measuring it requires upsampling before you measure. A 2× interpolated estimate catches most of it cheaply; the standardised measurement uses 4× oversampling and reads a bit higher. Either way the practical response is to leave headroom — a true-peak ceiling somewhere around −1 dBFS is the common default, and more if you know your output is going through a lossy encoder.

Why clipper first, limiter last

The standard arrangement — clipper, then limiter — is not arbitrary. It follows directly from what each device is good at.

The clipper takes the isolated, short, tall peaks. These are the ones that cost the most headroom (a single sample can define your whole peak level) and are the cheapest to distort (they're transients, so the distortion is masked). It does this with no gain movement whatsoever, so the density it creates has no pumping in it.

What reaches the limiter afterwards is only what the clipper's ceiling and knee let through — occasional, small, and mostly the band-limited overshoot described above. A limiter that acts occasionally and by fractions of a dB is genuinely inaudible. That is the goal: the limiter as a safety net, not as a loudness tool.

Reverse the order and you get the worst of both. The limiter pulls gain down for peaks the clipper would have removed for free, so you hear gain movement and end up with the clipper working on an already-flattened signal where there is less transient left to mask its distortion. The density comes from the chain in front of the limiter, never from the limiter itself. If your loudness is coming out of the last device, you have built it wrong.

In the product, briefly

TRANSMIT Tower runs a clipper followed by a final true-peak limiter, and the clipper is active on every preset in normal mode — it is not something EXTREME turns on. What EXTREME changes is how hard it is driven and what domain it works in (chapter 4). Drive varies across the presets from a firm push down to a negative setting on the most conservative preset, where the stage attenuates slightly before shaping and effectively stays out of the way.

The clipper works at twice the sample rate internally and uses the band-limited error-correction approach described above rather than relying on the rate conversion for its cleanliness. The final limiter has a short look-ahead and an interpolated inter-sample peak estimate rather than a full standardised four-times-oversampled true-peak meter, with per-preset ceilings sitting between roughly −0.3 and −1.8 dBFS. If you need a certified true-peak number for a delivery spec, measure it with a meter built for that.