Stereo in broadcast

From the broadcast.audio wiki · Chapter 06 of 9

Part of a series on
Broadcast audio processing

Chapters

Reference

Contents

1Mid and side
2The FM twist: you lose the width exactly when you need it
3Correlation
4Correlation gating
5Why width costs loudness
6Archive material will hurt you
7In the product
8See also

In a studio, mono compatibility is a box you tick at the end. In broadcast it is a design constraint, because a very large fraction of your audience is genuinely listening in mono and you will never know which fraction.

Kitchen radios are mono. Portable receivers are mono. Phone speakers are mono. A single Bluetooth cube on a shelf is mono. A car with a dead door speaker is effectively mono. And on FM, receivers deliberately blend toward mono as the signal weakens — which we'll come back to, because it is the cruellest part of the whole arrangement.

So anything you do to the stereo image has to survive being summed. That is the rule the rest of this chapter is about.

Mid and side

The useful way to think about a stereo signal is not left and right but mid and side:

M = (L + R) / 2     S = (L − R) / 2 L = M + S     R = M − S

Mid is what both channels have in common — the mono sum, and the thing a mono listener hears. Side is the difference — the part that only exists because the channels disagree.

A centred vocal lives entirely in mid. A hard-panned guitar is half in mid, half in side. A stereo reverb tail is mostly side. And a signal that is identical in both channels but polarity-inverted in one is entirely side: full scale on a stereo meter, and silence in mono.

"Widening" almost always means raising the side gain relative to mid. That is a genuinely effective trick: it increases channel separation and makes an image feel larger. It also, unavoidably, moves the signal toward the condition in which it disappears on summing. Widen far enough and a mono listener hears a mix with the middle scooped out of it.

The FM twist: you lose the width exactly when you need it

Here is the part that makes broadcast stereo particularly unforgiving.

The FM stereo system sends the mono sum as ordinary baseband audio and the difference signal on a subcarrier (chapter 8). Recovering that subcarrier costs signal-to-noise — roughly 20 dB worse than the mono baseband under the same conditions. That is why a station that sounds clean in mono goes hissy the moment the receiver locks stereo at the edge of coverage.

Receivers deal with it by blending: as the signal weakens, they progressively reduce the side channel until they are effectively mono, trading the image for a usable noise floor. Almost every receiver made in the last forty years does this, and it is the right engineering decision.

But look at what it means for processing. Your careful stereo widening is the first thing discarded — in exactly the conditions where the listener is straining, at the far edge of coverage, in the car, on a hill, with the hiss coming up. The width was never there for the listeners who needed the help most.

Streaming does not blend, which removes this particular problem, and replaces it with the fact that a lot of your audience is on a single phone speaker instead.

Correlation

The measurement that tells you how much trouble you are in is correlation — a running estimate of how similar the two channels are, computed over a window of a few hundred milliseconds and read on a scale from +1 to −1.

+1 — identical channels. Perfectly mono; the side channel is empty.
+0.5 to +0.9 — a normal, healthy stereo mix: common content in the middle, differences at the edges.
around 0 — unrelated channels. Wide, and already losing substantial energy on summing.
negative — the channels actively oppose each other and material cancels on summing. Sustained negative correlation across a whole mix means something is wrong, not artistic.

A correlation meter is the cheapest mono-compatibility check in existence and I would not run a station without one in view. Brief negative excursions on a cymbal or a reverb are normal. A needle that lives below zero during choruses is telling you that a mono listener is hearing a hole.

Correlation gating

Which leads to a design principle that shows up in every serious broadcast processor: do not widen material that cannot afford it.

Widening is a bet that the signal has healthy stereo content worth exaggerating. On a mix with strong positive correlation, that bet pays. On an out-of-phase mix, a badly transferred archive recording, or something already artificially widened by whoever mastered it, the same processing makes the mono result substantially worse.

So the width control is made conditional: the processor measures correlation continuously and scales the widening by it. Above some floor, widening applies normally. As correlation falls toward that floor, the widening fades out. Below it, the processor does nothing at all. The user sets a maximum, and the material decides how much of it is used.

It is also common to band-limit the side signal before widening it. Very high frequency side content is where phasey, swirling artefacts live, it is the least reliable part of the image, and on FM it is the most expensive part of the subcarrier. Low-frequency side content is handled separately and generally removed altogether (chapter 7). So the widening is applied to a middle region where it does the most good and the least damage.

Why width costs loudness

This one is not obvious and it catches people out.

Widening does not change the mono sum at all — M is untouched. But the peaks that a clipper and limiter have to control are not M; they are L and R, which are M + S and M − S. Raising S raises the peak of at least one of those channels on every sample where S is non-zero.

So you have increased the peak level of the signal without increasing its perceived loudness one bit. Everything downstream — the clipper, the limiter — now works harder for the same result. In a peak-constrained chain, which every broadcast chain is, stereo width is paid for directly in loudness. A heavily widened signal will end up measurably quieter than the same signal at neutral width, or equally loud with more distortion. There is no free version.

On FM it is worse still, because the difference signal's contribution counts toward the transmitter's total deviation budget. Width is literally spent out of the same allowance as loudness, and this is why FM stations with the most aggressive loudness processing tend not to be the ones with the widest image. They cannot afford both.

Archive material will hurt you

If your library goes back more than a couple of decades, expect the following, and expect a correlation meter to find them for you.

Azimuth error. On analogue tape, if the playback head is not aligned to exactly the same angle as the record head, the two channels are read with a small time offset. A time offset is a phase error that grows with frequency, so the top end goes progressively out of phase between channels. In stereo it sounds slightly smeared; in mono the highs partially cancel and the transfer sounds dull and hollow. This is extremely common in archive digitisations and it is the single biggest source of mysterious correlation problems in an old library.

Pseudo-stereo. Mono recordings processed into fake stereo by comb filtering or all-pass networks. They sound spacious and they collapse catastrophically — often back to something worse than the mono original, because the filtering does not undo cleanly.

Polarity errors. A channel inverted somewhere in a digitisation chain. Correlation pinned near −1 on a whole track; almost total cancellation in mono. Rare, obvious once you look, trivially fixable, and invisible if you never check.

Hard-panned early stereo. Drums entirely left, voice entirely right, in the fashion of the mid-1960s. Mono-compatible in the strict sense — nothing cancels — but strange on headphones and worth knowing about before you widen it further.

There is no processing answer to most of this. The answer is to check the library, fix what can be fixed at the file level, and configure the processing so that it declines to widen material it cannot trust. A station with a deep archive usually ends up narrower than it would like, on purpose, and that is the correct outcome.

In the product

TRANSMIT Tower's stereo stage follows the rules above fairly literally. Width is boost-only — there is no setting that narrows the image, and the maximum varies by preset, with one preset (the speech one) leaving width flat entirely. What width there is gets applied to a band-limited side signal and is scaled by the measured correlation between your channels, so on poorly correlated material it simply declines to widen. Low frequencies are centred below a preset-dependent corner, which is chapter 7's subject.

The display shows correlation before and after the stage, which is the readout I actually watch. If the "before" number is unhappy, the problem is upstream of the processor and no amount of processing is the answer.