How do we know a radio station drew two million listeners on a given morning? For a long time the answer came from listening diaries and phone surveys, self-reported and imprecise. For about twenty years, part of the measurement has rested on a far more discreet technique: an audio watermark inserted continuously into the broadcast sound. Inaudible to the listener, it is read by a small meter a panellist carries, and it tells to the second which station was on air. This is signal engineering, and it lives in the broadcast chain that radio techs know well.
A code hidden under the programme
The point of an audio watermark is to hide digital information in sound without altering it to the ear. The most widespread system in radio, Nielsen’s CBET (Critical Band Encoding Technology), exploits psychoacoustic masking: when a loud sound is present in a frequency band, the ear no longer hears the quiet sounds next to it. The encoder slips its code precisely into those masked regions, at levels the programme covers. The code carries a station identifier and a timestamp, repeated in a loop, so that a handful of seconds of listening is enough to recognise the station.
Where the encoder sits
The encoder’s position in the chain is an engineering topic in its own right. It is installed ahead of the on-air processing, because a heavy on-air chain, with multiband compression and a limiter, can crush or distort a code inserted too late. The watermarked signal then passes through the processor, the studio-to-transmitter link and the transmitter, and the code must survive all of it. The loudness and processing settings directly affect its robustness, which makes watermarking a concern for the whole technical operation, not just the research department.
The meter in your pocket
On the listener’s side, a panellist recruited by the research firm carries a small receiver, the Portable People Meter in the Nielsen system. The device does not capture what the person says, but what they actually hear, at home, in the car, in a shop. It detects the codes present in the surrounding air, timestamps them and sends them each night to the processing centre, which reconstructs listening station by station. Measurement becomes passive: no diary to fill in, no memory to rely on.
The weak point, silence and speech
Psychoacoustic masking has a catch: you need programme to hide the code. On dense material, energetic music or sustained speech, the watermark tucks in easily. On a quiet passage, a pause, sparse speech or a pianissimo classical piece, less sound remains to mask the code, and detection weakens. A heavily spoken station can therefore under-report its real audience.
It is this flaw that gave rise to devices such as 25-Seven’s Voltair, meant to monitor and strengthen the robustness of the encoding. Their arrival set off a fierce controversy: by pushing the code insertion, was there a risk of artificially inflating measured audiences rather than simply making them more reliable? The debate stirred the US industry and underlined an obvious point, a measurement tool that affects what it measures must be handled with care.
Different methods from country to country
Watermarking is not universal. It dominates radio measurement in the United States with Nielsen Audio, and also serves, in other forms, for content identification and rights tracking. Other markets still rely on self-reported panels, listening diaries or recognition by matching the captured sound against a reference fingerprint, an approach close to but distinct from watermarking. Each method has its biases: self-reporting suffers from memory, watermarking from masking, fingerprinting from capture quality. None measures listening perfectly; all seek the best trade-off between accuracy and cost.
My take
Audience watermarking is a fine example of invisible engineering: you have to hide data in sound without ever damaging it, and make it robust enough to cross an entire broadcast chain. For a radio tech, the key is knowing that one link in the chain works for the research department, and that processing that is too heavy or badly placed can get in its way. The Voltair controversy marks the system’s limit: as soon as a measurement carries commercial value, the temptation to optimise it appears. An honest code, well inserted and well monitored, still remains the most reliable way to know who is really listening.
