Automatic content recognition for radio identifies a station from its sound in one of two ways: a watermark is a hidden code added at the station, and a fingerprint is a pattern taken from the audio and matched against a library of what was on air. These methods answer one question, what was playing. Who heard it still has to come from a panel of people.
Key points:
- A watermark needs an encoder at each station. A fingerprint needs a recording of each station.
- Nielsen reads watermarks first and falls back on fingerprints. Barb's meters use fingerprints, and RAJAR's phone panel uses audio matching.
- Watermarks are hidden by masking, so quiet audio gives them less cover.
- Advert detection uses fingerprints too. That is monitoring, and it shows an advert ran, with no word on who heard it.
- Kenya's public record of passive measurement is a 400-person phone test in Nairobi in 2019.
What is automatic content recognition?
Automatic content recognition, or ACR, is any system that works out what is playing from the content itself. Nobody tells it the station.
It appears in three places: in audience meters, in smart TVs, and in the systems that log adverts. The audio science is shared, while the purposes differ. A device that identifies a song, a radio station and a commercial is doing the same job on different libraries.
What is an audio watermark?
A watermark is a code added to the audio before it goes out. People can't hear it. A meter can.
Arbitron's description of the Portable People Meter system sets out the parts. An encoder at the programming or distribution source inserts an inaudible identification code in real time. A station monitor checks that the over-the-air signal is carrying it properly, and warns engineers if it isn't.
Nielsen uses the same idea for TV, encoding watermarks into a broadcaster's audio and decoding them in panel homes, for live and time-shifted viewing.
How does a watermark stay inaudible?
Through masking. Louder programme sound covers the quieter code, much as a passing lorry covers a whisper.
A 2015 Radio World report on engineering papers explains the limits. The encoder adds the watermark only when its masking algorithms judge it won't be heard, and "during silence the encoder has no audio available for masking".
Watermark tones need about 400 milliseconds to decode cleanly. That means programme material matters. Speech with pauses and thin audio gives the code fewer chances than dense music, and a station's sound can affect how well it is credited.
What is an audio fingerprint?
A fingerprint is a compact summary of the sound itself. Nothing is added at the station.
The classic account is Avery Wang's 2003 Shazam paper. It marks peaks on a spectrogram, turns them into hash tokens and looks them up in a database. It could identify a short segment captured through a mobile phone in the presence of voices and other noise, and the paper names radio monitoring as an application.
That resistance to noise is the point. A phone in a pocket hears a poor copy of the broadcast, and a fingerprint system is built to tolerate that.
What is audio matching in radio measurement?
Audio matching is fingerprinting applied to live broadcast. RAJAR describes how MediaCell does it: an app grabs regular short samples of ambient sound, and they are compared with programmes airing at that moment on every RAJAR station.
An algorithm then decides whether there are enough matches of sufficient quality to credit a station. RAJAR says the science is similar to music-recognition apps such as Shazam.
The difference from a music app is the reference. A song library is fixed, while a radio library has to be recorded live, for every station, every minute. That is a big job.
How do watermarks and fingerprints compare?
Each one has a different cost and a different weak point. Neither is free.
| Watermark | Fingerprint or audio matching | |
|---|---|---|
| Where the work happens | At the station, before transmission | At the meter and at a reference centre |
| What the station does | Installs an encoder | Nothing, as long as someone records it |
| What the system needs | An encoder on every measured station | Live reference recordings of every station |
| What it handles badly | Silence and thin audio, which leave little masking | Stations carrying the same feed, which sound identical |
Arbitron was testing a "dual-function" meter that read codes and collected audio signatures as early as 2010, so that stations without encoders could be measured. Nielsen's current practice combines the two: watermark first, fingerprint when none can be detected.
Does passive measurement replace the panel?
It replaces the questionnaire. The panel stays.
ACR tells you what a device heard. To turn that into an audience you need devices carried by people chosen to represent a population, and weights to scale them up. Barb has 7,000 households for this reason and RAJAR about 5,000 people.
This is what buyers of passive technology tend to miss. Better identification doesn't fix a poor sample. People meters covers the panels themselves.
How does ACR work on smart TVs?
The set does the listening or looking. Nielsen lists ACR data from smart TVs among its big data sources, alongside return path data from set-top boxes.
This covers large numbers of screens with no panel recruitment. Its gap is the person, since the data doesn't register who is in front of the screen. That is why it is joined to a panel. Television audience measurement explains the hybrid.
How do advert detection systems use fingerprints?
Advert monitoring uses the same trick on a smaller library. The regulator's 2019 report describes Reelforge's process: an analyst loads a new advert, the software makes a digital fingerprint, and servers connected to antennas detect it every time it airs.
Presenter mentions can't be fingerprinted, because they are spoken live, so analysts log them by hand. The output is a list of what aired, and when.
Stations and advertisers often treat that as proof of reach. It isn't. It proves the spot went out, and it holds no information about listeners. See how a station proves its reach and media monitoring in Kenya.
Has passive radio measurement been tried in Kenya?
On a small scale, on the public record. TIFA Research presented a phone app called TIFAp+ in June 2019, built on audio frequency matching and calibrated to earshot, for Android phones with internet access.
Its Nairobi test ran through April and May 2019 with 400 media consumers, a stated margin of plus or minus 4.8%. It captured 75 radio stations and 33 TV stations, and found panel members touching an average of 5.2 radio stations a day.
The reference problem would be larger nationally. The Authority counted 265 FM stations and 253 TV stations operational at the end of March 2025, each needing a live feed in a system like this. Ipsos has also offered MediaCell for the Kenyan market.
Why do passive results differ from recall?
Because they catch what memory drops. TIFA Research concluded in 2019 that passive collection always captures more channels than recall, and warned against comparing the two.
RAJAR deals with the gap by calibration, matching its panel's listening levels to its diary sample so that all stations are treated equally. A market moving from recall to passive measurement should expect a break in its trend, and plan for one.
What happens to a station outside the system?
It disappears. A watermark system can credit only an encoded station, and a fingerprint system can credit only a station it records.
That is why Arbitron explored audio matching: to assure measurement of all broadcasts in a local market, even those not encoded. For a small FM or community station in Kenya, being in the reference set would be the first question to ask.
What about privacy?
Consent comes first. A phone that samples the room needs a clear promise. TIFA's panellists asked, "Do you listen to my conversation?", and its answer was that the app doesn't listen or store personal information.
Any Kenyan panel of this kind would need written consent and a clear account of what is sampled and kept. Data protection for audience research covers the duties.
What should a station do about passive measurement?
Understand it, and ask hard questions of anyone selling it. How many panel members sit in your county? Is your station in the reference library? What rule decides that a match counts as listening?
For most Kenyan stations the need this quarter is a fair estimate of who listened and why. That comes from asking people, as how radio listenership is measured explains.
Read next
- Audience measurement software for Kenyan media: the TI pack page for this work
- How media reach is measured in Kenya: all the pages in this section
- People meters and portable people meters explained
- How is radio listenership measured?
- Media monitoring in Kenya
- Takwimu Intelligence (TI) home page
- Audience measurement in Kenya: the guide to this library
- The TI platform and how we check our numbers
Where TI fits
Identifying a station's sound and understanding its audience are separate jobs. Takwimu Intelligence (TI) does the second by asking people: a cross-media day-after recall questionnaire for reach, and a programme panel for appreciation, weighted to KNBS census totals at county level with each margin of error shown.
To see it on your own station or title, book a walk-through.