My company, Immetrica, and I personally have been building audiovisual audience measurement systems for 33 years, auditing existing systems, and analysing ratings data.
In late 2017 we realised that as viewing fragmented among various platforms and devices, existing systems were becoming progressively blind to parts of overall viewing. This affected conventional people-meters (which could not measure VOD, OTT, or mobile devices), RPD (return path data) systems from pay-TV operators (no viewing outside each pay-TV system, usually no integrated VOD, no channel-operated VOD, no OTT, no mobile), OTT systems’ integrated measurement (nothing outside the system), and connected/smart TVs (nothing on other devices, usually no measurement of the big standalone services). In short, ratings services were not keeping up with how people watched television and video, and offering ever less complete accounts of viewing.
Networks and channels were unable to monetise an increasing number of impressions and viewers reached. Advertisers could not find target viewers on the newer platforms. Programme creators were denied feedback from their audiences.
Our solution was ACR fingerprinting by a code library in an app on a mobile device, usually a smartphone (for children without smartphones, the app could be installed on a tablet). This could theoretically capture any audio within hearing of the smartphone’s microphone, which survey research has shown to be in the vast majority of instances also within hearing of the viewer. We were not the first to think of this approach, but previous implementations all suffered from various severe flaws, chiefly poor ACR recognition and high costs.
Immetrica’s work on another project in 2015 compared detailed results in controlled conditions among several ACR fingerprinting providers and found substantial differences among the quality of recognition and pricing. Similar disparities among other ACR providers were found by other evaluators.
It was our good luck that our deployment plans concerned a price-sensitive part of the world where the best-known providers’ pricing was a nonstarter. Thus we found ACRCloud, which turned out to have among the best ACR implementations and also among its most affordable.
The initial results were superb: reliable recognition with almost no skipping of recognition cycles (in our implementation, a cycle is 10 seconds) with the phone in a protective pouch inside a coat pocket; recognition with little skipping at a distance of two meters in another room with the door open; reliable recognition of two concurrent sources with noise also present; recognition in an otherwise quiet house with a floor/ceiling between the source and the smartphone. There was even some correct recognition of ten concurrent sources–not tolerable by a human being. Multiple sources have confirmed that ACRCloud recognition is superior to that of many other providers, and of all regarding which we have become privy to such comparisons. Looking at the results, Bill Harvey, one of the most respected leaders of the audience measurement industry in the US, remarked that the constancy of recognition was more like that of RPD (which has no ACR but is fed by the set-top box’s tuner) and “the world’s best ACR”.
To be sure, great and affordable ACR is not sufficient for audience measurement. There has to be a practically unkillable library that ensures continuous measurement even during mobile device sleep. Otherwise the cooperation level (portion of time during which measurement is available) will be too poor to be useful, certainly for programme, rather than advertisement, measurement. Time has to be independently sourced, as the real-time clocks of many devices and even entire operators are unreliable. And there has to be a custom processing system. Without it, because of traits specific to ACR recognition, the results would be rubbish, and not even in the correct format. For competitive reasons, I cannot explain in greater detail, but if you want to know more, contact me at Boris (at) Immetrica (dot) com .
Here, again, the ACRCloud implementation deserves praise for its design. We can, for example, route our added data fields, such as various timestamps, through the ACRCloud system. Or we can reverse the flow and have the data come to us first, for subsequent recognition by ACRCloud. The latter route permits us to build redundancy that protects us from system failure at much lower cost than otherwise. And a very important consideration is ACRCloud’s ability to add preconfigured packages of computational resources as needed, to cover the maximum possible load at a given time. This again means cost savings as it is unnecessary to overspecify the system by a large margin in advance.
Overall the cost of ACRCloud is at least several times lower than the nonintroductory pricing of its competitors. Some of the largest and most expensive providers deliver an ACR quality far inferior to that of ACRCloud.
Another advantage of ACRCloud, which not all ACR providers have, is the clean set of API calls that cover every conceivable purpose. We use them to automate advance advertising content fingerprinting, OTT services’ content library automated ingestion at scale, signal monitoring from ACRCloud’s point of view (are fingerprint streams being received?), and configuration of realtime channel ingestion. This is just the start of the huge number of possible enhancements and automations that the ACRCloud API calls enable.
Finally, no ACR recognition will work well with substandard reference signals. The big hazards here are commercial IPTV services, especially the multinational kind. Their streams often fail after a few seconds, fail entirely for days or weeks, or are mislabeled. Many streams produced by broadcasters themselves are more reliable although some change URLs every few hours. On any channel received via stream, expect latency of between 10 seconds and 3 minutes. The most reliable reference signals are those obtained in the service territory through cable TV systems or terrestrial reception, but be sure to use an all-digital path: the analogue you get from the RCA sockets on the STB is sourced before the error correction using the correcting codes embedded in the digital data, and it is a struggle to get any recognition from those analogue sources at all (although it may work perfectly in the lab). Again, contact me for help implementing functional reference monitoring.
It was an indescribable feeling when our system was fully debugged and our results in a direct comparison (conventional TV in realtime only) tracked almost perfectly with those of the currency ratings provider. Sure, this result required much work besides fingerprinting. But without ACRCloud’s excellent fingerprinting, this result would not have been possible despite our best efforts.
Thank you, ACRCloud people, Peng, Olym, Pony and the others, for your devotion to your work!
Boris Levitan Founder and CEO Immetrica, Inc. Boston, Massachusetts, USA Boris (at) Immetrica (dot) com +1 857 891 4000