Better Closed Captioning
Closed captioning is much more than just an FCC requirement; it is an important community service – providing access to the approximately 15% of Americans who struggle with hearing. Accurate, timely, and affordable closed captioning is imperative for broadcasters of all sizes. The current solution, human transcription, can be great. It can also be expensive and, as we discuss below, inaccurate.
AppTek has analyzed multiple samples of human closed captioning and we have found that the most frequent errors are the errors of omission. Some of these are harmless (e.g., using a pronoun instead of a proper name in an effort to save time), others are more egregious. Entire sections of the news can be omitted while the transcriber tries to catch up. The average sample we have analyzed has below 60% accuracy. That means the hearing impaired community is missing more than 40% of the content! The worst part of this situation is that, by definition, the hearing impaired community does not know what they are missing. The error of omission by the human transcriber goes largely unnoticed and unreported. This may keep broadcasters technically compliant with FCC mandates, but is an embarrassing disservice to the community at large.
The Whole Story
AppTek’s Live Closed Captioning Appliance uses machine learning and artificial intelligence to automatically produce closed captions with low latency and high-accuracy. Out of the box, our appliance can provide 85%+ accuracy – improving to over 93% with training. This is better than human captions even if you forgive the majority of omissions.
Our machine learning platform trains on the data collected by the device as well as data provided by the station to improve accuracy. Each station has a custom language model for their specific needs. We work with our partners to train a model for them before we launch the appliance live.