The NER model was developed by Pablo Romero-Fresco of Roehampton University and Juan Martinez, a respeaking consultant. ‘Respeaking’ is the term used for a captioner repeating the dialogue of a TV program or other medium into a microphone, which is then turned into captions by text-to-speak software. In the NER Model, N stands for the number of words in the respoken text, E for ‘Edition’ errors introduced by the respeaker, and R for ‘Recognition’ errors caused by the software. The model makes allowances for the seriousness of errors, and produces an accuracy rate expressed as a percentage.
The model is currently being trialled by the UK communications regulator Ofcom, which recently announced that it will be requiring all broadcasters to measure the quality of their captions. It is also being used in several other countries including Germany, Spain and Italy.
In a media release, Ai-Media’s independent caption auditor Robert Scott said, “We have chosen the NER model because it is anchored in robust independent consumer research and produces quantitative scores that are consistent with viewers’ quality perceptions.”
Levels of live captioning, and in particular captioning created using respeaking, have been rapidly increasing around the world in recent years. This has led to growing concerns about caption quality, and the NER model is one of several that have been developed to measure this in a meaningful way.
In a white paper to be published in early 2014, Media Access Australia will look at these various models and their pros and cons, and the different approaches to maintaining and encouraging caption quality being taken by regulators around the world.
Top of page