NER Model

A caption-quality evaluation model developed by Pablo Romero-Fresco and Juan Martínez Pérez for measuring the accuracy of live subtitling and respeaking. Unlike Word Error Rate, which penalises all errors equally, the NER model weights each error by how much it affects the viewer's comprehension: 'N' is the total number of words in the caption, 'E' stands for edition errors (deliberate edits or omissions introduced by the respeaker or software), and 'R' stands for recognition errors (mistakes from speech recognition or typing). Errors are scored as minor, standard, or serious depending on semantic impact, and an accuracy rate is computed as (N – E – R) / N × 100. Ofcom and several European broadcasters use the NER model as the reference standard for live-subtitle quality audits.