NER model

The NER model is a method for determining the accuracy of live subtitles on television or at events that are created with speech recognition . The three letters stand for n umber, e dition error and r eCognition error. It is an alternative to the traditional WER model ( W ord E rror R ate, word accuracy ).

The NER model contains a formula for determining the quality of live subtitles: an NER value of 100 means that the content was reproduced perfectly correctly. For the calculation, the total number of words in the live subtitles is taken and the editing and recognition errors (caused by poor speech recognition) are subtracted from this. That number is divided by the total number of words in the live captioning and multiplied by a hundred.

${\displaystyle NER-Wert={\frac {N-E-R}{N}}*100}$.

Thereby means

• N (number) = total number of words in live subtitling
• E (Edition error) = editing error
• R (Recognition error) = recognition error

In Switzerland, this measurement method is already used in public television. Other countries have also shown interest.

The traditionally used WER model, on the other hand, is more static because it simply measures the literal deviation of what is said from what is written, without taking into account that there may be live edited subtitles.

See also

• Word recognition rate

literature

• Pablo Romero-Fresco: Subtitling through Speech Recognition: Respeaking. Manchester: St. Jerome 2011, ISBN 9781905763283