Abstract

Legislation defining caption accessibility states that automatically generated captions do not meet user needs and captions should be checked and corrected to be ‘verbatim’. Users of any auto-captioning service will have experienced errors, both minor and major, in the accuracy of the automatic transcription, and anecdotal evidence has suggested that some accents and the use of technical/academic language can further hinder the accuracy of the transcription.​

It is evident that automatic captions cannot be relied on to provide a suitable solution, particularly in meeting the ‘verbatim’ component of the WCAG success criteria. It is therefore important to review the accuracy of different automatic voice-to-text services, in order to give an initial transcript which can be manually corrected; considering the ease and time required in the workflow to produce fully accurate captions.​

How much work is required to produce accurate captioning, and are there ways to optimise the workflow? This work set out to review the accuracy of voice-to-text services and their suitability for captioning teaching content and decipher the most influential factors in captioning accuracy. Reviewing a sample of teaching videos including a mix of native/non-native accents, we tested the accuracy of 9 voice-to-text services (including 2 paid services).​

Participants will gain an insight into a range of, and the degree to which, different factors impact caption quality and strategies for developing the accessibility of video/audio based learning resources.​