Automatic Captions Quality

Since October 2021 the DEO have been working with students, academics and our suppliers to assess and improve the accuracy of the automatically generated closed captions within our Re/Play service.

Adding captions to media such as recordings of lectures provides a direct and tangible improvement to the learning experience of many students, and is essential for students who are D/deaf. This was recognised in legislation from 2018 (Public Sector Body Accessibility Regulations or PSBAR), which requires that all Re/Play content must be captioned to make sure it is accessible by meeting the international WCAG 2.1 AA accessibility standard. Pre-recorded video or recorded lectures which do not have captions do not meet this legal requirement.

Since September 2020, it has been possible to turn on automatic captions for all materials in Re/Play, which uses Automatic Speech Recognition (ASR) technology. This software’s ASR technology is based on three models - linguistic, acoustic and contextual events - and is specifically trained to work in our UK HE environment. Within this project, we will be turning on auto captioning for all recorded material in Re/Play using this ASR software.

However, there are concerns about ASR accuracy. From initial research, it’s clear that most students do not expect automatically generated captions to be perfectly accurate. Auto captioning is commonly used in platforms such as YouTube and MS Teams recordings, and we are all now more used to captions with minor errors in words or grammar, which do not affect their understanding. However, errors resulting in poor or unusable learning resources, or inaccuracies dramatically misrepresenting what someone has said, are not acceptable in an HE setting.

Therefore, this project aims to review the current quality of ARS captions, working with students to create a quality assurance process for sampling and assessing captions as a learning resource. We are also piloting student caption editors, who can work with specific units to hand-edit ASR captions where the ASR captions are known to be inaccurate (for example with terminology). Part of this project is to scope the viability of running a caption editing service such as this from with the Digital Education Office as a longer-term solution.

Captions via Disability Services

There is an established and well used system for providing captions by hand for students who require them, in collaboration with Disability Services, which ensures accuracy and provision of captions for those who rely on them as a learning resource.

I use captions whenever possible when I am using online recorded content. I find that captioning helps me follow the content and note down important information more easily.

Third Year UoB Student

How to improve caption accuracy

Project Aims

  • Establish an agreed way to assess quality of captions as a learning resource
  • Include the student voice in defining these criteria
  • Sample a selection of video content in order to benchmark the current quality
  • Better understand the impact a student caption editor can make to the accuracy of ASR captions over time for a specific subject area (ie ‘training the AI’), and define ways to measure this impact
  • Pilot the caption editor role, including how long it takes to caption, workflows and processes, what support they need, the impact they can make to a unit’s captions and viability of caption editors as a service
  • Establish a research project with UoB academics specialised in translation and subtitling
  • Work with other HEIs and relevant national stakeholders to develop standards and processes which meet accessibility standards for the sector