Deaf and hard of hearing individuals found ASR to be useful in low-stakes contexts, such as small group planning for recreational events. A low-stakes setting means that mistakes are inconsequential or can be easily rectified. 

ASR in Low-Stakes Settings

Some Deaf and hard of hearing users reported that they found ASR-produced transcripts to be physically draining because they had to expend more energy into filling in blanks of missed words or making sense of inaccurate transcriptions. High-stakes settings are situations in which inaccurate transcriptions may lead to large and long-term consequences. Examples include exam preparation, work training, job interviews, and health and safety informational seminars. As such, ASR at its current state may not be ideal for usage in a high-stakes setting. 


ASR in High-Stakes Settings

In a study conducted by researchers from Georgia Tech’s Center for Inclusive Design and Innovation, five ASR-produced transcripts of five 30-minute lectures containing about 16,000 words were compared to a human-written baseline transcript to determine the accuracy of ASR technology. Word-for-word accuracy between ASR transcripts and the human baseline transcript only showed a 69% - 88% level of correctness in ASR technology. As ASR-produced transcripts are not fully accurate, they do not create an equitable online environment for Deaf and hard of hearing individuals. 


Comparative Analysis of 5 ASR Transcripts


The "best guess" design of ASR can lead to cases of Deaf and hard of hearing individuals taking in the wrong information. Because ASR precisely matches verbal speech word-for-word, it will add in every filler word spoken, such as: uh, um, er, ah, like, right, okay, you know. Faithful and inflexible transcription of filler words and unusual speech patterns may be distracting for users of ASR.

Limitations of ASR

Through the use of artificial intelligence, Automatic Speech Recognition (ASR) is designed to capture every sound spoken, and will display a word for each utterance without regards to its accuracy or whether it makes sense in context. In order to do this, ASR makes a “best guess” for words and phrases. 


University of Michigan - Ann Arbor

A program has automated captioning, or Automatic Speech Recognition (ASR), if the text is generated by a speech recognition or translation program instead of a person. This is often used on social media websites, such as YouTube and Facebook, due to the vast amount of visual content on these websites. 

Automated Captioning

Ballenger, S. (2022). Access for deaf and hard of hearing individuals in informational and educational remote sessions. Assistive Technology Outcomes & Benefits, 16(2), 45-55. Retrieved from https://proxy.lib.umich.edu/login?url=https://www.proquest.com/scholarly-journals/access-deaf-hard-hearing-individuals/docview/2712291965/se-2


Access for Deaf and Hard of Hearing Individuals in Informational and Educational Remote Sessions

Since this type of captioning relies on technology that may struggle with accents and vocal inflections, automated captioning can often be poor, nonsensical, and unreliable. These captions also often leave out important non-verbal sounds, meaning that content may not be fully captioned. 

Problems with Automated Captioning

During the COVID-19 Pandemic, most American schools moved to an entirely virtual course model via the usage of video lectures. To comply with federal regulations for federally funded institutions, all lectures would need to be captioned. Since it would be too time and cost intensive to manually create captions for each lecture, automated captions are often used instead.

Learn Before

Related

Learn After