Question · Accessibility and health
Is automatic captioning accurate enough for deaf people?
Automatic captions are accurate enough for casual viewing, but not on their own where a missed word matters. Accuracy varies widely between products and recordings, and the streaming mode behind live captions measured lower in quality. Trained human captioners still lead in the research. Test any system on the speakers and rooms you use, and book a human captioner for high-stakes moments.
Updated October 16, 20264 min read8 cited facts, each traced to its source
Automatic captions are accurate enough for some uses, such as casual viewing of television. They are not accurate enough to rely on alone where a missed or wrong word would cost you something, such as a lecture, a medical visit or a legal hearing. The answer depends on the setting, and on whether the captions were produced live or from a recording, so no single accuracy figure settles it.
The answer depends on the setting
For watching television, the evidence is encouraging. A 2026 online survey of 216 validated deaf and hard of hearing participants found that live TV captions and automatic captions synchronized with the audio were rated similarly 1. That is a different question from a lecture or a clinic. The survey used clips of live television, so it did not test conversation or phone calls.
Conversation is harder. In a 2017 study, deaf and hard of hearing people named text accuracy and lag time as the main obstacles to using speech recognition in group conversation 2. A caption that is mostly right can still be unusable if it arrives after the moment has passed or gets the one word that changes the meaning.
Accuracy varies from one service to another
A 2024 study tested eleven common speech recognition services on university lecture recordings. It found that accuracy "ranges widely between vendors and for the individual audio samples" 3. So a claim about one product says little about another, and a good result on one recording does not predict the next one.
Live captions are the weaker case. The same study found significantly lower quality for streaming recognition, the mode used for live events 4. A figure measured on recorded audio does not tell you how the live captions from the same product will perform, so ask which mode a figure describes. The study used lectures, not conversation, so its numbers do not transfer directly to a family dinner or a meeting room.
Word error rate does not predict what deaf readers find usable
The usual measure of automatic accuracy is word error rate, which counts every wrong word the same way. A 2017 study with 30 deaf and hard of hearing participants tested a new metric called ACE, which the authors proposed for judging caption quality. Its correlation with the participants' ratings was 0.74 5. Word error rate correlated only 0.11 with the same ratings 6. The two statistics are not identical, so compare them with care, but the gap is the point: a low word error rate does not guarantee captions that readers find easy to use.
A 2026 survey found the same pattern. The standard caption metrics, including word error rate, tracked viewer ratings moderately to highly for TV captions, but far less so for automatic captions 7. The authors conclude these metrics are not technology-neutral, so a figure that predicts quality for one kind of caption may not predict it for another.
Trained human captioners still lead in the research
When a missed word matters, the standard is still a trained human captioner. The authors of a 2024 study report that existing approaches for improving automatic recognition still fail to perform at par with trained CART captioners 8. Human captioning has its own limits, and you need to book it ahead of time. For the service itself, see CART captioning.
What to do before you rely on automatic captions
- Test with your own speakers. Include the accents, names and technical words you meet every week, because one good demonstration is not enough.
- Test in the real room and on the real device. Background noise and microphone position change the result.
- Look for meaning errors. A wrong name, number or "not" can change the meaning, while a dropped filler word usually does not.
- Check the delay. Captions that arrive late are read after the moment has passed, so watch whether they keep up with the speaker.
- Ask what the figure measured. Find out whether it covers live or recorded audio, and which language and conditions it used.
- Book a human captioner for high-stakes moments. For a lecture, a medical visit or a hearing, ask for a human captioner in advance rather than relying on a phone's automatic captions.
Automatic captions are a useful everyday tool, but test them before you depend on them. For the wider set of options, see the guide to speech to text for deaf and hard of hearing people. For captions in person, read about speech to text glasses, and for phone calls, read what a captioned telephone is. The accessibility and health overview lists the rest of the guides.
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
Go deeper
Related questions
- Do speech to text glasses work for deaf people?
- What is a captioned telephone?
- Can a pinched nerve in the neck feel like carpal tunnel?
- Can Access to Work pay for dictation software?
- Can ADHD affect a child's speech?
- Can carpal tunnel come back after surgery?
- Can carpal tunnel go away on its own?
- Can carpal tunnel syndrome be cured without surgery?
Referenced by
- Speech to Text for Accessibility, RSI and DisabilitiesTopic
- Speech to Text for Deaf and Hard of Hearing PeopleGuide
- Are automatic captions good enough for accessibility rules?Question
- Do speech to text glasses work for deaf people?Question
- Is there Live Transcribe for iPhone?Question
- What is a captioned telephone?Question
- CART captioningGlossary