Table of Contents
Ask a vendor whether CART or AI captioning is "more accurate" and you will usually get a number back. Ask where that number came from and the conversation gets shorter. Most of the accuracy percentages attached to captioning, on either side of the human-versus-machine line, trace back to a vendor's own marketing rather than an independent test. That is not a reason to avoid the comparison. It is a reason to make it on the basis of what CART and AI captioning actually are, what they are each built to do well, and what happens when they fail — rather than on a figure that will not survive contact with a different vendor or a different room.
Two different processes, producing text that looks the same
CART — Communication Access Realtime Translation — is the instant translation of spoken word into text using a stenotype machine, a notebook computer, and realtime software, typed live by a trained human stenographer. The person doing it is listening, translating what they hear into phonetic steno strokes, and having those strokes converted to English text, all in the seconds between someone speaking and the words appearing on screen.
That skill is measured directly. The National Court Reporters Association's Certified Realtime Captioner exam requires candidates to caption a five-minute unfamiliar passage at 180 words per minute with 96% accuracy and no post-editing allowed. The Certified Realtime Reporter credential, which requires holding an RPR certification first, raises the bar to 200 words per minute at the same accuracy floor on live Q&A and testimony content. NCRA's own guidelines for working CART captioners set an ongoing target of 98% accuracy or better, "with a goal of 99% and above" — a working standard, not a guarantee, and one assumed to hold on clear audio and familiar subject matter.
Canada has no national equivalent certifying body. The only NCRA-approved court reporting program in the country is at the Northern Alberta Institute of Technology; two other Canadian programs exist without NCRA approval. In practice, most Canadian CART providers hold US NCRA credentials, because that is where the recognised testing standard lives.
What AI captioning actually does
AI or ASR (automatic speech recognition) captioning runs audio through a model trained to predict text from sound, with no person transcribing in the loop. There is no stenographer, no certification exam, and no fixed accuracy floor a vendor has to clear before shipping — accuracy is whatever that model produces on that audio, that day.
That makes ASR accuracy inherently harder to state as one number, because it depends on the vendor and the audio, not on a fixed skill. A 2024 academic benchmark tested 11 ASR vendors against 120 lecture and speech recordings and found word error rates ranging from 2.9% (the best-performing vendor, in batch mode, on clean audio) to 20.1% (the worst) — a roughly seven-fold spread on the same source material, reported in ACM Transactions on Accessible Computing. The same study found non-native and ESL speech pushed the average error rate up to 10.2%, and that streaming (live) transcription performed consistently worse than batch processing of a finished recording — an average word error rate of 10.9% live versus 9.37% after the fact, across vendors. That gap matters directly for captioning, which is almost always the live case.
A separate, peer-reviewed 2023 study in JMIR Mental Health compared three ASR services against a human transcription service on 65 real clinical interviews. Amazon Transcribe's median word error rate (8.9%) came close to the human service's (7.6%) — a small gap, though still statistically significant. Zoom and Otter.ai, tested on the same audio, came in at 19.2%. Three ASR products, one set of recordings, and a spread wide enough that "AI captioning" alone tells you almost nothing about which of those three you are getting.
Why neither side gets to claim one accuracy number
This is where marketing on both sides breaks down. 3Play Media, itself a captioning vendor, tested two unnamed competitors who advertised 99% accuracy "on all files" and measured their real output at 84.7% to 94.4% — five to fifteen points below the claim, on the vendor's own test. That is one of the only documented cases of anyone actually checking a captioning accuracy claim against reality, and it did not hold up.
The honest position is that CART accuracy depends on the individual captioner, the audio, and the subject matter, and that no independent study measuring real-world CART accuracy across providers appears to exist — NCRA's figures are certification and guideline thresholds, not field measurements. ASR accuracy depends on the vendor and the audio condition, and swings by fifteen to twenty points or more between vendors on the same material, as the studies above show. A single "X% accurate" figure, quoted without naming the vendor or service, the study that measured it, and the audio it was measured on, is not a data point. It is a marketing claim wearing a data point's clothes. Where this article cites a number, it names the study and what was actually tested — that is the standard worth holding any claim to, including your own vendor's.
Where CART is the right tool
Some settings do not tolerate a misheard word, and that is a reasonable place to stop optimising for speed and cost. Legal proceedings and depositions, medical appointments and clinical interviews, and formal human-rights accommodation are the clearest cases: a certified human producing a verbatim, accountable record — someone you can identify, and who stands behind what they typed — is worth its added cost and lead time precisely because the consequence of an error is not "an awkward caption," it is a wrong record.
This is also where Canadian accommodation law enters, separately from any accuracy comparison. Tribunals Ontario lists CART as a formal accommodation option alongside ASL/LSQ interpretation, and the Ontario Human Rights Code's duty to accommodate can require CART for a specific person regardless of what a website accessibility regulation does or does not exempt — we covered what the AODA actually requires, and does not require, for live captioning in more detail elsewhere, and it is worth reading if you are weighing an accommodation request against a general compliance obligation, because they are not the same test.
Where AI captioning wins decisively
None of that makes CART the right default for most events. AI captioning wins on latency, cost, and scale in ways that are structural, not marginal: software does not need to be booked days in advance, does not bill by the hour with a minimum call, and does not have a ceiling on how many rooms it can caption at once the way a finite pool of certified stenographers does. For a webinar, an internal town hall, a conference session, or a hybrid meeting where the goal is that most of the room can follow along without straining, that combination is usually what actually matters, and it is what most events need far more often than they need a certified verbatim record.
CART also has a real availability constraint in Canada that AI does not: with no national certifying body and most credentialed captioners holding US NCRA credentials, booking a qualified CART captioner on short notice, especially outside major centres, is a genuinely harder logistical problem than provisioning an ASR service.
How to choose
Start from consequence, not preference. If a transcript needs to stand up as a legal or medical record, or you are meeting a specific accommodation request, book CART and plan the lead time it needs. If the goal is that a live audience or a remote crowd can follow a streamed event or a meeting in real time, and the cost of an occasional misheard word is low, AI captioning is very likely the better fit on speed, cost, and availability alone. A meaningful number of organisations use both — CART for the sessions where accuracy is non-negotiable, AI captioning for the higher-volume, lower-stakes content around them — rather than treating it as a single either/or decision across every event they run.
If you are weighing this for a specific event, our transcription and captioning services page covers what live captioning and post-event transcripts involve, and our team can help you work through the specifics for your event, your audience, and what is actually at stake if a word gets missed.
Frequently Asked Questions
There is no single honest answer to that question. CART accuracy depends on the individual captioner, the audio, and the subject matter, and no independent study measuring real-world CART accuracy across providers appears to exist. ASR accuracy depends heavily on the vendor and the audio condition — independent benchmarks have measured word error rates from under 3% to over 20% across different ASR services on comparable material. Ask which vendor, on what kind of audio, measured by whom, before trusting any percentage from either side.
Communication Access Realtime Translation: a trained stenographer listens to speech and produces text in real time using a stenotype machine, a computer, and realtime software, typically displayed as an on-screen transcript or captions. It is a human service, not a piece of software.
There is no single Canadian rule mandating CART across every legal or medical setting, but it is the standard typically used where verbatim accuracy and an accountable, identifiable record matter — a certified stenographer produces a record they stand behind, which AI captioning does not. Human-rights accommodation processes, including through Tribunals Ontario, list CART as a formal accommodation option. If a specific proceeding or clinician has a stated requirement, follow that requirement directly.
Ontario's accessibility regulation does not specify a captioning method at all, and its live-captioning requirement (WCAG success criterion 1.2.4) is exempted for most large and designated public sector organisations. That is a separate question from the Ontario Human Rights Code's duty to accommodate, which can require CART for a specific individual regardless of that exemption. See our [full breakdown of what the AODA actually requires](/blog/aoda-captioning-requirements-ontario) for the detail.
For most webinars, conferences, and internal meetings, yes — it is the far more common choice, mainly on cost, speed, and availability rather than a claimed accuracy edge. Independent testing has also found that live ASR transcription consistently performs worse than the same software processing a finished recording, so if you are choosing an ASR vendor, test it against your actual speakers and audio setup rather than a demo reel before the event, not after.
Because a bare accuracy percentage, for either CART or AI captioning, is not a meaningful number on its own — it depends on which provider, which vendor, and what audio it was measured on, and the documented spread between best and worst cases is large enough (fifteen to twenty points or more in independent testing) that a single figure would mislead more than it would inform.