Live captioning vendors and the accuracy threshold that matters

Last January I watched a cardiology society’s annual meeting spend $14,000 on a captioning setup that still delivered a five-minute lag during the keynote. The vendor’s sales sheet promised 99 percent accuracy, which sounds great until you’re staring at a screen full of misspelled drug names and the speaker is already three slides ahead. I’ve booked live captioners for roughly forty healthcare and finance events since 2017, and I can tell you the accuracy number that actually matters isn’t the one on the marketing page.

The industry standard you’ll hear quoted is 95 to 98 percent accuracy for real-time captioning. At 95 percent, you’re serving your audience 300 incorrect words on a 6,000-word keynote. At 98 percent, it’s 120. Neither is acceptable when the speaker is naming a clinical trial or discussing a FINRA filing, so I push for vendors who contractually commit to a 98 percent floor on live sessions, measured against the actual event audio, not a studio demo. Most of my pharma clients cap dinner around $75 to $125 a head, and some finance firms that serve alcohol cut off bar service before the evening general session starts, and the captioning vendor needs to know those timings on the front end.

How I vet a captioning vendor before I send the RFP

I don’t start with price. First I ask for a sample transcript from a live event, not a YouTube video they captioned after the fact. Then I ask who actually does the typing. Is it a human CART provider, a stenographer with a real-time dictionary, or is it an AI engine with a human editor watching from a lag of thirty seconds? A stenographer builds a personal dictionary for your speaker ahead of time. An AI system can’t be trained on the spot when your presenter starts quoting obscure Latin anatomical terms. For medical meetings, I will not book a pure AI solution unless a human is actively correcting within two seconds.

I also ask how they handle the speaker’s slides. Some vendors will request the slide deck a week early and pre-program the technical terms. That alone jumps my accuracy from the mid-90s to the high 90s in practice. One vendor in Orlando told me they had their own proprietary medical dictionary, and they did, but it didn’t have the name of the specific surgical device the keynote surgeon had invented. We had a back channel where my production manager typed the term into a chat window for the captioner mid-speech.

The contract language matters more than the sales call. I have a standard rider now that specifies a latency cap of three seconds from spoken word to on-screen text, and it defines accuracy as the percentage of words captioned exactly as spoken, not including filler words. I negotiate a service credit if they breach the threshold. I’ve had three vendors walk away from that clause, and every one of them had something to hide about their live performance.

What the accuracy threshold actually means for your content mix

The threshold you need shifts based on what’s on stage. A breakfast keynote from a motivational speaker can survive a 96 percent captioning rate. A finance panel discussing a recent SEC rule interpretation cannot. For my budget-conscious clients, I often book a basic automated caption feed for the general keynotes and then spend the real money on a professional human CART provider for the breakout sessions that are actually being recorded. That split approach has saved clients as much as $6,000 on a two-day event.

I have a rule of thumb based on the session type. If the content will be edited into a highlight reel or shared with regulators later, I need the highest possible accuracy. If it’s a town hall where people are mostly listening for the broad strokes, I let the automated system do its thing and accept the occasional word salad. There’s no point paying for a stenographer to caption a trivia night at the welcome reception.

The content mix also drives how I seat the captioning team. I learned that the hard way at a theater venue in St. Louis where the only available power outlet placed the captioner right next to the HVAC vent, and the audio feed picked up a constant hum for the first hour. Many performing arts venues have surprisingly good production infrastructure if you know how to access it, so when I’m scouting a site, I ask about their audio patch points before I even discuss the catering minimum. I’ve run a successful captioning setup for a 200-person investor day in an industrial loft where the venue staff helped me run a clean XLR cable across the ceiling.

Working the live feed and handling the inevitable glitch

On show day, I set a fifteen-minute pre-session test with the actual speaker audio. I sit in the back row and watch the captions while listening to the audio. If there’s a discrepancy, we fix it right then. It costs me fifteen minutes of setup time, and it has caught real problems in about a quarter of my events. One such fix was a vendor whose feed was pulling from the house mix rather than the podium mic, which meant the captions picked up the audience laughter and missed the speaker’s quieter asides entirely.

When the session goes live, I assign a dedicated runner to the captioner with a direct chat line. The runner’s only job is to feed proper nouns, acronyms and product names to the captioner before and during the session. A human can see a speaker jump around a slide and type a correction to the captioner before the next sentence lands. This costs me one extra staffer per room, roughly $250 a day including labor burden.

A glitch will still happen. I had an event in Boston where the vendor’s cloud connection dropped for a full ninety seconds in the middle of a panel on compliance risks. My runner caught it immediately, a paper sign went up on the screen table with a CART Notice asking the audience to use the companion app captions instead, and we bridged to the secondary feed. My takeaway was to insist on a local machine as a backup recorder, so the vendor has a secondary source even if their cloud feed fails.

Running the numbers on quality versus the invoice

Captioning vendors price themselves anywhere from $150 an hour for a basic automated feed to $350 an hour for a certified CART provider with a stenography license. For a full-day general session that runs eight hours, I budget $2,800 to $3,500 for the premium tier. Add in the production feed, the pre-session test and the runner, and the total line item lands near $4,200 for a day of highly accurate live captions. That’s roughly 3 percent of a mid-sized corporate event’s total budget.

I have an upcoming healthcare client meeting in the Orlando area, and because that metro has a deep bench of AES and CART professionals, I expect to land a high-quality vendor in the $320 an hour range rather than the coastal agency markup. I’d estimate a comparable vendor in Manhattan or San Francisco runs 15 to 20 percent higher for the same certification tier.

Accessibility compliance intersects with captioning quality in ways planners often ignore. The ADA itself doesn’t demand a specific accuracy percentage for live captioning, but plaintiff’s attorneys do read the FCC’s guidance on closed captioning for broadcast content, which generally requires a high degree of accuracy, and they will apply that lens to a recorded event posted publicly afterward. Most corporate counsel I’ve worked with treat anything under what the vendor sold you as a potential liability if that recording is shared externally.

What I ask every client before I draft the captioning RFP is simple: what is the headcount for the sessions that need captions, what are the dates, and will any session be recorded for later distribution? If the recording is for internal training only, I can flex the accuracy target down a point and save you money. Send me the brief and I’ll tell you which vendor tier is genuinely necessary based on the room size and the compliance stakes.