A viewer watches your evening bulletin. The anchor says a court ruling will not take effect this year. The caption drops one word. Now the ruling takes effect this year.
The transcript is 91% correct. The story is 100% wrong.
That one dropped word is why caption accuracy is not a quality nicety. It sits between a broadcast that informs and one that misinforms, between an audit you pass and a complaint you answer. Which leaves media teams asking the same two questions. What does accurate actually mean, and how do you measure it?
What caption accuracy actually means
Accuracy is not just spelling. Since 2014, the FCC has held US television captions to four quality standards under 47 CFR 79.1(j)(2): accuracy, synchronicity, completeness and placement.
Accuracy means captions match the spoken words in the order spoken, including speaker IDs, sound effects and music cues. Synchronicity means captions land with the audio at a readable rate. Completeness means they run start to finish, including the cold open nobody remembered to send. Placement means they never cover a scoreboard or a lower third.
Miss one and the program is non-compliant, however clean the word-level transcript looks. Our breakdown of FCC caption rules for TV and streaming covers who carries the obligation at each stage. The bar also keeps moving: from 17 August 2026, device manufacturers and MVPDs must make caption display settings readily accessible, judged on proximity, discoverability, previewability, and consistency.
What breaks when captions are only mostly right
Ask a compliance lead what a bad caption costs and you get a legal answer. Ask content operations and you get a longer list.
- Compliance exposure. Complaints trigger records requests, corrective action logs and unbudgeted remediation.
- Audience loss. Captions serve deaf and hard-of-hearing viewers first, sound-off mobile viewers second. Both leave when captions garble.
- Broken search. A wrong player name means that clip is effectively lost to your archive.
- Bad downstream data. Captions now feed recommendations, contextual ad targeting and AI summaries. Errors compound quietly.
A caption file is no longer just an accessibility artifact. It is the text layer your whole operation reads from.
So what is word error rate?
Word error rate, or WER, is the standard metric for scoring automatic speech recognition against a verified human reference transcript.
WER = (S + D + I) / N
- S = substitutions, where a wrong word replaces a right one
- D = deletions, where a spoken word is missing
- I = insertions, where a word appears that nobody said
- N = total words in the reference transcript
A WER of 4% means four errors per hundred reference words, which vendors market as 96% accuracy. Lower is better. Zero is theoretical.
A worked example, because the maths matters
Reference transcript, 12 words:
The senator said the bill will not pass before the winter recess
ASR output:
The senator said the bill will pass before the winter recess
One deletion. S = 0, D = 1, I = 0, N = 12. WER = 1/12 = 8.3%, or 91.7% accuracy.
By the numbers, respectable. In practice, the caption says the opposite of what the senator said.
Now flip it. Had the engine inserted a stray “um” instead, the WER would be identical. Same score, entirely different consequence. That gap explains most of what goes wrong when teams treat one percentage as a quality guarantee.
What counts as a good WER
There is no single number, which frustrates procurement. Thresholds should follow content type and risk.
| Content scenario | Typical raw ASR result | Realistic target after QC |
| Scripted studio audio, single speaker | 3 to 8% WER | Under 1% WER (99%+ accuracy) |
| Prerecorded long-form, mixed speakers | 8 to 15% WER | Under 1% WER |
| Live news with re-speaking | Variable | 98% NER, the Ofcom live subtitling threshold |
| Sports, panels, crowd noise, accents | 15 to 30% WER | Set by error severity, not a flat number |
Two things to keep straight. The 99% figure is an industry benchmark, not an FCC rule: the Commission declined to set a numerical threshold and instead applies a de minimis test weighing the type of failure, whether it recurred, how understandable the program stayed and how fast it was fixed. And human parity on conversational speech benchmarks has historically sat near 5 to 6% WER, so “better than human” claims deserve a look at the test set.
Where WER falls apart
WER treats every word as equally valuable. Your audience does not.
Dropping “not” from a legal ruling scores the same as dropping “the” from a weather report. Fumbling a sponsor’s brand name scores the same as fumbling a filler word. WER also punishes paraphrase, which makes it a poor fit for live re-speaking where captioners condense deliberately to keep pace.
That is why the NER model exists. Developed by Dr Pablo Romero-Fresco at the University of Roehampton, it scores accuracy as (N – E – R) / N, where E covers edition errors from omission or condensing and R covers recognition errors, each weighted by how much it damages comprehension. Ofcom and Canada’s CRTC both treat 98% NER as the live subtitling threshold.
NER sits closer to the viewer’s experience. It is also slower and more subjective, and scoring a program can take ten to fifteen times its runtime. Neither model is sufficient alone.
The metrics worth knowing
| Metric | What it measures | Best used for |
| WER | Substitutions, deletions, insertions per reference word | Benchmarking ASR engines, prerecorded QC |
| FER | Formatted error rate: adds punctuation, casing, speaker labels | Broadcast-ready caption files |
| NER | Comprehension-weighted errors | Live captioning and re-speaking |
| CER | Character-level errors | Names, numbers, non-spaced languages |
| DER | Diarization error rate, speaker attribution | Panels, interviews, debates |
Speaker attribution deserves its own check. A file can score beautifully on words and still hand the guest’s quote to the anchor. Our guide to evaluating speaker diarization in media workflows covers testing those as separate capabilities.
What actually drives caption errors
Most caption failures trace back to the audio, not the engine. Overlapping speech in debates. Stadium noise. Rapid anchor delivery. Accents the engine was never tuned for. Proper nouns, tickers, drug names, player names. Mixed-language segments. Encoder drops mid-feed.
Fixing the input is usually cheaper than fixing the output. Clean multi-channel capture, vocabulary biasing toward the terms that recur in your content, and a short delay buffer on live workflows will move your numbers further than switching vendors. We go deeper in our guide to transcription QC for noisy live content.
How to set accuracy thresholds that hold up
Stop chasing one number across all content. Define acceptance criteria by use case.
- Build a gold-standard reference set from your real content mix, worst audio included
- Score WER and FER separately, since formatting failures fail conformance on their own
- Put names, numbers, negations, sponsor mentions and compliance phrases in a zero-tolerance tier
- Use NER for live targets, WER plus FER for prerecorded
- Measure synchronicity as latency, with 1 to 3 seconds a common live streaming target
- Log every caption failure and its corrective action the same day
- Re-benchmark quarterly, because content mix and engine versions both drift
Why hybrid workflows still win
Automatic captioning has improved enormously. It has not removed the review layer, and pretending otherwise is how teams end up in complaint queues.
The workflow that holds up runs in four stages. ASR produces a first-pass transcript. Automated QC checks formatting, reading rate, line length and placement against a style guide. A caption specialist reviews accuracy, tone, speaker labels and context. A final conformance check validates the file against the destination platform spec.
Human review is not evidence that automation failed. It is the control that makes automation safe to scale.
Common objections, answered
“Our vendor guarantees 99% accuracy.” Ask which model produced the number. A 99% NER score permits condensing that a WER-based contract would flag.
“Manual captioning is more accurate.” At low volume, possibly. At scale, manual-only workflows fail on turnaround, cost and consistency long before they fail on accuracy.
“AI accuracy is too risky for compliance content.” That argues for governed AI with human review, not for avoiding automation. The risk lives in unreviewed output.
How Digital Nirvana approaches caption accuracy
Digital Nirvana was built around this exact problem: getting machine speed and human judgment into one workflow without trading either away.
TranceIQ handles cloud transcription, caption and subtitle generation, translation and conformance checking, with human review inside the pipeline rather than bolted on after. Media Enrichment supplies the caption specialists who apply your style guide and clear the severity-weighted errors no automated score catches. The underlying speech and vision capabilities sit in MediaServicesIQ, available through APIs so accuracy work happens inside the systems your team already runs. The same stack supports live captioning for broadcast events with latency and conformance monitoring, while MonitorIQ confirms what actually aired, which is the evidence you need when a complaint arrives.
Where caption accuracy connects to everything else
Accurate captions are the cheapest metadata your operation will ever produce.
Once a transcript is trustworthy, it stops being a compliance file and becomes an index. MetadataIQ turns time-coded transcripts into searchable metadata inside PAM and MAM environments, so producers find the moment instead of scrubbing the file. The same text layer improves archive licensing, contextual ad placement and AI summarization. It travels beyond media too: learning management workflows depend on lecture captions students can rely on. Same discipline, different audience. Our customer success stories show what that looks like in production.
Frequently asked questions
What is a good word error rate for captions? It depends on content. Clean prerecorded audio should reach under 1% WER after QC, matching the 99% accuracy benchmark. Live and noisy content is better judged on NER, where 98% is the accepted threshold.
Does the FCC require 99% caption accuracy? No. The FCC requires captions to be accurate, synchronous, complete and properly placed, and applies a de minimis test to errors rather than a fixed percentage.
What is the difference between WER and NER? WER counts every substitution, deletion and insertion equally. NER weights errors by how much they damage comprehension and allows deliberate condensing, which suits live captioning.
Can automatic captioning alone meet compliance standards? Rarely, for regulated content. It is an effective first pass, but style conformance, speaker labelling, placement and severity-weighted error clearance still need human review.
Ready to find out where your captions actually stand? Book a 15-minute caption accuracy workflow review and we will benchmark your output against a gold-standard reference set built from your own content.
Conclusion
Caption accuracy is not a vanity metric, and WER is not a verdict. WER is a fast, comparable signal telling you how far machine output sits from a verified reference. What it cannot tell you is whether the errors it counted mattered.
The teams getting this right measure more than one thing, weight errors by consequence, fix their audio before blaming their engine, and keep a human in the loop wherever being wrong is expensive. That is what turns a percentage on a vendor slide into captions your audience can trust.
Key takeaways
- WER = (S + D + I) / N, scored against a verified human reference transcript.
- Every word scores the same in WER, so a dropped “not” costs what a stray “um” costs. Weight errors by severity.
- The FCC sets four quality standards but no numerical accuracy threshold. The 99% figure is industry practice.
- Use WER and FER for prerecorded content, NER for live. Ofcom treats 98% NER as the live threshold.
- Most errors originate in the audio. Clean capture and vocabulary biasing beat engine-switching.
- AI first pass, automated QC, human review, conformance check. That sequence scales without breaking compliance.
