In US Spanish-language network newscasts, 85% of the analyzed samples exceeded a 7-second caption latency ceiling, and the average caption arrived 8.2 seconds after the words were spoken (Fresno, JoSTrans 2024). Accuracy tells a parallel story: on news programming, human captioners reached 99.4% NER accuracy while unedited automatic captions reached 97.4%, below the 98% quality floor (Romero-Fresco and Fresno, Linguistica Antverpiensia 2023). Meanwhile, complaints about closed captioning on TV filed with the FCC rose from 310 in 2023 to 683 in 2024 (FCC CGB Consumer Complaints Data). We aggregated data from the FCC, Gallaudet University researchers, the Universidade de Vigo and University of Texas Rio Grande Valley NER studies, the Open ASR Leaderboard, WHO, and more than a dozen peer-reviewed papers. For the wider picture on delay across all live formats, see our companion roundup on real-time subtitling latency.
TL;DR
- Human captions on US news scored 99.4% NER accuracy versus 97.4% for unedited automatic captions (Romero-Fresco and Fresno, Linguistica Antverpiensia 2023).
- US Spanish-language newscasts averaged 97.04% NER accuracy, below the 98% threshold (Fresno, JoSTrans 2024).
- 85% of the Spanish newscast samples exceeded a 7-second caption latency ceiling (Fresno, JoSTrans 2024).
- Original live TV captions ran 7 to 12 seconds behind speech; automatic captions in the same test ran about 2 seconds behind (Thompson et al., arXiv 2026).
- Viewers lost about 1 quality rating point per 7.5 seconds of TV caption delay (Thompson et al., arXiv 2026).
- Only 11 of 70 live TV clips and 4 of 70 ASR clips reached 98% NER (Thompson et al., arXiv 2026).
- Automatic captions made about 22 errors per minute versus 5 for stenography in a 2020 US case study (Romero-Fresco and Fresno, 2023).
- FCC complaints about closed captioning on TV hit 683 in 2024, up from 310 in 2023 (FCC CGB Consumer Complaints Data).
- 65% of all errors in a certified automatic caption system were punctuation errors (Romero-Fresco and Van Gauwbergen, Applied Sciences 2025).
- Live captioners’ mental workload rated similar to that of neurosurgeons, in a study of 30 captioners (Nam et al., Applied Ergonomics 2023).
- The best long-form English ASR result on the Open ASR Leaderboard was 7.32% WER (Open ASR Leaderboard paper, arXiv 2025, data as of March 2026).
1. Live News Caption Accuracy: Human vs Automatic
The headline finding from the largest human-versus-machine comparison so far is not that automatic captions fail, but that they fail in a specific band. Unedited automatic captions on news reached 97.4% NER accuracy in a 2021 US case study, against 99.4% for human captions on the same genre (Romero-Fresco and Fresno, Linguistica Antverpiensia 2023). Two points of NER sounds small, but it is the difference between passing and failing: 98% is the acceptable threshold, and the automatic score translated to just 3.5 on the 10-point NER scale.
News also turns out to be the friendliest genre for machines. In the same case study, automatic captions on sports fell to 95.4% while human captions on sports climbed to 99.6%, so the gap widens precisely where speech gets fast and overlapping. The study analyzed about 17,000 live captions across four case studies in the UK, the US and Canada between 2018 and 2022.
| Metric | Value | Source |
|---|---|---|
| Human captions, US news genre, NER accuracy | 99.4% | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Enco case study, 2021) |
| Automatic captions, US news genre, NER accuracy | 97.4% (3.5 out of 10) | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Enco case study, 2021) |
| Stenography vs respeaking vs automatic, US | 99.1% vs 98.6% vs 96.3% | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Vitac case study, 2020) |
| UK respoken vs Microsoft vs Google captions | 97% vs 96% vs 95.5% | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Sky case study, 2018) |
| Canadian human live captions, average NER | 98.9% (7 out of 10) | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Canada, 2019-2021) |
| Canadian sample size | 440 ten-minute clips, 9,300 captions | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 |
| UK live subtitles, Ofcom-commissioned assessment, overall accuracy | 98.4% across 78,000 subtitles | Romero-Fresco and Fresno, 2023, citing the Ofcom 2013-2015 project |
Context note: the case studies date from 2018 to 2021, so they predate the newest speech engines. The same authors report that newer engines reached NER accuracy between 99.1% and 99.9% under controlled conditions with a single speaker and clear audio, which is the opposite of a breaking-news remote. Full paper: The accuracy of automatic and human live captions in English. Most recent available data for the Ofcom project: 2013-2015.
2. Caption Display Delay in Live News
Latency is where live news captions lose viewers before a single word is wrong. 85% of the US Spanish-language newscast samples analyzed exceeded a 7-second latency ceiling (Fresno, JoSTrans 2024), and individual caption delays ranged from 0.7 to 41 seconds. A 41-second caption on a fast news segment is describing a different story than the one on screen.
The automatic side of the market has a structural advantage here. Automatic captions typically run 2 to 3 seconds behind speech, against 4 to 10 seconds for human captions (Romero-Fresco and Van Gauwbergen, Applied Sciences 2025). The 2026 Gallaudet study measured original TV caption delays of 7 to 12 seconds on its live clips while its ASR condition, produced by broadcast vendor AppTek, ran at a mean of 2 seconds (Thompson et al., arXiv 2026). The trade-off for broadcasters is therefore not accuracy versus nothing, but accuracy versus time.
| Metric | Value | Source |
|---|---|---|
| Average caption delay, Telemundo vs Univision newscasts | 8.2 vs 8.3 seconds | Fresno, JoSTrans 2024 |
| Share of Spanish newscast samples exceeding 7-second latency | 85% | Fresno, JoSTrans 2024 |
| Range of individual caption delays | 0.7 to 41 seconds | Fresno, JoSTrans 2024 |
| Latency benchmark cited for US captions | 6 to 7 seconds acceptable, 4 to 5 seconds target | Fresno, JoSTrans 2024 (as cited in the study) |
| Original live TV caption delay in test clips | 7 to 12 seconds | Thompson et al., Gallaudet University, arXiv 2026 |
| ASR caption delay in the same test | 2 seconds mean (plus or minus 0.05 s jitter) | Thompson et al., Gallaudet University, arXiv 2026 |
| Typical delay, automatic vs human captions | 2-3 s vs 4-10 s | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| Super Bowl LII and 2016 presidential debate caption latency | 5.3 s and 6.1 s | Fresno, JoSTrans 2024 (summarizing earlier studies) |
Outlier note: the FCC rule itself contains no number of seconds. The 6-to-7-second figure is a benchmark reported by Fresno in the JoSTrans study of Spanish-language newscasts, not a codified limit (see section 6).
3. Word Error Rate vs NER: Why the Metrics Disagree
Word error rate is the default number in speech recognition, and it punishes human captioners for doing their job. Professional live captioners condense speech to keep up; WER counts every dropped word as an error, whether or not meaning survives. In the 2026 Gallaudet study, TV captions had a mean WER of 33.8% yet a mean NER of 95.28%, while ASR captions had a mean WER of 17.2% and a mean NER of 94.34% (Thompson et al., arXiv 2026). The machine “wins” on WER and loses on NER.
The more important result is that no metric tracks viewers equally across technologies. For TV captions, WER correlated with viewer quality ratings at r = -0.687 (p < .001), but for ASR captions the correlation fell to r = -0.390 (p = .003). NER showed the same pattern: r = 0.647 for TV versus r = 0.383 for ASR. A regulator that picks one metric is also, without saying so, picking a technology.
| Metric | Value | Source |
|---|---|---|
| Mean NER, TV captions vs ASR captions | 95.28 vs 94.34 | Thompson et al., arXiv 2026 |
| Clips reaching the 98 NER threshold, TV vs ASR | 11 of 70 vs 4 of 70 | Thompson et al., arXiv 2026 |
| NER correlation with viewer ratings, TV vs ASR | r = 0.647 vs r = 0.383 | Thompson et al., arXiv 2026 |
| ACE2 correlation with viewer ratings, TV vs ASR | r = -0.631 vs r = -0.291 | Thompson et al., arXiv 2026 |
| Mean ACE2 score, TV vs ASR (lower is better) | 0.67 vs 0.52 | Thompson et al., arXiv 2026 |
| Effect size of WER difference, TV vs ASR | d = 0.86 | Thompson et al., arXiv 2026 |
| Effect size of quality-rating difference, TV vs ASR | d = -0.13 | Thompson et al., arXiv 2026 |
Context note: viewers rated TV and ASR captions almost identically (d = -0.13) even though the WER gap was large (d = 0.86). An earlier CHI 2024 study by the same Gallaudet group found only weak correlation between WER, weighted WER, ACE and ACE2 and viewer ratings (Arroyo Chavez et al., arXiv 2404.10153). The 2026 paper is a preprint: Are Caption Metrics Broken?.
4. What Caption Errors Look Like in Practice
Error counts hide error types. In US Spanish-language newscasts, between 41% and 46% of captions contained at least one error, but 82% of those errors were minor, 16% standard and 2% serious (Fresno, JoSTrans 2024). Minor errors slow reading; serious errors create false information, such as a wrong number in a market report.
Automatic systems fail differently. In the 2025 NER certification of an automatic captioning system, 65% of all errors came from punctuation, and speaker identification raised error rates by 50% to 100% in multi-speaker videos (Romero-Fresco and Van Gauwbergen, Applied Sciences 2025). Human captioners edit; machines transcribe. The automatic system’s edition rate was 0.07%, against 6% to 33% for human captions across regions, which explains both its higher WER fidelity and its weakness at keeping pace with readers.
| Metric | Value | Source |
|---|---|---|
| Spanish newscast captions with at least one error | 41-46% | Fresno, JoSTrans 2024 |
| Error severity mix, Spanish newscasts (minor / standard / serious) | 82% / 16% / 2% | Fresno, JoSTrans 2024 |
| Error severity mix, automatic English captions (minor / standard / serious) | 89.2% / 10% / 0.8% | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| Share of automatic caption errors from punctuation | 65% | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| Error-rate increase with multiple speakers | 50-100% | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| Edition rate, automatic vs human captions | 0.07% vs 6-33% | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| Errors per minute: stenography vs respeaking vs automatic | 5 vs 7 vs 22 | Romero-Fresco and Fresno, Linguistica Antverpiensia 2023 (Vitac, 2020) |
| Text reduction, Spanish newscast captions | 26% | Fresno, JoSTrans 2024 |
Context note: reading speed explains the reduction. Spanish newscast anchors spoke at an average of 167 words per minute while captions averaged 121 words per minute (11.574 characters per second); 39% of captions exceeded 15 characters per second (Fresno, JoSTrans 2024). For comparison, the DCMP Captioning Key sets presentation rates not to exceed 160 wpm even for upper-level media. Source for the certification data: Fit for What Purpose? NER Certification of Automatic Captions.
5. Viewer Impact: Delay, Errors and Fatigue
The viewer data puts a price on every second of lag. TV captions with their original broadcast delay were rated 3.66 out of 7 for quality, against 5.09 when the same captions were synchronized, and ASR captions with a 2-second delay still scored 4.81 (Thompson et al., arXiv 2026). The study estimated roughly one rating point lost per 7.5 seconds of TV caption delay. One participant put it plainly: they would rather have incomplete captions in sync than better captions lagging behind.
Errors also cost the speaker, not only the viewer. In a preregistered experiment with 207 US participants, error-prone ASR subtitles consistently reduced evaluations of both the speaker and the content (Kadoma, Shrivastava and Naaman, arXiv 2026). And captions that help comprehension can still tire viewers: in an audiology lab test, imperfect, time-delayed captions raised sentence recognition from 63% to 77% but produced larger increases in subjective fatigue (Esser, Picou and Hornsby, American Journal of Audiology 2025).
| Metric | Value | Source |
|---|---|---|
| Quality rating, delayed vs synchronized TV captions (7-point scale) | 3.66 vs 5.09 | Thompson et al., arXiv 2026 |
| Understanding rating, delayed vs synchronized TV captions | 4.74 vs 5.65 | Thompson et al., arXiv 2026 |
| Rating loss per unit of TV caption delay | About 1 point per 7.5 seconds | Thompson et al., arXiv 2026 |
| Survey participants using captions daily | 84.7% of 216 | Thompson et al., arXiv 2026 |
| Sentence recognition with vs without imperfect captions | 77% vs 63% | Esser et al., American Journal of Audiology 2025 |
| Spoken words correctly shown in those test captions | 86-88% | Esser et al., American Journal of Audiology 2025 |
| Participants in subtitle-error evaluation experiment | 207 | Kadoma et al., arXiv 2026 |
| Crowdsourced viewers in caption stability evaluation | 123 | Liu et al., CHI Extended Abstracts 2023 |
Context note: the audience is large. WHO estimates that over 5% of the world’s population, or 430 million people, need rehabilitation for disabling hearing loss, rising to more than 700 million by 2050 (WHO deafness and hearing loss fact sheet, 2026). Flickering interim ASR text adds a separate burden; Liu et al. found stabilization techniques reduced viewer distraction and fatigue. Short-form platforms face the same issue at a different scale, as covered in our short-form video caption roundup.
6. FCC Rules and Caption Complaints
US rules define quality in words, not numbers. Under 47 CFR 79.1, captions must be accurate, synchronous, complete and properly placed, and “shall begin to appear at the time that the corresponding speech or sounds begin”. The rules do restrict the cheapest live news method: ABC, CBS, Fox and NBC and their affiliates in the top 25 markets cannot count teleprompter-based Electronic Newsroom Technique (ENT) captions toward compliance. Stations elsewhere that use ENT must script in-studio news, sports and weather and supplement unscripted breaking news with crawls or other text.
The complaint data shows rising friction. Complaints about closed captioning on TV in the FCC consumer complaint dataset rose from 310 in 2023 to 683 in 2024, an increase of 120% ((683 - 310) / 310), and stayed elevated at 604 in 2025 (FCC CGB Consumer Complaints Data). Complaints about captions over the internet peaked earlier, at 705 in 2023.
| Metric | Value | Source |
|---|---|---|
| Complaints: closed captioning on TV, 2023 / 2024 / 2025 | 310 / 683 / 604 | FCC CGB Consumer Complaints Data (queried October 2026) |
| Complaints: closed captioning over the internet, 2023 / 2024 / 2025 | 705 / 472 / 339 | FCC CGB Consumer Complaints Data |
| Closed captioning on TV complaints, 2026 year to date | 376 | FCC CGB Consumer Complaints Data |
| Cumulative complaints since October 31, 2014: TV vs internet captions | 4,139 vs 3,057 | FCC CGB Consumer Complaints Data |
| General captioning complaints, 2009-2013 | 2,323 (average 465 per year) | FCC Caption Quality Report and Order, FCC 14-12 (2014) |
| Window for filing a captioning complaint | 60 days after the problem | eCFR, 47 CFR 79.1(g)(1) |
| Distributor deadline to respond to an FCC-forwarded complaint | 30 days | eCFR, 47 CFR 79.1(g)(4) |
| Revenue level below which a channel is exempt from captioning | Under $3,000,000 per year | eCFR, 47 CFR 79.1(d)(12) |
Context note: the FCC states it does not verify the facts alleged in these informal complaints, and categories are chosen by the consumer, so the dataset cannot isolate live news from other programming (CGB Consumer Complaints Data). The 2009-2013 figure comes from a different intake system described in the FCC 14-12 caption quality order, so the two series should not be read as one trend line. Relay services for deaf and hard of hearing callers are a related FCC program, detailed in our TRS and VRS statistics.
7. The Workforce and Technology Shift
The human side of live captioning is under strain. Thirty live captioners rated their subjective mental workload as high, similar to that of neurosurgeons on the NASA-TLX scale, with paraphrasing and employment status as the main drivers (Nam et al., Applied Ergonomics 2023). The authors suggest that captioners move toward supervising AI-generated captions rather than typing every word. Supply pressure on realtime captioners is covered in our CART statistics roundup.
Hybrid workflows already show measurable gains. CARTGPT, which uses large language models and ASR to correct human CART transcripts, improved word accuracy by 5.6% over baseline CART and by 17.3% over state-of-the-art ASR on a 39.7-hour dataset (Wu, Kleiver and Jain, ASSETS 2025). Benchmarks show how far raw engines have come: the best short-form English result on the Open ASR Leaderboard was 5.42% WER, and the best long-form result was 7.32% (Open ASR Leaderboard paper, arXiv 2025, data as of March 27, 2026).
| Metric | Value | Source |
|---|---|---|
| Live captioners in workload study | 30 | Nam et al., Applied Ergonomics 2023 |
| CARTGPT word accuracy gain vs baseline CART / vs ASR | 5.6% / 17.3% | Wu et al., ASSETS 2025 |
| ASR systems and datasets on the Open ASR Leaderboard | 86 models, 12 datasets | Open ASR Leaderboard paper, arXiv 2025 |
| Best short-form English WER on the leaderboard | 5.42% (Cohere Labs Transcribe) | Open ASR Leaderboard paper, arXiv 2025 |
| Long-form English WER: best proprietary vs Whisper Large v3 | 7.32% vs 11.2% | Open ASR Leaderboard paper, arXiv 2025 |
| Certified automatic captions, overall NER: English vs Spanish | 99.03% vs 98.77% | Romero-Fresco and Van Gauwbergen, Applied Sciences 2025 |
| WER range of real-time clinical captions for deaf and hard of hearing adults | 12.7% to 22.8% | Hughes et al., JMIR Rehabilitation and Assistive Technology 2026 |
| Clinical caption users who found captions easy to follow | 90% (26 of 29 simulations) | Hughes et al., JMIR Rehabilitation and Assistive Technology 2026 |
Context note: leaderboard WER is measured on benchmark audio, not on live, multi-speaker news with crosstalk and remote feeds, and it says nothing about delay. Sources: Open ASR Leaderboard paper, CARTGPT, Nam et al. and Hughes et al.. Courtroom transcription shows a similar gap between benchmark and field accuracy, see our courtroom speech recognition statistics.
Summary: Closed Caption Display Delay and Word Error Rate in Live News by the Numbers
| Metric | Value | Source |
|---|---|---|
| Human captions, US news, NER accuracy | 99.4% | Romero-Fresco and Fresno, 2023 |
| Automatic captions, US news, NER accuracy | 97.4% | Romero-Fresco and Fresno, 2023 |
| US Spanish-language newscasts, NER accuracy | 97.04% | Fresno, JoSTrans 2024 |
| Spanish newscast samples over 7-second latency | 85% | Fresno, JoSTrans 2024 |
| Average Spanish newscast caption delay | 8.2 seconds | Fresno, JoSTrans 2024 |
| Original live TV caption delay in test clips | 7 to 12 seconds | Thompson et al., arXiv 2026 |
| Typical delay, automatic vs human captions | 2-3 s vs 4-10 s | Romero-Fresco and Van Gauwbergen, 2025 |
| Mean WER, TV vs ASR captions | 33.8% vs 17.2% | Thompson et al., arXiv 2026 |
| Clips reaching 98 NER, TV vs ASR | 11 of 70 vs 4 of 70 | Thompson et al., arXiv 2026 |
| Quality rating, delayed vs synchronized TV captions | 3.66 vs 5.09 out of 7 | Thompson et al., arXiv 2026 |
| Errors per minute, stenography vs automatic | 5 vs 22 | Romero-Fresco and Fresno, 2023 |
| Automatic caption errors from punctuation | 65% | Romero-Fresco and Van Gauwbergen, 2025 |
| Spanish newscast captions with at least one error | 41-46% | Fresno, JoSTrans 2024 |
| UK live subtitles, overall accuracy (2013-2015) | 98.4% | Romero-Fresco and Fresno, 2023, citing Ofcom project |
| FCC TV caption complaints, 2023 vs 2024 | 310 vs 683 | FCC CGB Consumer Complaints Data |
| FCC TV caption complaints, 2025 | 604 | FCC CGB Consumer Complaints Data |
| Captioning complaints filed with FCC, 2009-2013 | 2,323 | FCC 14-12 (2014) |
| Live captioner workload | Similar to neurosurgeons | Nam et al., Applied Ergonomics 2023 |
| Best long-form English ASR WER | 7.32% | Open ASR Leaderboard paper, arXiv 2025 |
| People needing rehabilitation for disabling hearing loss | 430 million | WHO fact sheet, 2026 |
Methodology and Sources
- Romero-Fresco, P. and Fresno, N.: The accuracy of automatic and human live captions in English, Linguistica Antverpiensia 22 (2023)
- Fresno, N.: Closed Captions en espanol. Live Captioning Quality in Spanish-language Newscasts in the U.S., The Journal of Specialised Translation (2024)
- Romero-Fresco, P. and Van Gauwbergen, Y.: Fit for What Purpose? NER Certification of Automatic Captions in English and Spanish, Applied Sciences 15(3) (2025)
- Thompson, B. et al. (Gallaudet University): Are Caption Metrics Broken? Latency, Deaf and Hard of Hearing User Ratings, and Bias across Technologies, arXiv preprint (2026)
- Arroyo Chavez, M. et al. (Gallaudet University): How Users Experience Closed Captions on Live Television, CHI 2024
- Romero-Fresco, P.: Negotiating quality assessment in media accessibility: the case of live subtitling, Universal Access in the Information Society (2020)
- FCC: CGB Consumer Complaints Data, queried by issue and year in October 2026
- FCC: Closed Captioning Quality Report and Order, FCC 14-12 (2014)
- eCFR: 47 CFR 79.1, Closed captioning of televised video programming
- DCMP: Captioning Key, Presentation Rate
- Esser, K., Picou, E. and Hornsby, B.: Imperfect Captions Paired to Audiovisual Stimuli Improve Speech Recognition but Do Not Reduce Listening-Related Fatigue, American Journal of Audiology (2025)
- Kadoma, K., Shrivastava, P. and Naaman, M.: Lost in Transcription, arXiv (2026)
- Liu, X. et al.: Modeling and Improving Text Stability in Live Captions, CHI Extended Abstracts (2023)
- Nam, S. et al.: Assessing subjective workload for live captioners, Applied Ergonomics (2023)
- Wu, L.-Y., Kleiver, A. and Jain, D.: CARTGPT: Real-Time Correction of CART Captions Using Large Language Models, ASSETS 2025
- Hughes, S. et al.: Assessing the Role of Medical Caption Technology, JMIR Rehabilitation and Assistive Technology (2026)
- Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation, arXiv (2025, results as of March 27, 2026)
- WHO: Deafness and hearing loss fact sheet (2026)
- Data watch: the Gallaudet latency and metric study (Thompson et al.) and the Kadoma et al. experiment are preprints and may change after peer review. The case studies in Romero-Fresco and Fresno date from 2018 to 2022, and the Ofcom live subtitling project ran from 2013 to 2015 (most recent available data for UK accuracy). Ofcom and CRTC pages returned bot-protection or 403 errors during research, so UK and Canadian figures are cited through the peer-reviewed papers that report them. The 6-to-7-second latency benchmark and the Super Bowl LII and 2016 debate figures are as reported by Fresno (2024); 47 CFR 79.1 itself sets no numeric delay limit. FCC complaint counts are informal, unverified and consumer-categorized, do not isolate live news, and the 2026 figure is year to date; the 120% rise is our own calculation from the 2023 and 2024 counts. The Spanish newscast study (2024) covers two networks and 200 minutes. Leaderboard WER figures are benchmark results and do not measure live-news delay.
Last updated: October 4, 2026. We update this roundup quarterly, and the next refresh is expected when the Gallaudet caption metrics study completes peer review and the FCC complaint dataset closes out calendar year 2026.