UK live subtitles by respeaking are accurate but late, Ofcom finds

In Ofcom's sampling of live subtitles from late 2013, accuracy was generally good and above 98% for news, but the median delay of 5.6 seconds exceeded its 3-second guideline.

  • 2 for a person

    The reviewers found that accuracy averaged 98.5%, a considerable improvement on their 2012 analysis, but that long delays caused omissions and misreporting.

Depth runs from 1, a convenience or time saved, to 5, a life or a livelihood preserved.

Status
Partly reported benefit.
Evidence

foundstudiedreportedclaimed

An independent regulatory document supports the benefit.

Compared with

The 98% accuracy threshold that Ofcom's external reviewers treat as acceptable. Measured.

Without the system
98%
With it
generally higher for news than 98%
The system’s part
Not isolated. Ofcom measured subtitles made by respeakers, who rely on speech recognition software, and by stenographers, without separating the two techniques. The reviewers attributed 37% of errors to recognition, through the subtitlers' interaction with steno machines or speech recognition software. The report names no software and does not show that it learns from data. A later study by the lead reviewer reports that almost all captions in Ofcom's study came from respeaking.
Later
Persisted, 13 December 2023. A study by the lead reviewer reports that the 78,000 captions produced by UK broadcasters for Ofcom's study between 2013 and 2015 averaged 98.3%.
Not yet checked
LimitsCosts and harmsLearns from data

Download data Cite Suggest a correction

Benefit type
Improved Accessibility
System
Live subtitling by respeakers using speech recognition software, alongside stenographers
Made by
Not named (speech recognition software used by subtitling providers)
Used by
BBC, ITV, Channel 4, Channel 5 and Sky and their subtitling providers
Where
United Kingdom
Stage
In routine use
Date
, when the source was published
Kind
Aggregate result
Broad benefit
Documenting/Recording Information
Collections
Scrutiny and findings
Last revised
1 November 2026

Sources

  1. Peer-reviewed study, by authors with no tie to the system.

    Romero-Fresco P, Fresno N. The accuracy of automatic and human live captions in English. Linguistica Antverpiensia, New Series: Themes in Translation Studies 22 (2023): 114-133, section 4 (final thoughts). Published 13 December 2023. doi:10.52034/lans-tts.v22i.774.

    higher than the 98.3% (6/10) obtained by the 78,000 captions produced by UK broadcasters between 2013 and 2015 for the Ofcom study (Romero-Fresco, 2016).
    However, it is important to note that while almost all the captions in the Ofcom study were produced by respeaking, at least half of the North American captions analysed in this study were produced by stenographers.

  2. Regulatory document, by authors with no tie to the system.

    Ofcom. Measuring live subtitling quality: Results from the first sampling exercise. 30 April 2014, with an external review of the measurements by Pablo Romero-Fresco and Inma Pedregosa (University of Roehampton) in Annex 1. Archived copy of https://www.ofcom.org.uk/__data/assets/pdf_file/0019/45136/sampling-report.pdf. Section 1, para. 1.10.

    In the first set of samples, accuracy was generally good, but rather variable. Accuracy rates by genre and broadcaster are shown in the graphs in section 2. Accuracy varied significantly by genre, and was generally higher for news than 98%.
    On the whole, the quality of the live subtitles analysed in this report may be regarded as good, with an average accuracy rate of 98.5%, including some very good examples (one fifth of the total sample) and also some subtitles that fall below the required quality threshold (also one fifth of the programmes).
    For all broadcasters, latency in the first set of samples was generally worse than the maximum recommended 3 seconds in Ofcom's guidelines. Typically, it was 75% more (the median latency was 5.6 seconds), with spikes of up to 24 seconds.
    In the sample analysed, 63% of the errors included in the subtitles were editing errors (those caused by incorrect omissions made by the subtitlers) and 37% were recognition errors, those caused by the interaction between the subtitler and the steno machine or the speech recognition software.

Details

AI system

Workflow

Respeakers repeated live speech, with punctuation and commands, into speech recognition software, and stenographers typed it on steno machines. Subtitlers edited fast speech to keep up, most of all in chat shows. Some news programs combined subtitles prepared from scripts with live subtitles.

Follow-up

Conditions for replication

Subtitlers must prepare names and terms in advance to avoid misrecognitions. Delays must stay short, since long delays caused omissions and misreporting. Serious errors that introduce misleading information must be corrected live.

Entry details

Sector

Broadcasting / live subtitles for deaf and hard-of-hearing viewers.

Purpose

Unnamed developers built the speech recognition software that respeakers use to turn their repetition of live speech into subtitles for deaf and hard-of-hearing viewers.

31 more questions not yet checked

model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, AI contribution, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.

Cite this record

Peter H. Vartanian. (2026). UK live subtitles by respeaking are accurate but late, Ofcom finds (AIBD-1150). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-1150.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)

BibTeX Data Field definitions