Broadcasters, publishers and podcasters use speech recognition for live captions, subtitles on recorded programmes, searchable archives and newsroom tools. Live captioning trades accuracy against delay; pre-recorded subtitling trades speed against polish. In both, the difficult words are the names in today's news, and the output must be split into readable, well-timed captions rather than a block of text.
Key takeaways
- Live captions have a few seconds to appear. Pre-recorded subtitles can use the whole programme as context, a second pass and human review.
- Captions are a formatting problem as well as a recognition problem: line length, reading speed, shot changes and speaker identification all matter.
- News vocabulary changes daily. Feeding running orders, scripts and rundowns to the recogniser before transmission is the cheapest accuracy gain available.
- Accuracy for live subtitling is often scored with the NER model, which weights errors by how much they mislead the viewer, rather than plain WER.
- EU law, through the Audiovisual Media Services Directive and the European Accessibility Act, pushes providers to make content accessible, including captioning.
Workflows#
| Workflow | Mode | Time budget | What matters most |
|---|---|---|---|
| Live captions (news, sport, parliament) | Streaming | Seconds | Delay, names, stability on screen |
| Pre-recorded subtitles | Batch | Hours to days | Timing, segmentation, reading speed |
| Archive indexing | Batch | Weeks, in bulk | Cost per hour, searchability of names |
| Newsroom search and clipping | Streaming or batch | Minutes | Word timestamps for editing |
| Podcast and video transcripts | Batch | Hours | Speaker labels, readability |
| Translated subtitles | Batch | Days | Accurate source text and timing |
Archive work is often the largest by volume. Broadcasters hold decades of programmes with little metadata; transcription makes them searchable by what was said. Here, cost and long-form reliability matter more than latency.
Live captioning#
Live subtitles have traditionally been produced by respeaking: a subtitler listens to the programme and repeats it, with spoken punctuation, into a recogniser trained on their voice, correcting errors on the fly. Respeaking can condense fast speech and fix names, at the cost of a delay of several seconds and staff for every live hour. Stenography is used in some markets.
Direct ASR on programme audio removes the respeaker. The trade-offs are different:
- Delay. ASR can emit words within a second of speech, faster than a respeaker. But the text rewrites itself as the model revises its hypothesis (see partial vs final results). Showing unstable partials on screen distracts viewers; waiting for finals adds delay.
- Verbatim speed. Presenters and panel guests can speak faster than viewers can comfortably read. ASR transcribes everything; a respeaker can edit down.
- Overlap. Panel discussions, studio audiences and sports commentary over crowd noise.
- Music and effects. Beds under speech, jingles and adverts, which a recogniser must not caption as words.
Measuring caption quality#
WER is used for ASR engines, but live subtitling quality in broadcasting is often scored with the NER model (Romero-Fresco and Martínez Pérez, 2015). It counts edition errors, where a subtitler's condensing changed meaning, and recognition errors, weighting each as minor, standard or serious, and computes an accuracy rate from the number of words. A rate of 98% is commonly used as the threshold for acceptable live subtitles. Ofcom has used the NER model in its reviews of live subtitling quality on UK television.
The NER model captures something WER does not: a misrecognised name in a news headline misleads the viewer more than a dropped "just". This makes it a close relative of semantic WER.
From transcript to subtitles#
Turning a transcript into subtitles means deciding where each caption starts and ends, where lines break, and how long each stays on screen. Good segmentation breaks at clause boundaries, avoids splitting a name across lines, respects shot changes where possible, and keeps reading speed within the broadcaster's limits. This needs accurate word timestamps.
Output formats depend on the destination: SRT for simple workflows, WebVTT for the web, TTML and its profile EBU-TT-D for European broadcast distribution. Details are in captions and subtitles: SRT, VTT and timing. Speaker changes are marked with colour, dashes or labels according to house style, which requires diarisation that respects programme structure.
Names in the news#
The words that matter most in broadcast are proper nouns that change every day: a newly appointed minister, a town hit by flooding, a transfer signing. A general model has never seen many of them. Kyriakos Mitsotakis becomes Carrie across Mitsu Takis; a Polish footballer's surname becomes three English words. See rare words and proper nouns.
Newsrooms already hold the text needed to fix this: running orders, autocue scripts, wire copy and team sheets. Supplying them to the recogniser before the bulletin gives it the day's names. Sport adds scores, times and statistics that must be formatted correctly ("two-nil" as 2–0); finance programmes add tickers and figures. Many programmes also mix languages: interviews with foreign guests, or regional broadcasters working in two languages (see code-switching).
Accessibility obligations#
The Audiovisual Media Services Directive, as amended by Directive (EU) 2018/1808, requires Member States to ensure that media service providers make their services continuously and progressively more accessible to persons with disabilities (Article 7), including through subtitling and audio description. National regulators set the detail, often as quotas by channel size. The European Accessibility Act (Directive (EU) 2019/882), applicable from 28 June 2025, adds accessibility requirements for services giving access to audiovisual media, such as players and electronic programme guides, including support for subtitles. In the UK, Ofcom's Code on Television Access Services sets subtitling targets for qualifying channels.
Public broadcasters' online content may also fall under the Web Accessibility Directive; see speech recognition in the public sector.
How Reson8 approaches this#
Resonant-1 scored 2.32% WER on TED-LIUM and 8.76% on GigaSpeech, which includes podcasts and online video, in the Open ASR Leaderboard short-form English sets (April 2026). On long-form English, its average was 8.58%, third in that table behind ElevenLabs Scribe v2 and AssemblyAI Universal 3 Pro. Scripts and running orders can be supplied as text through domain adaptation, which takes effect in under 60 seconds, so a newsroom can update names before each bulletin. Batch transcription costs 1 credit per minute, streaming 2.
Common questions#
Is ASR good enough for live broadcast captions?
For many programme types, yes, and it is widely used. Fast, overlapping or noisy programmes remain difficult, and names need preparation. Measure on your own output with a severity-weighted metric such as NER.
Should subtitles be verbatim?
House styles differ. Many broadcasters allow light editing for reading speed, especially for fast speech. Pure ASR output is verbatim by default, so condensing needs an extra step.
How are subtitles synchronised with video?
From word timestamps produced during recognition or by forced alignment against the audio, then adjusted for shot changes and minimum display times.
What about translated subtitles?
They start from an accurate source-language transcript with good segmentation. Errors in the source, especially names, carry into every translation.
References#
- Romero-Fresco, P., & Martínez Pérez, J. (2015). Accuracy rate in live subtitling: The NER model. In Audiovisual Translation in a Global Context. Palgrave Macmillan.
- European Broadcasting Union. EBU-TT-D Subtitling Distribution Format, Tech 3380.
- W3C. WebVTT: The Web Video Text Tracks Format. W3C.
- Directive (EU) 2018/1808 amending the Audiovisual Media Services Directive. EUR-Lex.
- Directive (EU) 2019/882 (European Accessibility Act). EUR-Lex.
- Ofcom. Code on Television Access Services.
- Rousseau, A., Deléglise, P., & Estève, Y. (2012). TED-LIUM: an Automatic Speech Recognition dedicated corpus. LREC.
- Chen, G. et al. (2021). GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio. Interspeech. arXiv:2106.06909.