· Court Transcript Platform
How to read a transcription accuracy claim
Everyone says 99%. Here is what the number actually describes at each service, and why 'claimed' and 'measured' are different things — including how to apply both tests to us.
Shop for transcription and you will meet the same number everywhere: 99%. It is not a lie — but it rarely describes what you think it describes. Here is the landscape as the vendors themselves publish it (as of August 2026), and the two questions that make any accuracy number legible. Ours included.
What the industry claims, in its own words
| Service | Published claim | Which tier |
|---|---|---|
| Rev | "99%+ accurate," guaranteed for clear audio; 99.6% for the court-ready product | Human |
| Rev | "96%+ accuracy guaranteed" | AI-only |
| GoTranscript | 99% guaranteed, money-back below it | Human |
| Verbit | "99% accuracy" for legal | Hybrid AI + human |
| TranscribeMe | 99%, including AI-generated | AI / hybrid |
| Otter and most AI-only tools | Usually no published number; third-party tests land around 85–90% | AI-only |
Two questions turn this table into information.
Question 1: which tier does the number describe?
Every "99%" above is a human or hybrid tier — a person read the transcript. The AI-only numbers, where published at all, sit visibly lower. That is not scandal; it is the honest shape of the technology.
It is also why we publish our numbers the way we do. Our 97.9% and 96.6% describe the raw machine draft; our 98.6% describes that draft after a human checks only the words the system itself marked as uncertain — one reviewed word in twenty-five. Our certified tier — the one comparable to everyone's 99% row — adds a licensed human's full pass on top of that floor. We publish the machine layers separately because that is the part we can measure without grading our own homework.
Question 2: claimed on whose audio?
An accuracy claim is measured on something — usually audio the vendor chose. We measured one leading service's AI tier, the one marketed at "96%+ guaranteed," on certified court audio with the official transcript as ground truth: it scored 97.2% on one hearing and 96.4% on the other — above its guarantee on this clean audio, and still behind our draft on both hearings under the same blind scoring.
A guarantee met on the easiest court audio there is says little about a noisy motion calendar — which is why every number we publish names its corpus — certified Supreme Court transcripts, clean audio, stated as our ceiling condition — and why the same scoring is applied blind to everyone we measure, ourselves included. Ask any vendor, including us: measured on what, against what reference? The answer is the claim.
The test, applied to us
Tier: stated per number, machine layers published separately from the human tier. Corpus: certified public court transcripts, named, with the ceiling-condition caveat attached. Measurement: same blind scoring for us and anyone we compare against, run on every release; full reports available on request.
Hold every accuracy claim to those two questions — especially ours.