What "99% accurate" actually means

Go through the marketing pages of ten transcription products and you will find roughly the same claim on nine of them: 99% accurate. Sometimes 99.5%. Occasionally "up to 99%", which is a fine piece of writing, since "up to" is doing all the work.

The number is not a lie, exactly. It's a real measurement of a real thing. It's just a measurement of something other than what you're about to use the product for. Here's what's underneath it.

The number is one minus the word error rate

Accuracy claims in speech recognition are almost always 100% minus word error rate, and WER has a specific definition. Line up the machine's transcript against a correct one and count three kinds of mistake: words it got wrong (substitutions), words it missed (deletions), and words it invented (insertions). Add those up, divide by the number of words in the correct transcript.

WER = (substitutions + deletions + insertions) / total words

If you want to watch this happen on your own text, we built a word error rate calculator that shows the number and then breaks it down by where the errors landed.

Two percent WER means two errors per hundred words. On a one-hour recording — call it nine thousand words — that's about a hundred and eighty errors. "99% accurate" is a page with a mistake every two or three lines. Stated that way it sounds considerably less impressive, which is why nobody states it that way.

It's measured on audio that sounds nothing like yours

The low numbers come from benchmark datasets, and the most commonly used ones are far easier than real recordings. The classic benchmark is read speech — someone reading an audiobook aloud, alone, into a good microphone, in a quiet room, with no interruptions and no crosstalk. Modern systems do extremely well on it, in the low single digits of WER and sometimes under two percent.

Your recording is not that. Your recording is four people in a room with a laptop in the middle, one of them on speakerphone, somebody's chair squeaking, two people starting sentences at the same time, and a technical vocabulary the model has never encountered. On conversational audio of that kind, error rates several times higher than the benchmark are entirely normal, for every vendor.

So the honest translation of "99% accurate" is: on audio we chose, our system made one error per hundred words. That is a true sentence about a lab. It is not a prediction about your interview.

Every word counts the same, which is the real problem

Here's the part that matters more than the benchmark critique, and it's the reason I don't think WER should be the number you shop on at all.

WER treats every word as equally important. A missed "um" costs exactly what a missed "not" costs. Consider two transcripts of the same hour:

• Transcript A has 3% WER. The errors are filler words, false starts, and "gonna" written as "going to". Every name, every number, every negation is correct.
• Transcript B has 1% WER. The errors are the client's surname, one figure in a budget discussion, and a dropped "don't".

Transcript B wins the benchmark by a factor of three. Transcript A is the one you can publish. Transcript B is a document that quotes someone as agreeing to something they refused, with a great score.

This isn't a hypothetical edge case; it's the standard shape of the problem. The errors that matter are concentrated in exactly the words that carry the meaning — proper nouns, numbers, domain terms, negations — because those are the words with the least redundancy for a language model to lean on. "The" is easy to recover from context. Your interviewee's surname is not.

A single metric that averages over all words will systematically underweight the errors that ruin your work.

The reference transcript isn't perfect either

One more wrinkle, mostly for people who enjoy this sort of thing. WER is measured against a "correct" transcript that a human produced. But humans transcribing the same conversational audio disagree with each other — on where a sentence ends, on whether a mumble was a word, on how to render a false start. Studies on conversational speech have put human-to-human disagreement at several percent.

Which means once a system is reporting WER in that same range, a meaningful part of what's being measured is disagreement about what counts as correct, not error. Claims of "better than human accuracy" that come out of this territory are measuring against a ruler that wobbles about as much as the thing being measured.

The question I'd ask instead

Accuracy is table stakes and roughly commoditised. Most serious products are built on a handful of comparable engines, and the gaps between them are much smaller than the gaps between an easy recording and a hard one. If you're choosing a tool, differences in raw accuracy are unlikely to be what you actually feel.

What you'll feel is the twenty minutes after the transcript arrives. So:

How fast can I fix an error? You will be fixing errors regardless of vendor. Can you click a paragraph and type, or are you in a modal? Does correcting text destroy the timing data, so that click-to-play stops working on the lines you've edited?

How do I find what needs fixing? Can you click any line and hear that exact second? Without that, proofreading means scrubbing a waveform, and nobody does it, which means nobody catches anything.

Does the fix stick? If it mishears a term once it'll mishear it forty times. Can you give it the names and the jargon up front, so the same correction isn't retyped all afternoon?

Can I check what it tells me? When the product summarises the recording or answers a question about it, can you get from what it says back to the moment it came from, or is it a confident paragraph you have to take on trust? This is the one that separates a research tool from a convenience.

What does it do when it's wrong? Does it say "unclear", or does it produce fluent text that happens to be invented? Some failure modes are visible, and some are not: a model that quietly translates a sentence instead of transcribing it, or decodes one language as though it were another, returns something fluent and plausible at full coverage. Nothing looks wrong, so proofreading never catches it — which makes it far worse than an error you can see.

What we claim

We don't publish an accuracy number. We could pick a favourable benchmark and generate a big one, the same way everybody else does, and it would tell you nothing you could use.

What we'll tell you instead is where ours is weak. Speaker labels drift on long recordings, for structural reasons we've written up. A number of our languages don't support speaker separation at all. Automatic language detection can misfire on bilingual audio. Names and jargon need a word list to land reliably. And our summaries don't cite the moment each point came from — answers to questions do, so when a summary says something you need to rely on, ask about it and follow the citation.

And then the product is built around that reality rather than against it: every word timestamped so you can hear any line, edits that don't break the timings, speaker renames that propagate through the whole document, a custom vocabulary so a correction only happens once, and answers that cite the second they came from.

The pitch isn't that we're never wrong. It's that you can find out in two seconds — which, unlike 99%, is a claim you can test on your own recording before you pay us anything.

---

Corrected 20 September 2026: this post said our summaries cite the second they came from. Answers to questions do; summaries don't.