Word error rate (WER): what transcription accuracy numbers measure

Word error rate (WER) is the standard measure of speech-to-text accuracy: the number of word substitutions, deletions, and insertions needed to turn a machine transcript into a correct reference transcript, divided by the number of words in the reference. Lower is better. When a transcription vendor says its engine is “95% accurate”, it almost always means a WER of around 5% on some test set.

It is a good metric for the job it was designed for, which is comparing speech models against each other on the same audio. The trouble starts when it gets read as an answer to a different question — how good will my meeting notes be? — and the reason is worth stating up front: WER counts how many words are wrong, not which ones, and it only ever describes the transcript. Everything a meeting-notes reader cares about sits in one of those two blind spots.

How it’s calculated

The machine transcript (the hypothesis) is lined up word by word against a human-checked transcript of the same audio (the reference), using the alignment that needs the fewest edits. Each mismatch is one of three kinds:

Substitution (S) — a word heard as a different word.
Deletion (D) — a word that was said and is missing.
Insertion (I) — a word in the transcript that nobody said.

WER = (S + D + I) ÷ N, where N is the number of words in the reference.

Two consequences fall straight out of the formula. First, because insertions are counted but N isn’t increased by them, WER can go above 100% — a model that turns a one-word “yes” into “yes yes okay” has a WER of 200% on that clip. Second, “accuracy” expressed as one minus WER isn’t a true percentage of words right, and it isn’t always defined the same way; if a vendor quotes an accuracy figure, it’s reasonable to ask how it was computed.

The worked example that explains the problem

Take a six-word reference: “so um we can’t ship Friday.”

Transcript A: “so we can’t ship Friday.” One deletion. WER = 1 ÷ 6 ≈ 17%.
Transcript B: “so um we can ship Friday.” One substitution. WER = 1 ÷ 6 ≈ 17%.

Same score. Transcript A lost a filler word nobody needed; Transcript B inverted the decision. Many benchmarks strip fillers before scoring, in which case Transcript A scores a perfect 0% and Transcript B still scores 17% — the harmless error disappears from the metric and the one that changes the meeting stays exactly as heavy as a misheard “the”.

That is the metric’s core property, not an edge case. WER weights every word equally, and meetings don’t. A misheard person’s name is one error in a thousand-word transcript and the only word in the action item. A dropped “not” is one error and the whole meaning. Proper nouns, numbers, and negations — the tokens that carry consequences — are also the ones speech models find hardest, which is why how accurate AI meeting notes are can’t be read off a single percentage.

What a WER figure doesn’t tell you

A WER number is silent on…	Why	Where it matters
Which words were wrong	Every error counts as one, whatever it was	A wrong name or a lost negation costs as much as a lost “um”
Your audio	It’s measured on a specific test set, usually recorded speech under reasonable conditions	Crosstalk, a speakerphone across a room, and platform compression aren’t in most benchmarks
Who said it	Speaker labels are scored separately, if at all	A perfect transcript attributed to the wrong people reads as authoritative and is wrong — see speaker diarization
Live vs final text	Scores are normally computed on finished transcripts	What a live tool shows first is an interim result that gets revised
Punctuation, casing, numbers	Text is usually normalized before scoring, and normalization choices differ	”Twenty” vs “20”, or a question read as a statement, may or may not count
The summary	WER measures the transcript and nothing downstream of it	The notes you actually read have no equivalent number at all

Why two WER figures rarely compare

Because the number depends as much on the test as on the model, two published WERs are only comparable if they share three things:

The same audio. Read speech, broadcast news, and conversational meeting audio produce very different scores from the same engine. Conversational audio, with people interrupting and trailing off, is among the hardest. A figure from one kind of test set says little about another — and very little about your calls, which is why a tool that tested well can still be rough in a conference room.
The same normalization. Whether fillers are removed, how numbers and contractions are handled, and whether casing and punctuation count can move the score noticeably without the transcript changing at all.
The same mode. Streaming transcription commits to words with less of the sentence still to come, so a streaming model generally trails an offline model of the same family on the same audio. A batch benchmark tells you little about a live product.

Language matters too. Speech models are unevenly good across languages and accents, and a headline figure is usually for English on well-represented speech; for languages written without spaces between words, such as Chinese or Japanese, the comparable metric is character error rate (CER). The languages answer covers what a long supported-languages list does and doesn’t promise.

Human transcription isn’t zero WER either — careful transcribers disagree with each other on messy conversational audio — so “human-level” means comparable to a person, not perfect. When verbatim exact words genuinely matter, human-verified transcription is still the benchmark worth paying for, which is the honest case made in Canary vs Rev.

The failure WER can’t see: insertions from nothing

One kind of insertion deserves its own mention because it behaves differently from mishearing. Some speech models, given silence or background noise, produce plausible words anyway — Whisper’s tendency to do this is documented, and it’s why a local setup puts voice activity detection in front of the model (covered in Canary vs Whisper). On a benchmark those are just insertions in the count. In a meeting transcript they’re sentences nobody said, and a summary built on them can’t tell the difference.

The number that matters most doesn’t exist

Meeting notes are two layers: a transcript, and a summary written from it. WER describes the first. Nobody publishes a WER for the second, because there isn’t one — a summary is a rewrite, so there’s no word-for-word reference to line it up against.

The two accuracies are related but not the same. A summary can absorb small transcription errors: a dropped article or a misheard filler rarely survives into a note anyway. It can also do the opposite and launder a garbled passage into a clean, confident sentence. The accuracy answer calls this the difference between failing visibly and failing invisibly: a bad transcript looks like nonsense and gets noticed; a bad summary is a well-written paragraph that happens not to match the meeting. See live transcript vs live summary for why both layers are worth keeping reachable.

So the useful vendor question isn’t only “what’s your WER” but “on what audio, normalized how, for which mode — and how do I check the summary against the words underneath it?”

Where Canary sits

Canary is a real-time, bot-free meeting summarizer. It captures your computer’s system audio (no bot in the call, no plugin) and shows a live, multi-resolution rolling summary — from what’s being said right now to the whole call — so you can catch up the instant your name is called.

Two things follow for WER. On the input side, system-audio capture reads the stream your computer is already playing, so remote voices arrive as cleanly as they reached your ears rather than re-recorded through a room — and headphones don’t change that. Honestly, though, it is one mixed stream: Canary is subject to every factor above, and on attribution specifically a tool with access to the platform’s participant list has an advantage, as covered in how notetakers know who is speaking.

On the output side, Canary’s product is the layer WER doesn’t measure. A rolling summary can’t be scored against a reference, but it can be checked in a way a post-meeting note can’t: while the meeting is still running, when a “now” line that says the date moved — when it didn’t — costs one sentence to correct. The multi-resolution views help the same way, since a line that looks wrong at two minutes can be checked against the full-call view without leaving the summary. That isn’t a claim that the model is more accurate. It’s that the error has a shorter half-life, which is the one kind of accuracy a benchmark can’t report.

Character error rate (CER) — the same calculation at the character level, used for languages without word spacing.
Diarization error rate (DER) — the share of speech time assigned to the wrong speaker, missed, or falsely detected; the separate score for “who said it”.
Interim transcription results — the provisional words a streaming model shows before it commits.
Action-item detection — only as good as the names and verbs underneath it, which is exactly where WER is least informative.
On-device transcription — local models trade some accuracy for privacy; a WER comparison is how that trade usually gets described.

Word error rate (WER)