WER vs word recall: how to measure AI meeting transcription accuracy in 2026

Every speech-to-text vendor prints a number for meeting transcription accuracy: 99% accurate, 4% WER, best in class. Then you read the transcript of your own meeting, and it does not feel like 99%.

The number is not necessarily wrong. It measures something different from what you care about in a meeting. Word error rate (WER) tells you how much you must edit. Word recall tells you how much the transcript lost. For meeting notes, the second number usually matters more.

We measured both on a real one-hour call: the d5s meeting transcript against four well-known open-weight speech models, on the same audio.

Key findings

• The lowest WER did not catch the most words. Whisper large-v3-turbo had the lowest WER (14.1%). The d5s transcript caught the most words (94.5% word recall, 95.8% of content words).
• d5s missed the fewest words: 187, against 276 to 345 for the four models. Most of its WER came from repeats and false starts that people really said.
• Speaker names were right for 96.7% of the speech time in the d5s transcript.
• Published WER depends on the audio more than on the model. The same commercial models score 1% to 45% WER, from clean read speech to phone calls.

What is word error rate (WER), and how do you calculate it?

WER is the number of word errors divided by the number of words in a reference transcript.

A reference transcript is the "correct" text of the audio. Three types of error count:

• Substitution: a wrong word in place of the right one.
• Deletion: a word that was said, but is not in the transcript.
• Insertion: a word in the transcript that is not in the reference.

WER = (substitutions + deletions + insertions) / reference words. For example, a transcript of a 10-word sentence with 2 extra words has a WER of 2 / 10 = 20%. That is about 20 edits per 100 words to get the reference text. WER can go above 100%, because insertions have no limit.

Accuracy = 100% − WER. So "99% accurate" means a WER of 1%. When a vendor gives an accuracy number, ask which audio and which reference it comes from.

What is word recall?

Word recall is the share of reference words that the transcript caught, in the right order.

Extra words do not lower recall. Only missed and wrong words do. So word recall answers one question: did the transcript lose anything?

A variant is content-word recall. It skips small words such as "the", "and" and "yeah", and counts only names, verbs and terms. That is close to what a reader of meeting notes needs.

The two transcripts above have the same WER. But the verbatim one kept every word, and the clean one lost "signed" and "on". In a meeting, a lost word is often a lost name, number or commitment.

WER vs word recall on a real one-hour meeting

On our call, the transcript with the lowest WER was not the one that caught the most words.

The call was a 57-minute internal meeting in English, with 3 speakers, product names and some cross-talk. We compared the d5s meeting transcript with four open-weight speech models on the same audio: Whisper large-v3, Whisper large-v3-turbo, NVIDIA Canary 1B v2 and NVIDIA Parakeet TDT 0.6B v3.

Whisper large-v3-turbo wrote the cleanest text. The d5s transcript caught the most words. Both results are true at the same time, because the two numbers measure different things.

The d5s transcript is verbatim. When a person says "so, so we", it writes both words. The reference keeps one, so WER counts an error. In a meeting transcript, we think this is the right trade: a repeated word costs a reader almost nothing, but a missed word can hide a decision.

The speaker matters too. For 96.7% of the speech time, the d5s transcript named the same person as the meeting itself. The agent writes its meeting notes from that transcript.

Why vendor benchmarks look better than your meetings

The same models score from 1% to 45% WER. The type of audio explains most of the gap.

Most published numbers come from short, clean clips with one speaker. A meeting is long. People talk at the same time, use names and product terms, and share one microphone.

Some vendors publish no method at all. A "99% accurate" claim with no test set, no reference and no WER cannot be compared with anything. On meeting sets where the speaker must also be right, the best result in AssemblyAI's October 2026 tables is about 30% cpWER.

How to measure meeting transcription accuracy

Measure on your own long meetings, and report WER, word recall and speaker accuracy together. One number on a short clip hides most meeting problems.

• Use long, real meetings. Take calls with your accents, languages, names and cross-talk. A 30-second clip hides the problems of a one-hour call.
• Build a good reference. A person checks the text by ear. If you merge several transcripts, say so, because the merge favors the transcripts it came from.
• Normalize before you score. Remove case, punctuation and fillers in the same way for every transcript, for example with the Whisper text normalizer.
• Report WER and word recall together. WER tells you the editing work. Recall tells you what is lost.
• Check the speakers. Report cpWER, or the share of speech time with the right speaker.
• Count names and terms separately. A transcript can have a low WER and still miss every customer name.

We test agents the same way: real tasks and independent checks, not a benchmark table. Read how our agent swarms help test d5s.

How to improve meeting transcription accuracy

Most gains come from complete audio, key terms, speaker names from the meeting and set languages. These are the methods with the best evidence, from vendor docs and recent research:

• Hear all of the audio. No model can recover words that never reached it. Measure how much of the call the transcriber received.
• Give the model your key terms. Several engines accept a list of names and product terms, for example Deepgram keyterm prompting and ElevenLabs Scribe keyterms.
• Take speaker names from the meeting. Meet, Teams and Zoom know who speaks. That is more reliable than a guess from one mixed audio track.
• Set the call languages. Language detection can fail on short or mixed-language turns. In d5s, you can select up to 5 languages per call.
• Combine transcripts, but do not let an LLM rewrite them. Each model makes different errors, so a merge can beat each single model. In a March 2026 study, an LLM fixed names best when it selected between the versions that the models wrote, and WER did not get worse.
• Filter silence. Whisper models can invent text such as "thank you" in silence. A voice activity detector before the model, and a list of known phrases after it, remove most of these.

What this means for AI meeting follow-up

Ask for WER and word recall on long meetings like yours, then check the names, numbers and commitments yourself.

Ask a transcription vendor for two numbers, not one: WER and word recall, on long meetings like yours. Then read a transcript of your own call, and look for the names, numbers and commitments.

In d5s, an agent joins your Google Meet, Microsoft Teams or Zoom call as a listener. It writes notes with decisions and owners from the transcript, and it asks before it acts on them. Read how to bring an agent into a meeting, and see how d5s protects your data.

Questions and answers