How accurate are AI meeting minutes? Three layers can go wrong: the words, the speakers and the summary.
Short answer. Accurate enough to draft from, not to file unchecked. In open benchmarks, Whisper large-v3 gets about 1.6 words in 100 wrong on read audiobooks but 13.6 in 100 on the AMI meeting recordings, and open speaker-labelling software has an error rate of 17% to 20% of speech time on the same recordings. Accuracy also varies by language: OpenAI’s paper reports 4.2% word error for English and 21.5% for Hindi on an earlier Whisper model. The summary and the action-item owners add a third layer of risk. Good minutes tools therefore link each item to the transcript line and ask a person to confirm owners, which is how ANVE’s minutes work.
01
Layer 1: the transcript
A word error rate (WER) counts wrong, missing and extra words as a share of the words spoken, so 13.63% means about 14 errors in every 100 words. These are Whisper large-v3 results from the Hugging Face Open ASR Leaderboard, last updated on 2 October 2026.
Test
Word error rate
What the audio is
LibriSpeech clean
1.56%
Audiobooks read aloud, clean recording
LibriSpeech other
3.52%
Audiobooks read aloud, noisier
SPGISpeech
2.71%
Financial meetings
Earnings22
7.99%
Earnings calls
Voice Arena Monsoon
3.95%
Conversational English with an Indian accent
AMI-Cleaned
13.63%
Spontaneous multi-speaker meetings
Private (conversational)
15.13%
Aggregated private conversational data
02
Language and accent change the result
OpenAI’s Whisper paper reports word error on the FLEURS read-speech test for Whisper large-v2 of 4.2% in English and 21.5% in Hindi. OpenAI’s model card says large-v3 cuts errors by 10% to 20% compared with large-v2, but also that the models perform unevenly across languages and show different error rates across accents and dialects. It adds that the model may produce text that was not spoken and can repeat phrases.
03
Layer 2: who said what
Diarization is the step that labels each stretch of speech with a speaker. The pyannote.audio project publishes diarization error rates, where lower is better, for its open pipelines. The AMI corpus is 100 hours of meetings recorded with close-talking and far-field microphones; about two-thirds are scenario-based role-plays and the rest are natural meetings. IHM means individual headset microphones and SDM a single distant microphone.
AMI meeting corpus (pyannote.audio README, September 2025)
Older pipeline 3.1
community-1
Individual headset microphones (IHM)
18.8%
17.0%
Single distant microphone (SDM)
22.7%
19.9%
04
Layer 3: the summary and the action items
A language model can drop an item, merge two decisions, or give a task to the person who asked for it instead of the person who must do it. That last error is the one that matters most in minutes. ANVE’s own published test, five synthetic meetings that its note calls a smoke test and not a benchmark, caught a 3B-class model assigning an owner wrongly, which is why ANVE uses nothing smaller than 4B-class for minutes.
05
What improves accuracy
Record close to the speakers: the table shows distant microphones are harder.
Keep one person speaking at a time and state names at the start.
Keep the language consistent where you can, and check names and figures first.
Review the minutes while people still remember the meeting, and send action items to their owners for confirmation.
Require every item to quote the transcript line it came from.
06
How ANVE’s minutes check themselves
ANVE’s minutes maker produces transcripts in Hindi, English and Hinglish. Every action item must quote the transcript line it came from, and its owner must be confirmed by a person before the item is approved; each confirmation is logged. No accuracy figure for a shipping unit is published. ANVE is pre-release, and the software runs today on a development machine.
07
Questions people ask
What word error rate is good enough for minutes?
There is no universal threshold. Minutes need names, numbers and decisions to be right, so check those items first rather than relying on an average rate. Treat the transcript as a draft.
Does Whisper work for Hindi meetings?
Yes, with a higher error rate than English: OpenAI’s paper reports 21.5% word error for Hindi against 4.2% for English on FLEURS for Whisper large-v2, and says large-v3 cuts errors by 10% to 20% compared with it.
Why do minutes sometimes contain things nobody said?
Speech models can produce text that was not spoken or repeat phrases, as OpenAI notes for Whisper, and language models can invent details when summarising. Linking each item to its transcript line makes these visible.
Should a person review every set of minutes?
Yes for anything that creates an obligation or goes to a client or a regulator, and at least for every action item’s owner and due date.
08
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
