How Accurate Is AI Transcription? We Tested It on Real Meetings

AI transcription accuracy is close to human level on clear audio, but on real meetings AI makes several times more errors, mostly because people talk over each other. You measure it word by word against a human-made transcript, like these examples of transcription. Below, we answer "how accurate is AI transcription?" in order: benchmarks, word error rate, our test, what causes mistakes, AI versus humans, fixes.

How accurate is AI transcription?

On clear, read-aloud audio, modern speech recognition gets 97 to 99 of every 100 words right; on meeting recordings, published benchmarks show several times more errors, especially on a distant microphone. AI transcription accuracy is reported as word error rate, the share of words a system gets wrong.

In the Whisper paper (Radford et al., 2022), Whisper Large V2 scored 2.7% on LibriSpeech, a benchmark of audiobooks read aloud. Newer transcription software does better: the top English model on the Hugging Face Open ASR Leaderboard, last updated October 9, 2026, scores 1.11% on the same audiobooks and averages 4.7% across 11 test sets, from audiobooks to earnings calls and meetings.

The meeting rows come from the AMI Meeting Corpus, 100 hours of meetings recorded on close-talking and far-field microphones. They date from 2022, so read them as the gap between a headset and a room microphone, not as today's error rate. Our test, below, runs 9-minute excerpts from the same corpus through Meetlens in 2026; it scores whole stretches with crosstalk, so its numbers are not directly comparable with these rows.

How is transcription accuracy measured?

Transcription accuracy is measured with word error rate (WER): substitutions, deletions and insertions divided by the number of words in a human reference transcript. Accuracy is 1 − WER. Microsoft's speech documentation defines the three transcription errors: a substitution is a wrong word, a deletion is a missing word, and an insertion is a word nobody said.

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

Word error rate, part by part

WER =

SSubstitutions

a wrong word

e.g. “Tuesday” for “Thursday”

+

DDeletions

a missing word

e.g. “revised” left out

+

IInsertions

a word nobody said

e.g. “over” added

divided by

NWords in the human reference (N)

e.g. 10 words

Accuracy = 1 − WER

Illustrative example 1 deletion, 1 insertion, 1 substitution: 3 / 10 = 30% WER, or 70% accuracy

Because insertions count, WER can exceed 100%. Before scoring, both texts go through text normalization: lowercase, no punctuation, one form for numbers ("zero" becomes "0"), and filler words such as "um" and "uh" removed. The Open ASR Leaderboard documents these steps on its About tab.

What does "95% accurate" actually mean?

95% accurate means 1 wrong word in every 20 spoken; 90% accurate means 1 in 10. Our 9-minute test excerpts averaged 1,780 words, so even 95% accuracy leaves about 89 wrong words per excerpt. Microsoft's guide, on the same page, calls a WER of 5–10% good quality and 30% or more poor quality.

WER weighs every mistake the same: "Tuesday" for "Thursday" costs as much as "a" for "the". It also ignores context, punctuation and who spoke, so our test adds metrics for critical terms and speaker labels.

Our test: how accurate is Meetlens on real meetings?

On a four-person design meeting where people mostly took turns, Meetlens reached 85% to 86% transcription accuracy: 14.0% WER on clip-on microphones, 14.5% on one microphone in the middle of the table and 14.8% on headsets. On clear speech, with nobody else talking, Meetlens missed or got wrong 12 words in 100 (11.8% across all eight test excerpts).

When people talked over each other, the error rate rose: on the five excerpts where 36% to 55% of all words overlapped another speaker, WER was 28.7% to 42.4%, up to about 42 in 100 words on a lively student meeting. Crosstalk was the main cause: across all eight excerpts, words spoken over someone else were missed or wrong half the time (50.4%).

Meetlens is the product we build: a Chrome extension that records a browser meeting's audio without a bot joining the call and produces a transcript with speaker labels and timestamps. For this test we uploaded recordings as files, which go through the same transcription step as recordings made with the extension. Scoring our own tool cuts both ways, so every result stays in.

How we tested:

1. Test recordings. Eight excerpts of about 9 minutes, cut at pauses from four four-person meetings in the AMI Meeting Corpus (AMI Meeting Corpus, University of Edinburgh et al., CC BY 4.0): a product design meeting (E1–E3), a design meeting full of crosstalk (E4), a design meeting of non-native speakers (E5–E6) and a real project meeting of MSc students (E7–E8). In the three design meetings, participants play a product team without a script. AMI recorded every meeting on several microphones at once, so three stretches appear more than once: on headsets, clip-on microphones or one microphone in the middle of the table. We uploaded the files to Meetlens as MP3s; the only change to the audio was one fixed volume level per file.
2. Reference transcript. AMI's own manual word-level transcript, made by the corpus's transcribers. We dropped cut-off word fragments and unified AMI's spelling conventions ("L_C_D_" becomes "LCD").
3. Scoring. Both texts went through Whisper's English text normalizer, and the open-source jiwer library (version 4.0.0) counted substitutions, deletions and insertions. Where people talk at once, the reference lists their words in the order they started speaking.
4. Extra metrics. Critical terms: 29 to 31 names, numbers, dates and product terms per meeting, listed from the reference before any upload (519 occurrences in all), each scored as exactly right or not. Speaker label errors: reference turns whose words landed under another person's speaker label, after matching each label to the participant it shares most words with. Invented phrases: inserted runs of 3 or more words that are not in the audio, checked by hand.

Test date: October 16, 2026. Meetlens web app: production build of October 14, 2026 (files uploaded through Import, General Meeting format). Time from upload to a finished transcript: about 2 to 4 minutes for a 9-minute file.

The archive holds the eight audio files, AMI's reference transcripts, the Meetlens output, per-excerpt results and our scoring scripts: download the test data (ZIP, 50 MB).

Results for the design meeting, where people mostly took turns:

E1, E2 and E3 are the same 8.7-minute stretch of the design meeting, recorded at the same moment on three kinds of microphone. Names and product terms came out right more often than numbers. Most speaker label errors were short replies of under five words, such as "yeah"; among turns of five words or more, 2% to 6% carried the wrong label, and Meetlens found all four speakers in every excerpt. None of the inserted runs of three or more words was invented: each was real speech, misheard or moved.

Limits. Eight excerpts from four meetings is a small sample: 71 minutes, all in English, recorded in quiet meeting rooms, and the scores describe one product version. Three of the four meetings are role-played, though unscripted. AMI's reference keeps every repeated word and "yeah", which a transcript may leave out, and those count as errors; so do spelling differences such as "touch screen" against "touchscreen" (about 1.3 points on E1–E3). Where people talk at once, any word order is a choice; ordering the reference differently moved each score by less than one point. We did not test a noisy participant, a client call or the extension inside Google Meet, Zoom or Teams. The Meetlens transcription page itself names overlapping speech, distant microphones, heavy noise and unfamiliar terminology as sources of mistakes; this test measures what overlapping speech, distance, accents and names cost in AI transcription accuracy.

To check accurate AI transcription on your own setup, record one meeting on the Free plan (180 minutes a month, no card needed) and compare a few minutes of the transcript with the audio.

What affects AI transcription accuracy?

Overlapping speech hurts AI transcription accuracy most in meetings; names and numbers, accents, microphone distance and noise add errors on top. Most of it comes down to audio quality. Microsoft's guide ties many missing words to a weak audio signal, insertions to noise or to people talking over each other, and substitutions to domain-specific terms.

Overlapping speech

Overlapping speech loses words when multiple people talk at once and blurs who said what. Microsoft's guide notes that overlapping speakers can confuse the model and make words drop out. In our test, words spoken while someone else was talking were missed several times more often than clear speech, and the excerpts with the most crosstalk also had the most speaker label errors.

Names, numbers and jargon

Proper names, numbers and complex jargon are where one error costs the most. Microsoft's guide links substitutions to domain terminology and to product and people's names. That is why our test also scores critical terms; in it, names and product terms came out right more often than numbers.

Accents and dialects

Transcription accuracy is not equal across speakers. A 2020 PNAS study of five systems from Amazon, Apple, Google, IBM and Microsoft found an average WER of 0.35 (35%) for Black speakers and 0.19 (19%) for white speakers. The Black speakers' interviews came from a corpus of African American Language, so dialect is part of that bias. In our test, a meeting of non-native speakers (French, Wolof and Chinese) scored worse overall, but on stretches where only one person was speaking, non-native speakers' words were missed about as often as native speakers'; most of the gap came from crosstalk.

Microphone distance

A distant microphone doubled the error rate in 2022 benchmarks; in our quiet-room test it barely mattered. In the AMI rows of the first table, the same meetings had more than twice the error rate on one distant microphone as on headsets. In our test, the same stretches recorded on one microphone in the middle of the table scored almost the same as on headsets, though speaker labels slipped slightly.

Background noise

Background noise adds words nobody said and hides quiet ones. In the same Whisper paper, Whisper Large V2 scored 25.5% on CHiME-6, conversation recorded at dinner parties in everyday homes. Our test did not include noise: the AMI meetings were recorded in quiet rooms.

Switching languages

Calls that switch languages mid-conversation add their own mistakes; this test was English-only.

Is AI transcription as accurate as a human?

On recorded phone conversations, AI transcription has nearly caught up with professional transcribers, who make about 5 to 11 errors per 100 words, but the best human transcribers still make fewer errors on harder conversation, and AI can occasionally invent whole phrases. Whether AI transcription accuracy has reached human parity depends on which humans you measure.

Microsoft researchers measured professional transcribers at 5.9% WER on the Switchboard speech corpus, phone calls between strangers on an assigned topic, and 11.3% on CallHome, calls between friends and family; by the paper's 2017 revision, their system scored 5.8% and 11.0%. An IBM team re-measured humans in 2017: its best transcriber reached 5.1% and 6.8%, against 5.5% and 10.3% for IBM's own system.

In the Whisper paper, five professional transcription services transcribed 25 recordings, including meetings and phone calls. The computer-assisted service beat Whisper by 1.15 percentage points; for the four purely human services, the difference was a fraction of a point.

AI also fails in a way people don't: hallucination. In a study presented at ACM FAccT 2024, roughly 1% of Whisper transcriptions from spring 2023 contained whole phrases absent from the audio, more often for speakers with long pauses, and 38% of those carried harms such as violence or implied false authority. A December 2023 retest found improvement.

Automated transcription is fast and manual transcription costs more, so human work is shifting to review. The Bureau of Labor Statistics projects employment of medical transcriptionists to decline 4% from 2025 to 2035 and lists editing drafts from speech recognition software among their duties.

Is any AI 100% accurate?

No AI transcription is 100% accurate, and neither are professional transcribers. The Whisper paper notes that indistinct speech and labeling errors leave irreducible error in every dataset. In medical transcription, a 2025 systematic review of 29 studies found WERs from 8.7% when clinicians dictated under controlled conditions to over 50% in conversational or multi-speaker settings. Legal and medical records still need human proofreading.

How to get more accurate AI transcripts

Use a headset or a close microphone, avoid talking over each other, say names and numbers clearly, and review only the critical terms afterwards. Each step targets a factor that lowers AI transcription accuracy:

1. Wear a headset or sit close to your microphone.
2. Join from your own device rather than sharing one laptop's microphone in a room.
3. Mute when you're not speaking and find a quiet spot.
4. Take speaking turns, and answer with a full sentence ("Yes, I'll send it Friday") rather than a bare "yes", which is easy to lose or credit to the wrong speaker.
5. Say names, numbers and dates clearly, and spell an unusual name or acronym once.
6. Record from the browser tab rather than through speakers, which add room noise and distance.
7. Review the critical terms: names, numbers, dates, task owners and quotes. Limit editing to those, except in legal and medical records.

---

A correct transcript is the input — what about the notes made from it?

Transcript accuracy vs. AI meeting notes accuracy

Transcription accuracy covers the words; the accuracy of AI meeting notes built from them is a separate question. A word-perfect transcript can still produce a summary that drops a "not", names the wrong owner for a task or records a suggestion as a decision. WER doesn't measure those mistakes, so check AI meeting notes against the transcript on their own; how accurate AI meeting notes are is a separate test.

Is it legal to use AI transcription?

The legal question is the recording, not the transcription software. Federal law allows recording when one participant consents (18 U.S.C. § 2511(2)(d)), but some states, such as California, require every party's consent to record a confidential conversation. Get recording consent from everyone before you press record.