By whispertranscribe.ai Editorial Team · Last updated: November 2026 · 13 min read · Category: Transcription accuracy
How accurate is Whisper? On clean, read English, OpenAI's paper reports a 2.5% word error rate (WER) for Whisper large-v2, so about 97 to 98 words in 100 are right. Harder audio scores worse: 13.6% on phone calls, 24.9% at a noisy dinner party, and 39.9% in a meeting recorded by one distant microphone.
Whisper is OpenAI's open-source speech recognition model and the engine behind many transcription apps, including ours. This guide covers accuracy only: what WER measures, what the benchmarks say by audio type and language, where Whisper fails, and how to test it yourself. For the model's background, see our explainer on what Whisper AI is.
Most figures come from the Whisper paper (Radford et al., 2022), which tested every model "zero-shot", with no fine-tuning on the test sets. The newer large-v3 model card reports 10% to 20% fewer errors than large-v2 across many languages, but without a full per-dataset table. whispertranscribe.ai publishes this article. We didn't run a benchmark for it, and no number here is our own test result.
Whisper accuracy at a glance
No single test set answers the question, because the audio moves the number more than anything else. The table shows OpenAI's results for large-v2 with beam search and temperature fallback (appendix Table 9), scored after OpenAI's text normalizer.
| Audio type | Test set in the paper | Whisper large-v2 WER | Rough share of words right |
|---|---|---|---|
| Audiobook, clear reader | LibriSpeech test-clean | 2.5% | About 97.5 in 100 |
| Read news text | WSJ | 2.6% | About 97.4 in 100 |
| TED talks | TED-LIUM 3 | 3.7% | About 96 in 100 |
| Audiobook, harder speakers | LibriSpeech test-other | 4.9% | About 95 in 100 |
| Volunteers on home microphones | Common Voice 5.1 | 8.2% | About 92 in 100 |
| Phone conversations | Switchboard | 13.6% | About 86 in 100 |
| Long-form sociolinguistic interviews | CORAAL | 14.2% | About 86 in 100 |
| Phone calls between family and friends | CallHome | 16.4% | About 84 in 100 |
| Meetings, each person on a headset mic | AMI IHM | 17.4% | About 83 in 100 |
| Dinner parties, far-field mics | CHiME-6 | 24.9% | About 75 in 100 |
| Meetings, one distant table mic | AMI SDM1 | 39.9% | About 60 in 100 |
"Share of words right" is 100 minus WER, a shortcut, since WER also counts added words.
You'll often see 2.7% quoted instead: that's the same model with simpler greedy decoding (the paper's Table 2). Also note the AMI rows are the same meetings recorded two ways. Moving the mic from each speaker's head to the middle of the table more than doubles the error rate.
What is word error rate (WER), and how is it calculated?
Word error rate is the share of words a transcript gets wrong compared with a human-checked reference. Count three kinds of error and divide by the words in the reference:
WER = (substitutions + deletions + insertions) ÷ words in the reference
- Substitution: a wrong word in place of the right one ("lunch" for "launch").
- Deletion: a spoken word that's missing from the transcript.
- Insertion: a word in the transcript that nobody said.
Here's a self-made example, scored with the open-source jiwer library:
| Line | Text |
|---|---|
| Reference (22 words) | we moved the product launch to march ninth because the vendor missed two deadlines and doctor patel wants a new test plan |
| Transcript | we moved the product lunch to march ninth because vendor missed two deadlines and doctor battle wants a new test plan today |
| Errors | 2 substitutions (lunch, battle), 1 deletion (the), 1 insertion (today) |
| WER | 4 ÷ 22 = 18.2% |
Three things about WER catch people out:
- It can go above 100%. Insertions have no ceiling, and a few low-resource languages in the paper's Fleurs results score above 100%.
- Normalization changes the score. "Dr." versus "doctor" counts as an error unless both texts are cleaned the same way. OpenAI says its normalizer cut WER by up to 50% on some datasets, mostly by removing formatting differences.
- Some languages use character error rate (CER). For Chinese, Japanese, Thai, Lao, and Burmese, which don't separate words with spaces, the paper scores characters instead.
An accuracy claim without a named test set and scoring method isn't a measurement, whoever makes it, us included.
How accurate is Whisper in English?
Whisper is very accurate on clear English and stays usable on messy English. In the paper, it made 55.2% fewer errors across 12 other English datasets than a supervised wav2vec 2.0 model with almost the same LibriSpeech score.
Long recordings
Whisper reads audio in 30-second windows, so the paper tested long files separately (appendix Table 16, large-v2):
| Long-form test set | What's in it | WER |
|---|---|---|
| TED-LIUM 3 | 11 full TED talks | 3.5% |
| Meanwhile | 64 late-night TV segments, heavy on unusual words | 5.1% |
| Kincaid46 | 46 mixed recordings from a 2018 comparison | 8.8% |
| Earnings-21 | Company earnings calls | 9.7% |
| Rev16 | 16 podcast episodes | 11.3% |
| Earnings-22 | More earnings calls, many accents | 12.6% |
| CORAAL | 231 full interviews in African American English | 19.6% |
Whisper versus human transcribers
OpenAI had 25 Kincaid46 recordings transcribed by four human services and one computer-assisted service. The computer-assisted service scored 1.15 percentage points better than Whisper, and the purely human services only a fraction of a point better. The paper's verdict: "not perfect but very close to human-level accuracy."
Model size matters
Whisper comes in several sizes, and apps don't always say which one they run (appendix Table 9):
| Model | Parameters | LibriSpeech clean | Common Voice 5.1 | Switchboard | CHiME-6 |
|---|---|---|---|---|---|
| tiny | 39M | 6.7% | 26.1% | 18.3% | 48.0% |
| base | 74M | 4.9% | 19.4% | 15.6% | 40.0% |
| small | 244M | 3.3% | 12.8% | 13.3% | 30.9% |
| medium | 769M | 2.7% | 9.7% | 13.2% | 27.6% |
| large-v2 | 1550M | 2.5% | 8.2% | 13.6% | 24.9% |
The gap grows as audio gets harder: about 4 points on clean audiobooks, 23 points at the dinner party. The newer turbo model, an optimized large-v3, accepts what OpenAI's GitHub README calls a "minimal degradation in accuracy" for much faster transcription.
How accurate is Whisper in other languages?
Outside English, accuracy depends mostly on how much training audio Whisper had in that language. The paper found a squared correlation of 0.83 between log training hours and log WER, and estimates WER halves for every 16-fold increase in data. Most of the 96 non-English languages had under 1,000 hours.
Selected Fleurs results (read sentences, large-v2, appendix Table 13):
| Language | Fleurs WER | Note |
|---|---|---|
| Spanish | 3.0% | Lowest in the table |
| English | 4.2% | |
| Portuguese | 4.3% | |
| German | 4.5% | |
| Japanese | 5.3% | Character error rate |
| Russian | 5.6% | |
| French | 8.3% | |
| Korean | 14.3% | Flagged as an outlier |
| Chinese | 14.7% | Character error rate |
| Arabic | 16.0% | |
| Hindi | 21.5% | |
| Hebrew | 27.1% | Flagged as an outlier |
| Swahili | 39.3% | |
| Telugu | 99.0% | Flagged as an outlier |
The paper flags Hebrew, Telugu, Chinese, and Korean as doing worse than their training data predicted, possibly because of linguistic distance, a tokenizer that fits those scripts poorly, or data quality. Fleurs is read speech, so expect higher error rates on conversation. The README has a newer chart for large-v3 and large-v2 on Common Voice 15 and Fleurs.
Spanish results are encouraging if you transcribe Spanish audio to text. For lower-resource languages, test a short clip before committing a whole project.
What affects Whisper's accuracy?
Recording conditions matter more than anything else. Evidence from the paper and model card:
| Factor | What the evidence shows | What to do |
|---|---|---|
| Background noise | Under pub noise below 10 dB signal-to-noise ratio, Whisper beat all 14 comparison models, but every model got worse as noise rose | Record in a quiet room, away from fans and music |
| Microphone distance | Same AMI meetings: 17.4% on headset mics, 39.9% on one table mic | Keep the mic close to whoever is talking |
| Conversation and crosstalk | Phone calls and dinner parties score 13.6% to 24.9%, against 2.5% for read speech | One speaker at a time; one mic per speaker |
| Accents and dialects | The model card reports uneven accuracy across accents and dialects; long CORAAL interviews scored 19.6% | Test with speakers who sound like your real recordings |
| Rare words and jargon | Good on the jargon-heavy Meanwhile set (5.1%), yet names and technical terms remain common errors | Check names and terms against a list |
| Language resources | Fleurs WER runs from 3.0% (Spanish) to 99.0% (Telugu) | Choose the language yourself; plan more review for low-resource languages |
| Model size | tiny scores 48.0% on CHiME-6, large-v2 24.9% | Ask which model a tool runs, or use the largest you can |
| File format and sample rate | Whisper's code resamples all audio to 16 kHz mono | A 48 kHz WAV gains nothing over a clean 16 kHz file |
Converting an MP3 to a big WAV doesn't restore lost detail. A clean recording at a modest bitrate beats a noisy one at a high bitrate.
Known Whisper problems: hallucinations, loops, and timestamps
OpenAI documents each of these failure modes, which matter more than their share of WER suggests.
Hallucinated text
Whisper can write words nobody said. The model card says predictions "may include texts that are not actually spoken in the audio input," likely because the model also predicts probable next words. The paper's limitations section even mentions "complete hallucination," output unrelated to the audio.
An independent study, Careless Whisper (Koenecke et al., 2024), ran 13,140 audio segments through OpenAI's Whisper API in April and May 2023. About 1% of transcriptions (1.4% on average) contained whole invented phrases or sentences, and 38% of those included harmful content such as violence or false associations. Speakers with longer pauses, here people with aphasia, triggered more. The API has changed since, so treat this as a 2023 measurement.
Repetition loops and dropped words
The model card says Whisper is "prone to generating repetitive texts," which beam search and temperature scheduling reduce "but not perfectly." The paper also lists skipping the first or last words of a segment. Its decoding fixes cut average long-form WER from 11.0% to 10.0%: better, not solved.
Silence, music, and filler
Invented text tends to appear in long silences and music. The paper found Whisper's own "no speech" signal unreliable alone, and the open-source code now has an option to skip long silences when a likely hallucination is detected.
Timestamp drift and long files
Whisper moves through long audio one 30-second window at a time, using its own predicted timestamps to place the next window, so the paper notes a mistake in one window can carry into the next. Word-level timestamps are marked "experimental" in the code. For subtitles, spot-check timing around pauses.
During development, OpenAI also saw Whisper guess speaker names that were "almost always incorrect." The released models were fine-tuned to curb this, but check any name in the text.
10 ways to improve Whisper transcription accuracy
Most gains come before you press transcribe. Ordered roughly from biggest effect to smallest:
- Put the microphone close. Distance alone more than doubled the error rate in the AMI meetings.
- Cut noise at the source. Close windows, turn off fans, pause music.
- Avoid crosstalk. One person speaking at a time beats any setting.
- Trim long silences and music intros, where invented text most often appears. Cut at pauses, not mid-sentence.
- Set the language yourself. Whisper identified the language correctly 80.3% of the time across 82 Fleurs languages, so auto-detect can miss on short or accented clips.
- Use the largest model available, especially for noisy recordings.
- Prompt with names and terms if you self-host. Open-source Whisper accepts an
--initial_promptwith the spellings you expect. - Upload the original file. Re-recording through speakers or re-compressing repeatedly throws detail away.
- Run a term check. Search for every name, number, date, and acronym, and verify it against the audio.
- Review in two passes. First fix errors at normal speed. Then scan for hallucination signs: repeated lines, sentences over silence, sign-offs like "Thanks for watching." For legal, medical, or published work, have a person review everything.
How to measure Whisper's accuracy on your own audio
The best estimate for your work is a 30-minute test on your own recordings. These steps are for you to run; they aren't a whispertranscribe.ai test.
- Pick three 2-minute clips like your real work: one easy, one typical, one hard.
- Write a reference transcript by hand, word for word. Set rules first: keep or drop "um," digits or words for numbers.
- Transcribe each clip with the tool you're testing and save plain text.
- Normalize both texts the same way. OpenAI's normalizer ships with
openai-whisper. - Calculate WER with jiwer:
# pip install jiwer openai-whisper
import jiwer
from whisper.normalizers import EnglishTextNormalizer
norm = EnglishTextNormalizer()
ref = norm(open("reference.txt", encoding="utf-8").read())
hyp = norm(open("transcript.txt", encoding="utf-8").read())
out = jiwer.process_words(ref, hyp)
print(f"WER {out.wer:.1%} S={out.substitutions} D={out.deletions} I={out.insertions}")
- Count critical errors separately. Five wrong fillers and five wrong client names give the same WER but aren't the same problem.
- Log results so you can compare tools or settings:
| Clip | Conditions | Reference words | S | D | I | WER | Critical errors |
|---|---|---|---|---|---|---|---|
| Clip A | Quiet room, one speaker | 310 | |||||
| Clip B | Video call, two speakers | 285 | |||||
| Clip C | Café, accented speaker | 296 |
Word counts are placeholders. A caution on step 4: when we ran OpenAI's English normalizer on a self-made sentence, it matched "Dr." to "doctor" and "two" to "2" but still scored "9th" against "ninth" as an error. Read flagged differences before trusting the score.
How accurate is whispertranscribe.ai?
whispertranscribe.ai is a web-based AI transcription tool powered by Whisper that supports 130+ languages, with uploads up to 5GB and 10 hours. Accuracy reaches up to 99% on clear audio; background noise, strong accents, and crosstalk lower it, as the benchmarks above show for Whisper itself. We don't claim to match or beat the paper's figures, and we haven't published our own benchmark, so test your own clips with the steps above.
- Noise / Low-Quality Audio/Video options in the language picker for many languages. On a noisy file, try the standard and noise options and compare.
- 130+ languages, a Multilingual Mix option, and translation of audio into English.
- Files up to 5GB and 10 hours, speaker detection, and timestamps (all features).
- Clean input: record in your browser with the mic close, or convert video to MP3 to upload only the audio.
- Exports to DOCX, PDF, TXT, CSV, SRT, and VTT, plus a subtitle generator for timed captions.
You can start free with no account required. Paid plans start from $9.99/month billed annually; see pricing.
FAQs
Is Whisper 99% accurate?
Not in general. Whisper's best published result is a 2.5% word error rate on LibriSpeech test-clean, a set of clearly read audiobooks, which is about 97.5% of words correct. On phone calls, meetings, and noisy rooms, OpenAI's own figures range from about 13% to 40% WER. A "99%" claim only means something if it names the audio it was measured on. Tools built on Whisper, including whispertranscribe.ai, should be judged on your recordings, not on a headline percentage.
Which Whisper model is the most accurate?
The large models are the most accurate. In OpenAI's paper, large-v2 had the lowest error rates on most test sets, and the large-v3 model card reports 10% to 20% fewer errors than large-v2 across many languages. The turbo model is an optimized large-v3 that OpenAI says loses only a little accuracy in exchange for much more speed. The smallest model, tiny, is far less accurate on hard audio: 48.0% WER at a noisy dinner party, against 24.9% for large-v2.
Why does Whisper write words that weren't said?
Whisper can hallucinate because it's trained to predict likely text as well as transcribe sound, and OpenAI's model card warns that it may output text that isn't in the audio. It happens most often during long silences, music, or noise. A 2023 independent study found hallucinated phrases in about 1% of transcriptions, and pauses made them more likely. Trimming silence, setting the language, and reviewing quiet stretches of the transcript reduce the risk.
How accurate is Whisper with accents?
Whisper handles accents better than many older systems, but not evenly. OpenAI's model card says accuracy varies across accents and dialects of the same language. In the paper, long interviews in African American English (CORAAL) scored 19.6% WER, compared with 3.5% for full TED talks, though recording conditions differ too. The reliable approach is to test a clip with speakers who sound like your real recordings, and to plan for more review when accents are strong.
Is Whisper as accurate as a human transcriber?
For English, Whisper comes close but doesn't match the best. OpenAI compared it with five professional services on 25 recordings. A computer-assisted service had an aggregate WER 1.15 points lower than Whisper's, and four fully human services were a fraction of a point better. Humans still tend to handle names, crosstalk, and unclear speech better. For anything quoted, filed, or published, a person should review the transcript against the recording.
How accurate is Whisper for languages other than English?
It varies widely by language. On the Fleurs benchmark, large-v2 scored 3.0% WER in Spanish, 4.5% in German, and 8.3% in French, but 27.1% in Hebrew and 99.0% in Telugu. Chinese and Japanese are scored by character error rate: 14.7% and 5.3%. The paper found that accuracy tracks how much training audio each language had. For lower-resource languages, test a short clip first and budget more time for corrections.
Test accuracy on your own recording
Benchmarks describe other people's audio. Upload a clip to whispertranscribe.ai in any of 130+ languages and score it with the steps in this guide. Start free, no account required, or compare plans.
Related articles: What is Whisper AI? · Best free transcription software · Best TurboScribe alternatives · How to transcribe a YouTube video · What is an SRT file? · Hard subtitles vs soft subtitles
References
- Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (the Whisper paper, OpenAI, 2022): Tables 2, 9, 13, and 16 and Sections 3 to 6
- OpenAI, openai/whisper on GitHub: README model table, turbo note, and per-language WER chart
- OpenAI and Hugging Face, whisper-large-v3 model card: large-v3 changes, evaluated use, and limitations
- Koenecke et al., Careless Whisper: Speech-to-Text Hallucination Harms (2024)
- jiwer: open-source library for WER and CER
- Hugging Face, Open ASR Leaderboard: independent, regularly updated comparison of open speech models
