How Accurate Is Whisper? Real Numbers, from Audiobooks to Meeting Rooms (2026)

Quick answer: There is no single accuracy number for Whisper — there is a ladder. On clean, single-speaker read speech, the large models reach roughly 2–3% word error rate (WER), or better than 97 words in 100. Across a mix of real English benchmarks, Whisper Large V3 averages about 7.4% WER. On meeting-room audio with distant microphones and overlapping speakers, the same model runs around 16%. And on low-resource languages the error rate can triple again. Whether Whisper is “accurate” depends on which rung your recording sits on — and, just as much, on which Whisper variant the app you use actually runs.

Every number in this article is taken from OpenAI’s published paper and model cards or from the public Open ASR Leaderboard, linked in the sources at the end. It was written by the LoroNote team on August 19, 2026. LoroNote runs Whisper Large V3 Turbo on the device, so we have a stake here — which is exactly why nothing below is our own marketing figure.

What word error rate actually measures

WER counts the words a transcription got wrong — substituted, dropped, or invented — as a share of the words spoken. A 5% WER means one error in every 20 words; 16% means one in six. Two things follow from the definition:

1. A WER only means something next to the audio it was measured on. “98% accurate” with no named test set is a slogan, not a measurement. The same model produces 97%+ on an audiobook and 84% in a conference room without anything being wrong.
2. Not all errors cost equally. A dropped “um” and a wrong drug name both count once. WER is the standard yardstick because it is objective, not because it captures how usable a transcript feels — which is why testing on your own audio, covered below, beats any table.

We unpack the full definition — the formula, why WER can top 100%, and the normalization step that quietly moves published numbers — in What is word error rate?

The accuracy ladder: one model, four rungs

These are published figures for Whisper’s large models on standard benchmarks:

Read the ladder top to bottom and the pattern is plain: the model does not get worse — the audio does. Dictation into a phone held near your mouth lives near the top rung. An interview across a café table lives in the middle. A meeting recorded from the far end of a table lives on the third rung, before a single software choice has been made.

Which Whisper? The variant changes the answer

“Powered by Whisper” covers a family, not a model. Whisper ships in sizes from tiny to large, and the spread between them is larger than the spread between vendors:

• Large V3 is the accuracy flagship. OpenAI trained it on 1 million hours of weakly labeled audio plus 4 million hours of pseudo-labeled audio, and reports 10–20% fewer errors than Large V2 across a wide range of languages.
• Large V3 Turbo is the speed-optimized derivative: 809 million parameters, with the decoder cut from 32 layers to 4. OpenAI describes the cost as “minor quality degradation” — on the Open ASR Leaderboard the averages sit close together (about 7.8% vs 7.4% mean WER; about 16.1% vs 16.0% on AMI). Combined across languages it performs similarly to Large V2, and it loses the most in a few languages such as Thai and Cantonese. We took apart that trade-off in our Whisper Large V3 Turbo guide.
• Small and medium run several points higher WER — a real gap you can see in a one-paragraph test. This matters on iPhones specifically, because several well-known “Whisper apps” run small or medium on the phone even when their Mac versions run large. Our Whisper apps comparison lists which variant each iOS app actually loads.

The practical rule: before comparing apps on accuracy, check which Whisper each one runs. An app running Large V3 Turbo and an app running small are not offering the same accuracy, whatever both listings say about “Whisper AI.”

Language moves the number more than anything else in software

Whisper’s training data is heavily skewed toward high-resource languages, and its accuracy follows. English, Spanish, German, Japanese, Korean, and other well-represented languages cluster near the averages above. Low-resource languages sit far higher — the Hindi figure in the table is one published example, and OpenAI’s model card notes uneven performance across accents and dialects within languages, too.

Whisper Large V3 covers about 100 languages (Cantonese was added with V3), so “supported” and “equally accurate” are different claims. If your language is outside the high-resource tier, the honest advice is the same as everywhere in this article: run a short recording of your own and look at the result — the full language roster tells you what LoroNote will attempt, and two minutes of your own audio tells you how well.

Hallucinations: the error WER understates

OpenAI’s model card states it directly: because the models are trained on large-scale noisy data, “the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination).” In practice this shows up on silence, background noise, and music — Whisper is a generative model, and when there is no speech to anchor it, it can produce fluent sentences nobody said.

Two things keep this manageable:

• Apps can mitigate it. Voice-activity detection that skips silent stretches removes the model’s main opportunity to invent text. This is an app-level behavior, not a model setting, and it is one of the real differences between apps running the same Whisper.
• You can spot it. Hallucinated text tends to appear at stretches you know were quiet — the end of a recording, a long pause. If a transcript’s middle is solid but its silences grew sentences, you are looking at hallucination, not mishearing.

We compared this behavior against a transducer-architecture model in Parakeet V3 vs Whisper Large V3 — transducers hallucinate less on silence, while Whisper covers vastly more languages; there is no free lunch.

What actually moves accuracy on your recordings

Benchmarks vary the audio; your recordings vary the same things. In rough order of impact:

1. Microphone distance. The single biggest lever. A phone at arm’s length on the table beats a phone across the room by more than any model upgrade.
2. Noise and music. Steady background noise degrades gracefully; music and clattering degrade badly.
3. Overlapping speakers. Cross-talk is why meeting corpora sit at ~16% while single-speaker audio sits at ~3%.
4. Specialized vocabulary. Names, drugs, tickers, and jargon fail disproportionately. LoroNote’s custom dictionary exists for exactly this: recurring terms it should stop getting wrong.
5. The model variant — the one factor above that is purely a software choice, which is why LoroNote runs the largest Whisper variant practical on the phone.

Speaker labels deserve a note: diarization does not change WER, but it changes whether a multi-speaker transcript is usable at all. An accurate transcript with no speakers is still homework; how speaker identification works covers that half of the problem.

How to measure Whisper’s accuracy yourself in ten minutes

A test on your own audio outranks every table in this article:

1. Record about two minutes that resemble your real use — same room, same distance, same language, real names and jargon.
2. Transcribe it and read the result against the audio.
3. Count the errors in a 200-word stretch: every substituted, missing, or invented word. Errors ÷ 200 = your WER.
4. Judge the errors, not just the count: five mangled filler words and five mangled client names are the same WER and a very different day.

The precise version of this method — including how to decide up front what counts as an error — is in our WER guide.

Because LoroNote transcribes on the device and its free version covers the first two minutes of every recording — exactly enough for this test — it costs nothing, and the recording never leaves your iPhone: no upload step, no account, and Airplane Mode changes nothing.

Frequently Asked Questions

Is Whisper accurate enough for professional transcripts?

For clean, close-miked recordings in high-resource languages, Whisper’s large models produce drafts most people only lightly edit — the 2–3% clean-speech WER is near the practical ceiling for automatic systems. For distant-mic meetings, heavy cross-talk, or critical exact wording, plan on review time: at 16% WER, one word in six needs your attention.

Does Whisper Large V3 Turbo lose accuracy compared to Large V3?

Slightly. OpenAI calls the degradation minor, and public leaderboard averages sit well under a point apart in English. Combined across languages, Turbo performs similarly to Large V2, with the largest losses in a few languages such as Thai and Cantonese. In exchange it is dramatically faster — the trade that makes a Large-series model practical on a phone.

Why does my transcript contain sentences nobody said?

That is hallucination, a documented behavior of Whisper on silence, noise, and music: a generative model with no speech to anchor it can produce fluent text anyway. Check the quiet stretches of your recording. Apps that run voice-activity detection before Whisper largely avoid it — it is an app behavior, not a user setting.

Does running Whisper on-device reduce its accuracy?

No. The same model weights produce the same transcription quality wherever they run; on-device inference is not a degraded mode. What differs between apps is which variant they run — cloud services can host the largest models easily, while on-device apps must fit the phone, which is why some run small or medium. LoroNote runs Large V3 Turbo locally, keeping a Large-series model and on-device privacy in the same app.

Conclusion

“How accurate is Whisper” has a shape, not a number: roughly 2–3% WER on clean read speech, about 7% across mixed English benchmarks, around 16% on meeting audio, and worse on low-resource languages — with hallucination as the caveat WER understates. The two choices that move your own results most are physical (get the microphone closer) and structural (know which Whisper variant your app runs). The rest is measurable in ten minutes with your own recording, which is the only benchmark that was ever about you.

Official Resources

• OpenAI: Robust Speech Recognition via Large-Scale Weak Supervision (Whisper paper)
• openai/whisper-large-v3 model card
• openai/whisper-large-v3-turbo model card
• Open ASR Leaderboard
• LoroNote: Why LoroNote runs Whisper Large V3 Turbo