Accuracy matters more than almost anything else about a transcription tool, which is exactly why "is Whisper accurate?" is the wrong question. Whisper's accuracy on a clean English podcast and its accuracy on a noisy recording in a low-resource language are not the same model having a good day and a bad day. They're two genuinely different results, and the gap between them is wide enough to decide whether the tool is useful to you at all.
One disclosure before the numbers: we didn't run a private lab study for this piece. We pulled the figures OpenAI and the public benchmarks already publish, and read them the way someone deciding whether to rely on the tool should. Every number here links to its primary source.
What WER actually measures
Word Error Rate is the number nearly every speech benchmark reports, and it's worth thirty seconds because the rest of this is numbers. Take the reference transcript — what was actually said — and count the edits needed to turn Whisper's output into it: every word substituted, deleted, or inserted. Divide by the number of words spoken. A WER of 6 means roughly six words in every hundred are wrong in some way.
Lower is better, zero is perfect, and — the detail that trips people up — WER can run past 100%. A model that invents extra words can rack up more errors than there were words to begin with. Keep that in your back pocket; it matters later.
The one number everybody quotes
The figure you'll see most is English, and it's genuinely good. On the
Open ASR Leaderboard — a public, reproducible benchmark that averages WER across several English datasets —
OpenAI's whisper-large-v3 lands at 6.43% WER.
large-v3 is no longer at the very top of that board — NVIDIA's Canary Qwen 2.5B leads at 5.63% — but a sub-7% average across varied English audio is the kind of number that gives you a transcript you mostly read rather than fix. The board also reports speed as RTFx (times faster than real time); large-v3 runs around 68× real time on the benchmark hardware, which is the other half of why it's the default workhorse for so many tools.
Why one number is misleading
Here's where the single-number framing falls apart. English isn't representative of Whisper — English is the best case, by a wide margin, and for a structural reason rather than an incidental one. Whisper learned to transcribe from the audio it was trained on, and that audio is distributed wildly unevenly across languages. The original model saw 680,000 hours of audio, and the paper is blunt about the split:
"Of those 680,000 hours of audio, 117,000 hours cover 96 other languages." — Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, 2022
Read that twice. English and a few close neighbours took the lion's share; ninety-six other languages split roughly a sixth of the data between them. The consequence is the single most important sentence in the paper, and OpenAI states it without hedging:
"performance on a language is very well predicted by the amount of training data for the language." — Radford et al., 2022 (arXiv:2212.04356), §3.4
That's close to a straight line: more hours, lower WER. You can roughly predict how well Whisper handles a language by asking how much of the internet's recorded audio happens to be in it. And OpenAI doesn't bury the takeaway — the official large-v3 model card repeats it in plain product language:
"Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data." — openai/whisper-large-v3 model card
So how does it do on your language?
Practically, that sorts languages into rough tiers — and the honest move is to point you at the source rather than invent a table. OpenAI publishes a full per-language breakdown for large-v3 (word error rate, or character error rate for languages that don't split on spaces) on the Common Voice 15 and FLEURS datasets, as a figure in the official Whisper repository. If a specific language matters to you, that chart is the thing to open. Its shape is exactly what everything above predicts:
- High-resource languages — English, Spanish, Italian, Portuguese, German, French, Polish — cluster near the bottom of the error range. Paste a clean recording and you can mostly trust the output.
- Mid-resource languages land in a usable-but-check-it middle. The transcript is a strong draft; names, technical terms, and homophones want a human pass.
- Low-resource languages — many African, South and Southeast Asian, and minority European languages — carry error rates high enough that the transcript is a starting point, not a finished product.
This is also why a tool's choice of Whisper model size hits some users far harder than others. If you work in English, almost any checkpoint is fine. If you work in Ukrainian, Telugu, or Swahili, the difference between a small and a large model is the difference between a usable draft and a frustrating one.
large-v3 vs the small model: what you trade
Model size is the other dial, and it trades accuracy for compute in a way that's easy to underrate. Whisper ships in a range of sizes, and the parameter counts tell the story:
| Model | Parameters |
|---|---|
| tiny | 39 M |
| base | 74 M |
| small | 244 M |
| medium | 769 M |
| turbo | 809 M |
| large‑v3 | 1,550 M |
Bigger is more accurate — most of all on accents and the lower-resource languages where the smaller checkpoints give up the most ground. large-v3 itself was a real step up: trained on five million hours of audio (one million weakly-labelled plus four million pseudo-labelled by its predecessor), with a finer 128-bin spectrogram input and a new Cantonese token, it delivers what OpenAI calls a "10% to 20% reduction of errors compared to Whisper large-v2."
The catch is cost. large-v3 carries 1.55 billion parameters against the small model's 244 million — roughly six times the weight — and it really wants a GPU to hold near-real-time pace, which small manages on modest hardware. large-v3-turbo sits between the two: a distilled decoder at 809 million parameters that keeps most of large-v3's accuracy for a fraction of its compute. That is what shrybe's free tier runs. The one place we still fall back to small is the iOS Shortcut, where a hard timeout on the share-sheet round trip leaves no room for anything heavier — and the Slavic and low-resource cases are exactly where you feel that drop.
The failure modes nobody averages
WER hides two failure modes that, in practice, matter more than a percentage point either way. The first is hallucination. Because Whisper was trained on large, noisy data, OpenAI warns that "the predictions may include texts that are not actually spoken in the audio input." Hand it a long silence, background music, or garbled audio, and it can confidently produce a sentence nobody said.
The second is repetition. The model card notes that "the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts" — the stuck-record loop where one phrase repeats for a paragraph. Neither shows up cleanly in an averaged WER, and both are why you proofread anything you're about to quote, regardless of how good the headline accuracy number looks.
What this means if you just need a transcript
So — is Whisper accurate? For mainstream-language audio that's reasonably clean, yes: accurate enough that the transcript is something you read, not something you repair. For low-resource languages, heavy accents, overlapping speakers, or rough audio, it's a fast first draft that still needs your eyes. The single most useful habit is to spot-check what WER under-weights — proper nouns, numbers, technical jargon, and anything headed for quotation marks.
That's the reasoning behind how shrybe works, too. Paste a YouTube link, a podcast episode, or an audio URL and you get a timestamped transcript in about thirty seconds — fast enough to be worth it, with timestamps so verifying a specific line means jumping straight to it instead of re-listening to the whole thing. If you need the text as on-screen captions, the caption generator wraps the same output into SRT and VTT.
Common questions
Is Whisper accurate enough for professional work?
For clean audio in a high-resource language, whisper-large-v3 averages around 6–7% word error rate — good enough that the transcript is a readable draft. For anything you will publish or quote, treat it as a draft and check proper nouns, numbers, and technical terms, which are exactly what error-rate averages under-weight.
Which languages does Whisper transcribe best?
The ones it saw the most training audio for. English is the strongest by a wide margin; major European languages such as Spanish, German, French, Italian, Portuguese, and Polish are close behind. Accuracy drops as you move to lower-resource languages, because — in OpenAI's words — performance on a language is very well predicted by the amount of training data for the language.
What counts as a good word error rate (WER)?
Under roughly 10% is generally a transcript you can read with light editing; under 5% is excellent. Above about 20% you are editing more than reading. WER can also exceed 100%, because a model that hallucinates extra words can produce more errors than there were words spoken.
Does a bigger Whisper model always transcribe better?
For accuracy, generally yes — large-v3 (1.55 billion parameters) beats the small checkpoint (244 million), most visibly on accents and low-resource languages. But bigger models need far more compute, so the right choice depends on whether you need speed and low cost or maximum accuracy.
Why does Whisper sometimes add words that were never spoken?
It is a known failure mode. OpenAI notes the model may include texts that are not actually spoken in the audio input — hallucination — and can also loop into repetitive text, most often on silence, music, or low-quality audio. It is the main reason to proofread before quoting.
Sources
Everything above is drawn from primary sources, not secondary summaries:
- Radford, Kim, Xu, Brockman, McLeavey, Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision (2022, arXiv:2212.04356) — training-hours split and the per-language prediction finding.
- openai/whisper-large-v3 model card — five-million-hour training mixture, 128 mel bins, Cantonese token, the 10–20% error reduction, and the hallucination / repetition / uneven-across-languages caveats.
- Open ASR Leaderboard — the 6.43% English WER and RTFx figures for large-v3 (8 October 2025 snapshot).
See it on your own audio.
Paste a link and read the result: YouTube · podcasts · audio URLs. Free, 15 a day, no signup.