Whisper Large v3
OpenAI · released Nov 7, 2023 · openai/whisper-large-v3
Speech to textTranscribes a recording into words
- Type
- Open weightsApache License 2.0
- Languages
- 99
- Size
- 1.5B
Context measured in tokens
Our take
Written Aug 4, 2026Whisper Large v3 turns recorded speech into written words, and is still the name most people reach for. It handles 99 languages and is small enough to run on a laptop, though newer models now beat it on accuracy.
Reach for it when the recording is clean and the language coverage matters — it gets about one word in sixty wrong on read-aloud audio, and handles far more languages than most alternatives. It is also small enough to run on your own machine rather than paying by the minute. Skip it if you are transcribing meetings or heavily accented speech, where its error rate rises more than eightfold.
The case for it
- 99 languages on one model, which very few transcription models match.
- Excellent on clean read-aloud audio, at about a quarter of its own average error rate.
- Runs on consumer hardware: it fits in memory on almost every device tracked.
The case against it
- Meeting audio is its weak spot, with more than eight times the error rate it manages on clean speech.
- Mid-table on headline accuracy now, at 54th of 74 on the Open ASR Leaderboard.
- Every accuracy figure here is English; nothing we hold measures the rest.
How good is it?
An open transcription model for turning speech into text, though it struggles more than most with recordings of meetings in a room.
- transcribing recordings of meetings in a roomRecorded meetings · 72nd of 92
TranscriptionTurning speech into text2.5 of 5Open ASR WER · 54th of 76
94.2%
Misses roughly one word in 17, averaged over nine English test sets.
470×50th of 74
an hour of audio in 8 seconds, on the board's own hardware. Your machine will differ.
99
Listed on the model card. The accuracy above is English only.
Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case; the placing beneath each rate is against every model measured on that set.
Which languages ↓ ↑
English · Chinese · German · Spanish · Russian · Korean · French · Japanese · Portuguese · Turkish · Polish · Catalan · Dutch · Arabic · Swedish · Italian · Indonesian · Hindi · Finnish · Vietnamese · Hebrew · Ukrainian · Greek · Malay · Czech · Romanian · Danish · Hungarian · Tamil · Norwegian · Thai · Urdu · Croatian · Bulgarian · Lithuanian · Latin · Māori · Malayalam · Welsh · Slovak · Telugu · Persian · Latvian · Bangla · Serbian · Azerbaijani · Slovenian · Kannada · Estonian · Macedonian · Breton · Basque · Icelandic · Armenian · Nepali · Mongolian · Bosnian · Kazakh · Albanian · Swahili · Galician · Marathi · Punjabi · Sinhala · Khmer · Shona · Yoruba · Somali · Afrikaans · Occitan · Georgian · Belarusian · Tajik · Sindhi · Gujarati · Amharic · Yiddish · Lao · Uzbek · Faroese · Haitian Creole · Pashto · Turkmen · Norwegian Nynorsk · Maltese · Sanskrit · Luxembourgish · Burmese · bo · Filipino · Malagasy · Assamese · Tatar · Hawaiian · Lingala · Hausa · ba · Javanese · Sundanese
The figures above come from the Open ASR Leaderboard, an independent public test that runs every model on the same recordings. It is the only measurement of transcription quality we know of, so there are no other scores to show.
Each of these is the same transcription job on a different kind of recording, so together they say where it holds up and where it slips — not how closely it follows an instruction.
Every published score for this model9 scoresEvery figure we hold, from 9 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 20.6 GB spare means a 10% error in the size would not change the answer.
Apple M1 Pro (16-core GPU) · 32 GB
Room to spare. 21.8 GB spare means a 10% error in the size would not change the answer.
Apple M2 (8-core GPU, 8GB unified) · 8 GB
Room to spare. 3.8 GB spare means a 10% error in the size would not change the answer.
These cards answer whether Whisper Large v3 loads, not how fast it transcribes. Throughput figures for a transcription model come from its text decoder, so treat this as a fit answer rather than a speed one.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first. This is a fit answer: whether it loads, not how fast it transcribes.
Check against your own machine → · Where to rent it hosted →
Or rent it from someone else
Prices checked 4 hours ago — each listing carries its own date.
- per minute of audio
- $0.002
- Context served
- —
- Throughput
- Not measured
| Provider | Price per minute of audio | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| GroqDirect | $0.002checked 4 hours ago | not reported | not measured | Unknown | Unknown | Unknown |
| DeepgramDirect | $0.006checked 4 hours ago | not reported | not measured | Unknown | Unknown | Unknown |
DeepgramAcross the 2 listings we hold: 0 say they do not train on prompts, 0 say they do and 2 do not say. 0 appear in the zero-retention registry we check; the rest are unknown to us.
When we formed this view
Recent changes
What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- 2 of 2 listings publish no parameter list, so what their API accepts is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 2 of 2 listings do not say whether they train on prompts.
- We hold no cached-input rate for any of its listings.
- We hold no batch or off-peak rate for any of its listings.
Licence and identifiers
What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- openai/whisper-large-v3
- Architecture
- Dense
- Takes in, gives back
- Audio in, text out
- Catalogue slug
- openai-whisper-large-v3