Whisper large-v3-turbo is large-v3 with most of its decoder removed: OpenAI cut the decoder from 32 layers to 4, then fine-tuned it again. That takes it from 1550M to 809M parameters, and OpenAI's README lists it at about 8 times the speed of large-v3, "with a minimal degradation in accuracy." The catch is translation. Turbo was not trained to translate speech into English, so use large-v3 (or medium) if you need that. For plain transcription, try turbo first and switch to large-v3 if your language is one where turbo is weaker.
We have not run a controlled benchmark. Every figure below comes from OpenAI, Hugging Face or the Open ASR Leaderboard, and describes their test audio, not your voice, accent, microphone or vocabulary.
What is Whisper large-v3-turbo?
The whisper-large-v3-turbo model card says turbo "is a finetuned version of a pruned Whisper large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation."
OpenAI's release discussion (1 October 2024) adds three details:
- Turbo has "only 4 decoder layers—just like the tiny model—down from the 32 in the large series."
- It was fine-tuned "for two more epochs over the same amount of multilingual transcription data used for training large-v3, i.e. excluding translation data."
- "Across languages, the turbo model performs similarly to large-v2, though it shows larger degradation on some languages like Thai and Cantonese."
The encoder, the part that listens to the audio, is unchanged. The decoder, the part that writes out the text one token at a time, is what shrank. That is why turbo is faster: the decoder runs once per token, so a smaller one saves time on every word.
Large-v3 vs turbo side by side
| large-v3 | large-v3-turbo | |
|---|---|---|
| Parameters | 1550M | 809M |
| Decoder layers | 32 | 4 |
| Required VRAM, OpenAI's README | ~10 GB | ~6 GB |
| Relative speed, OpenAI's README (A100 GPU) | 1x | ~8x |
| Translate speech into English | Yes | Not trained for it |
| Languages | Multilingual, 99 language codes on the card | Same language list |
| English-only version | No | No |
| Licence | MIT per OpenAI's README; the Hugging Face card's tag reads apache-2.0 | MIT per OpenAI's README and the Hugging Face tag |
On licences: the openai/whisper README says "Whisper's code and model weights are released under the MIT License." The large-v3 card on Hugging Face carries an apache-2.0 tag, while the turbo card carries MIT. Both are permissive licences, but if it matters to your organisation, read the LICENSE file and both cards yourself.
How do they compare on the Open ASR Leaderboard?
The Hugging Face Open ASR Leaderboard tests models on public English datasets and reports word error rate (WER, lower is better) and RTFx (a speed measure, higher is better). These figures are from the leaderboard's English short-form results file, version dated 25 September 2026:
| Dataset on the leaderboard | large-v3 WER | large-v3-turbo WER |
|---|---|---|
| LibriSpeech test-clean (read audiobooks) | 1.56 | 2.13 |
| LibriSpeech test-other (noisier audiobooks) | 3.52 | 3.71 |
| AMI, cleaned (meetings) | 13.63 | 13.88 |
| VoxPopuli, cleaned (European Parliament) | 4.53 | 7.02 |
| Voice Arena Monsoon (Indian-accented conversation) | 3.95 | 4.77 |
| RTFx (speed, higher is faster) | 470.2 | 797.03 |
Three things to notice. Turbo is slightly worse on every dataset here, by a little on most and more on VoxPopuli. The gap between datasets (under 2 on audiobooks, over 13 on meetings) is far larger than the gap between the two models. And the leaderboard's own text says it normalises text before scoring, removing punctuation and casing, so these numbers say nothing about how well commas and capitals come out. The table can change as the leaderboard updates; check the live page.
What changed from large-v2 to large-v3?
The large-v3 card says large-v3 "has the same architecture as the previous large and large-v2 models, except for the following minor differences":
- "The spectrogram input uses 128 Mel frequency bins instead of 80."
- "A new language token for Cantonese."
The training data changed more. Large-v3 "was trained on 1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio collected using Whisper large-v2." The card reports "10% to 20% reduction of errors compared to Whisper large-v2" across "a wide variety of languages."
Large-v2, per its card, was "trained for 2.5x more epochs with added regularization" compared with the original large, on the original 680,000 hours.
So which is better? On OpenAI's own figures, large-v3 is the newer and more accurate model. Some people keep large-v2 because a pipeline or fine-tune was built on it. WhisperX's README, for instance, still uses large-v2 in its examples. If you are starting fresh, there is no published reason in OpenAI's pages to prefer v2.
When does each model make sense?
| Your situation | Start with | Why |
|---|---|---|
| Transcribing long English files on a Mac | turbo | Much faster per OpenAI's README; small accuracy cost on the leaderboard datasets |
| Translating speech into English | large-v3 or medium | OpenAI's README: "the turbo model is not trained for translation tasks" and "will return the original language even if --task translate is specified" |
| Thai or Cantonese | large-v3 | OpenAI's release note names these as languages where turbo degrades more |
| Limited memory | turbo, or a smaller size | ~6 GB vs ~10 GB in OpenAI's table; see Whisper model sizes for smaller options |
| An existing large-v2 pipeline | Test large-v3 against it | Card reports fewer errors, but your audio decides |
On Apple Silicon, both run through Mac-focused projects. whisper.cpp's README lists large-v3 and large-v3-turbo among its downloadable models. Argmax's WhisperKit README recommends a compressed Large v3 Turbo (large-v3-v20240930_626MB) "across iOS and macOS for maximum accuracy" and the uncompressed turbo "on macOS for maximum speed and accuracy." Run Whisper on a Mac covers setup.
How to choose on your own audio in 15 minutes
- Record two minutes of yourself reading a made-up passage: "Priya Raman moved the 9 June review to 4:15 pm. The fee is 3,480 dollars, not 3,840. Do not send the Halvorsen file."
- Add one minute of the kind of audio you really transcribe: a call, a lecture, a voice memo.
- Run both models on the same file, for example
whisper test.m4a --model large-v3and then--model turbowith OpenAI's command-line tool. - Time each run and count wrong names, wrong numbers and dropped "not"s.
- If turbo's errors do not matter for your use, keep the faster model.
Questions people ask
Is Whisper turbo as accurate as large-v3?
Close, but not quite, on OpenAI's and the leaderboard's figures. OpenAI calls the loss "minimal" and says turbo performs similarly to large-v2, with larger drops on some languages such as Thai and Cantonese.
Can Whisper turbo translate?
Not reliably. OpenAI's README says turbo "is not trained for translation tasks" and returns the original language even when you ask it to translate. Use medium or large for translation.
Is large-v3 better than large-v2?
OpenAI's large-v3 card reports a 10% to 20% reduction in errors compared with large-v2 across many languages. Test on your own audio if you have a working v2 setup.
Does SuprFlow use large-v3 or turbo?
No. SuprFlow 1.0.0 uses a WhisperKit (Whisper) model with Small and Medium options, in multilingual and English-only variants.
Related guides
- Whisper model sizes compared
- How to run Whisper on a Mac
- Whisper vs faster-whisper vs WhisperX
- Speech to text accuracy and word error rate
- Parakeet vs Whisper
Sources
- OpenAI — whisper GitHub repository and README
- OpenAI — large-v3-turbo release discussion #2363
- OpenAI — whisper-large-v3-turbo model card on Hugging Face
- OpenAI — whisper-large-v3 model card on Hugging Face
- OpenAI — whisper-large-v2 model card on Hugging Face
- Hugging Face — Open ASR Leaderboard
- English results dataset
- ggml-org — whisper.cpp GitHub repository and README
- Argmax — argmax-oss-swift (WhisperKit) GitHub repository and README
- m-bain — WhisperX GitHub repository and README