Whisper Large-v3 vs Turbo

Turbo is Large-v3 with its decoder cut from 32 layers to 4. OpenAI says it is about 8 times faster and needs about 6 GB of GPU memory instead of 10, with accuracy close to the older Large-v2. It can't translate, and it slips further on a few languages.

By VexaScribe Editorial · Published August 23, 2026 · Rewritten November 17, 2026 · Figures from OpenAI

Short answer

  • Use Turbo to transcribe, especially on a smaller GPU or when speed matters. It's the default in openai-whisper.
  • Use Large-v3 to translate speech into English, for Thai or Cantonese, or when a test on your own audio shows Turbo making more mistakes.

Differences

Large-v3Turbo (large-v3-turbo)
ReleasedNovember 6, 2023October 1, 2024
Parameters1,550M809M
Decoder layers324
GPU memory~10 GB~6 GB
Speed (relative)1×~8×
Translate to EnglishYesNo
Name in openai-whisperlarge-v3 (large)turbo (the default)
LicenceMITMIT

Memory and speed are from the openai/whisper README, measured on English speech on an A100; your speed will vary with the language, the hardware and how fast people talk. The encoder is the same in both, which is why Turbo isn't 8 times smaller: most of the saving comes from the decoder, which runs once per output token.

Accuracy

OpenAI hasn't published a single headline WER for Turbo. What its release notes do say:

  • Large-v3 made 10% to 20% fewer errors than Large-v2 across a wide range of languages
  • Turbo performs similarly to Large-v2 across languages
  • Turbo shows larger degradation on some languages, such as Thai and Cantonese
  • Turbo does better on clean recordings (FLEURS) than on everyday ones (Common Voice)

So expect Large-v3 to be somewhat more accurate, with the gap depending on the language and the recording. Our own Whisper benchmark averaged three deployments, Turbo among them, and doesn't publish per-model results, so it can't settle this. The reliable test is ten minutes of your own audio through both.

Translation

Turbo was fine-tuned on transcription data only, without translation data. Ask it to translate with --task translate and it returns the original language. To translate speech into English, use Large-v3, or medium on a smaller GPU.

How to choose

  • Translating into English? Large-v3.
  • Thai, Cantonese or another language where tests show a gap? Large-v3.
  • GPU with less than about 10 GB? Turbo.
  • Lots of audio, or you need results fast? Turbo.
  • Otherwise, run both on a sample and compare.

Running each

Command line (openai-whisper)
pip install -U openai-whisper
whisper interview.mp3 --model turbo
whisper interview.mp3 --model large-v3
whisper japanese.mp3 --model large-v3 --task translate
Python
import whisper

model = whisper.load_model("turbo")  # or "large-v3"
print(model.transcribe("interview.mp3")["text"])

Full setup, including FFmpeg and a GPU, is in our install guide.

Don't want to run it yourself?

VexaScribe's Standard tier runs OpenAI Whisper and adds speaker labels, with nothing to install. Upload audio or video up to 5 GB, then export TXT, Word, SRT, VTT or JSON.

30 free minutes on signup, no card. Plans from $2 a month.

A VexaScribe transcript with a timestamp and the speaker, Sam or Rosa, on every line

Sources

Checked November 17, 2026. An earlier version cited a VexaScribe benchmark that was withdrawn on August 25, 2026. VexaScribe is independent and not affiliated with OpenAI. See our editorial standards.

Frequently Asked Questions

Is Whisper Turbo as accurate as Large-v3?

Close, but a little less. OpenAI says Turbo performs similarly to Large-v2 across languages, while Large-v3 made 10% to 20% fewer errors than Large-v2. The gap is larger on some languages, such as Thai and Cantonese, so test both on your own audio.

How much faster is Turbo?

About 8 times faster than the large models, according to the openai/whisper README, measured on English speech on an A100. The speed-up comes from cutting the decoder from 32 layers to 4.

Can Whisper Turbo translate to English?

No. Turbo was fine-tuned without translation data and returns the original language even with --task translate. Use Large-v3, or medium, to translate speech into English.

How much GPU memory do they need?

About 6 GB for Turbo and about 10 GB for Large-v3, according to the openai/whisper README. Turbo is the practical choice on an 8 GB card.

Which one does VexaScribe use?

Our Standard tier runs OpenAI Whisper through a hosting provider. We can't confirm which checkpoint it uses at any given time, so we don't name one.