Which Whisper Model Should You Use?

For English on a laptop or desktop, use turbo, the model Whisper loads when you don't name one. We ran all ten current Whisper models on the same recording on a Mac, and turbo got the fewest words wrong: 10 out of 1,108. large-v3, the biggest, was the slowest and the least accurate, and on one run it invented half a minute of text the speaker never said. To translate speech into English, use medium or large-v3.

Which Whisper model should I use?

What are the different Whisper models?

OpenAI's Whisper comes in six sizes, and four of them have an English-only version marked .en. Parameters count how big the model is, in millions.

turbo is the newest model, and the Whisper package hasn't had a release since June 2025. It's large-v3 with the decoder cut from 32 layers to 4, then fine-tuned for two more epochs. Ask for "large" and you get large-v3; the older large-v1 and large-v2 still load by name.

How fast and accurate is each Whisper model?

We ran each model on an Apple M5 Pro with 64 GB of memory, on the CPU with default settings and the language set to English. The recording was 5 minutes 52 seconds of clean English read aloud, from the LibriSpeech test set: 1,108 words, one speaker. Each model's error count comes from one run, and large-v3 ran twice.

tiny and base finished the recording in under 10 seconds, small in under 30, medium in about 55, turbo in 87 and large-v3 in 252. Scaled up to an hour of audio, that's about a minute for base, 15 minutes for turbo and 43 for large-v3.

OpenAI's own speed column says turbo is about 8 times faster than large and tiny about 10 times faster, measured on an A100, a data-center GPU. On our Mac's CPU, turbo was about 3 times faster than large-v3 and tiny more than 50 times.

On an 8 GB computer, medium and turbo leave little room for anything else. small and below fit easily.

Do the English-only .en models make fewer mistakes?

For the two smallest sizes, yes. tiny.en got 3.6% of words wrong against tiny's 5.6%, and base.en 2.7% against base's 3.6%. From small up the gap closes: small.en did a little worse than small, and medium.en a little better than medium. That matches OpenAI's note that the difference "becomes less significant for the small.en and medium.en models". None of them beat turbo, which has no English-only version.

Is large-v3 the most accurate Whisper model?

Not for English on our recording. OpenAI reports 10% to 20% fewer errors than large-v2 across a wide range of languages. On our clean English clip, large-v3 was the slowest model and the least accurate: 11.3% of words wrong on the first run, 9.1% on the second.

On the second run it wrote about 30 seconds of mixed-language nonsense while the speaker was reading plain English, and its memory jumped to 12.6 GB. The transcript says "capacity variation in 1974 for the following jargon" and "However your 오늘속 Übertragung derives fromורה". The speaker said neither. Whisper's model card warns that output "may include texts that are not actually spoken". A 2024 study of an earlier Whisper found made-up phrases in roughly 1% of transcripts.

One recording can't rank the models, but it's enough to stop assuming the biggest one is the safe pick for English.

Which Whisper model should I use for translation or other languages?

To turn speech in another language into English text, use medium or large, not turbo. OpenAI's README says turbo isn't trained for that, and asking it to translate gets you the original language back.

To transcribe other languages as spoken, turbo "performs similarly to large-v2", OpenAI says, but slips on a few, Thai and Cantonese among them. OpenAI's docs list 98 languages for Whisper, and accuracy varies by language.

Are faster-whisper and Distil-Whisper different models?

faster-whisper isn't a new model. It runs OpenAI's weights on a different engine, and its makers say it's up to 4 times faster for the same accuracy. Distil-Whisper is a smaller model trained from Whisper. Its makers say it's 6 times faster and within 1% word error rate on out-of-distribution test sets, and the distil-large-v3 version is for English only.

Is Whisper free, and is OpenAI's Whisper API still available?

The models are free. OpenAI released the code and the trained model files under the MIT License, so running them costs your hardware, your time and some setup: our Python environment took 805 MB before the first model downloaded, and ffmpeg is a separate install.

The API is still there, for now. OpenAI's whisper-1 model is deprecated and goes away on February 26, 2027, and OpenAI now points developers to gpt-transcribe. The API also caps files at 25 MB.

Can I skip choosing a model?

Yes, with WhisperAI. It's an independent transcription app, not made by OpenAI, and it doesn't run Whisper, so there's no model to pick, download or make room for. Upload audio or video in your browser, up to 5 GB and 10 hours per file on every plan, and you get a transcript in any of 98 languages. Speaker labels are on every plan, and you can export as DOCX, PDF or SRT.

The free plan gives you 5 minutes a month with no card, and Business Pro is $24.99 a month for unlimited minutes. If you'd rather run Whisper itself without installing anything, here are the ways to use Whisper online.