How can you use OpenAI Whisper online?
Three ways, and none of them needs an install: Whisper Web, OpenAI's demo on Hugging Face, or OpenAI's API. Each has a catch. Whisper Web says it runs on your own computer, the free demo runs on shared hardware, and the API is a developer product you pay for.
Whisper itself is free. On our test Mac, its default model, turbo, made 10 errors in about 1,100 words of a clean audiobook clip, the fewest of the six models. To get that, we needed an 805 MB Python setup and a 1.6 GB download, and the run took 87 seconds for under 6 minutes of audio.
What did it take to run Whisper on a Mac?
We installed OpenAI's official openai-whisper package on an Apple M5 Pro with 64 GB of memory. Then we ran each of its six models on the same recording: 5 minutes 52 seconds of someone reading aloud, from the LibriSpeech test set. Everything ran on the processor, which is what the package does on a Mac out of the box, with the language set to English.
Before it can transcribe anything, you need ffmpeg and Python. We used Python 3.12, and the environment took 805 MB on disk. Then you download a model: 76 MB for tiny, 1.6 GB for turbo and 3.1 GB for large-v3. Our guide to downloading and installing Whisper has the steps for Windows and Mac.
Seconds to transcribe the clip
Turbo took 87 seconds. At that pace, an hour of audio would take about 15 minutes. Large-v3 would take about 43 minutes, and tiny under a minute.
Which Whisper model should you run?
For English, run turbo. It's already the default, and it made the fewest errors in our test: a 0.9% word error rate, against 1.35% for medium and 1.8% for small.
Word error rate (lower is better)
The biggest model did worst here. Large-v3 was also about 3 times slower than turbo. On its second run it turned the last half minute of the clip into garbled text in several languages ("However your 오늘속 Übertragung derives fromורה"). Nobody on the recording said anything like that. It's one clean clip, so treat the chart as a guide, not a ranking.
To translate into English, use medium instead. Whisper's README says turbo "is not trained for translation tasks" and returns the original language even when you ask it to translate. Here's what we got from Whisper's translate task. On our test Mac, turbo and medium each needed about 6.5 GB of memory.
What are the ways to use Whisper online without installing it?
| Route | Where it runs | Account | File limit | Cost |
|---|---|---|---|---|
| OpenAI's API | OpenAI's servers | OpenAI developer account | 25 MB per file | Billed per use |
| OpenAI's Hugging Face demo | Hugging Face's shared graphics chips, on large-v3 | Not stated | Not stated | Free |
| Whisper Web | On your device, the site says | None | Your device's power and memory | Free |
The API's 25 MB cap means a bigger recording has to be compressed or split before you send it. If you're choosing between Whisper Web and an upload app, WhisperAI vs Whisper Web compares them. If you'd rather run Whisper offline in an app, see the Whisper desktop apps.
What if you only want the transcript?
Then you don't need a model at all. WhisperAI is an independent transcription app that works in a browser tab. Upload a recording up to 5 GB and 10 hours long, turn on speaker labels if more than one person talks, and export the transcript as DOCX, PDF, TXT or SRT. Nothing runs on your computer, so there's no setup and no model to pick. An hour of audio usually comes back in under a minute, depending on the language and the number of speakers. The Free plan covers 5 minutes a month, and Business Pro is unlimited for $24.99 a month.
Frequently asked questions
Is there a free version of Whisper?
Yes. Whisper's code and model weights are free under the MIT License, so running it on your own computer costs nothing but the setup. Of the three online routes on this page, only OpenAI's API charges.
Can you still use Whisper?
Yes. The open model on GitHub keeps working. What's ending is whisper-1, the Whisper model in OpenAI's API. OpenAI told developers on August 26, 2026 that it will remove it on February 26, 2027, and names gpt-transcribe and gpt-live-transcribe as the replacements.
How many languages does Whisper support?
Whisper's tokenizer lists 100 language codes (the large-v3 model added Cantonese), and OpenAI says the model was trained on 98 languages. Accuracy varies by language. WhisperAI transcribes 100+ languages and regional variants using Whisper's codes.
Is WhisperAI made by OpenAI?
No. WhisperAI is an independent product. It is not affiliated with, owned by, operated by, or endorsed by OpenAI.
Open a browser tab and transcribe.
Upload a file and read the transcript in the editor, with nothing to install.
See plans and pricing.