Is Whisper Free for Commercial Use? (2026)
Yes: OpenAI's Whisper code and weights are released under the MIT licence. What that allows, where the licence labels differ, and what it does not cover.
The Scriptorium team ·
Yes. OpenAI's Whisper README says "Whisper's code and model weights are released under the MIT License", and the MIT licence permits commercial use if you keep the copyright and permission notice with the software. Running Whisper yourself is free of licence fees; what it costs you is computing time. OpenAI's paid speech-to-text API is a separate service with its own terms. This page is a reading of the public licence texts, not legal advice.
Which licence applies to which piece?
| Piece | Licence shown | Where we read it |
|---|---|---|
| OpenAI Whisper (code and weights) | MIT; copyright line "Copyright (c) 2022 OpenAI" | GitHub README and LICENSE file |
| openai/whisper-small on Hugging Face | apache-2.0 (in the page's metadata) | Model card |
| openai/whisper-large-v3 on Hugging Face | apache-2.0 (in the page's metadata) | Model card |
| faster-whisper (a faster reimplementation) | MIT | GitHub repository |
| Systran/faster-whisper-small (converted weights) | MIT (in the page's metadata) | Model card |
Notice that the GitHub README says MIT while OpenAI's own Hugging Face model cards list Apache-2.0. Both are permissive licences that allow commercial use, but they are different documents with different conditions, so read the licence that is attached to the exact files you download. We did not find a statement on those pages that explains the difference.
What the MIT licence lets you do
The MIT licence is short. From the LICENSE file in the Whisper repository, its main condition is that the copyright notice and permission notice must be included in all copies or substantial portions of the software. In practice, if you ship Whisper inside a product or a service, you keep that notice, usually in a licences file or an about page. It does not require you to open-source your own code.
What "free" does not cover
- The recordings. The licence is about the software and the weights. Whether you may record, store and transcribe a person's voice is a question for privacy and consent law where you operate. The model card itself cautions against using Whisper to transcribe recordings of people taken without their consent.
- Accuracy. The same model card says Whisper can produce text that was not spoken (hallucination), performs unevenly across languages, and varies by accent and dialect. A permissive licence says nothing about whether a transcript is correct.
- Training data. The model card says the models were trained on about 680,000 hours of audio collected from the internet. The licence text we read says nothing about that data; if it matters to your legal review, ask a lawyer.
- OpenAI's hosted API. It is a paid service governed by OpenAI's terms. Its speech-to-text guide lists a 25 MB file limit and names newer models alongside whisper-1; check the guide for the current status.
- Compute. The weights are free; a processor or graphics card to run them is not. On a CPU, expect transcription to take about as long as the recording or longer (see how long an hour of audio takes).
How this site uses Whisper
TryScriptorium runs Whisper-family models through faster-whisper on its own server; the model names are tiny, base, small, medium, large-v3-turbo and large-v3, and the free no-account preview always uses small. Using the website does not require you to take a licence from anyone. If you would rather run it yourself, the local Whisper guide explains the options, and self-hosted Scriptorium wraps the same models in a web interface.
A short checklist before you ship it in a product
- Open the licence attached to the exact weights you download (GitHub and Hugging Face may differ) and copy the notice into your product's licences file.
- Decide where the audio is processed (your server, the user's device or a vendor's) and say so in your privacy notice.
- Get consent for recordings of people, and keep a human check on anything that will be published or relied on.
- Test on your own audio and language before you promise anything about accuracy.
Checked on 2026-10-26. OpenAI Whisper GitHub README (licence sentence) and LICENSE file (MIT, "Copyright (c) 2022 OpenAI", notice condition); Hugging Face cards for openai/whisper-small and openai/whisper-large-v3 (apache-2.0 metadata, training-data and caution wording) and Systran/faster-whisper-small (MIT metadata); the faster-whisper repository (MIT); OpenAI's speech-to-text guide (25 MB limit). We are not lawyers and this is not legal advice. Not read: OpenAI's API terms of use. Menus, prices and limits change; confirm on the vendor's page before you rely on them.