Can Whisper AI run offline?

Yes. The open-source Whisper model runs offline once its model file is on your computer, and nothing you say has to reach a server. Two steps need a connection: installing the software and downloading the model the first time. After that, transcription works with the network switched off. The hosted OpenAI API is the exception, because it is a cloud service and cannot run offline.

The model runs offline because OpenAI released it as open-source software

OpenAI released Whisper's code and model weights under the MIT License ¹. The license lets anyone download the model, run it on their own hardware and transcribe without sending anything to OpenAI. That is what makes an offline setup possible. The model is a file on your disk, and the recognition runs on your processor or graphics card.

The whisper.cpp project also shows the model running on an iPhone 13 fully offline, on-device ². That demonstration shows it can run offline. It does not tell you how fast the model runs on your machine or how accurate it is, so test your own setup.

The first run needs a connection, and the model file then stays on disk

The openai-whisper package fetches a model the first time you load it by name. Its loader downloads the file over the network when the file is not already present ³. By default the file is saved in a folder in your home directory, ~/.cache/whisper ⁴, unless you choose another folder.

So an offline setup follows three steps:

1. Install the software while you are online.
2. Load the model once while you are online, which downloads and saves it.
3. Switch off the network and transcribe a file.

Once the model is in the cache, the loader finds it and skips the download. That is why the test in step 3 is the proof that the setup works offline, not the setup steps themselves.

Offline depends on the app you use, not only on the model

Many apps are named after Whisper, and each one decides where the model runs. WhisperType downloads its models from Hugging Face when you choose Download, but it can instead use a model file you import, and the import route works offline ⁵. An app like that is offline after setup, but only if you use the import route or let the download finish first.

Whisper To Input, an Android keyboard, offers two backends: the OpenAI API with your own key, or a Whisper server you host yourself ⁶. Neither works offline on the phone. The self-hosted server keeps audio on infrastructure you run, but the phone still needs a connection to reach it. Whether your words stay private and whether the app works without a connection are separate questions.

Hardware is the other limit. The command-line package uses an NVIDIA graphics card only when PyTorch can detect one, and otherwise falls back to the processor ⁷. A processor-only run works offline, but it can be slower. Our question on whether Whisper runs without a GPU covers the hardware choice in more detail.

Offline is the right choice when audio is sensitive or the network is unreliable

Choose an offline setup when the recordings are confidential, when you work where connections are poor, or when a cloud service would break your rules. The guide to whether Whisper AI is safe explains where each setup sends its data. For the wider choice, see our guide to open-source dictation software and whether Apple dictation works offline.

What to do: test the setup with the network switched off

1. Choose a local option: the open-source model, or an app whose local mode you have tested.
2. Load the model while you are online, so the file is saved to disk.
3. Turn on airplane mode, then transcribe a real recording of a few minutes.
4. If the text appears, the recognition ran on the device. If the app stops or asks for a connection, its offline mode is not what the app claims.
5. Repeat the test after each app update, since a later version can add a cloud feature.

If you would rather not build a local setup, look for an app that documents an offline mode, and run the same test on it. Our dictation page for Mac covers the offline options on that platform. For the privacy side on every platform, see the overview of speech to text privacy.