Does Whisper AI have an API?

Yes. You can send an audio file to OpenAI's transcription endpoint and get a text transcript back, but the hosted service is paid by the minute and requires an API key. The open-source Whisper model needs no API at all. You can download it and run it on your own computer for free. Which of the two you need depends on whether you are building a product or simply transcribing recordings.

OpenAI's API transcribes a recorded file and returns text

OpenAI's speech-to-text guide documents the transcription endpoint. You send a file and a model name, and the service returns the transcript. The API offers more than one model, and the names differ. The guide names whisper-1 as the Whisper model, and its own quickstart example uses gpt-transcribe, a newer model.

The same API handles translation. The guide says that to translate a completed recording into English you use the translations endpoint with whisper-1 ¹. Translation is a developer feature. It is not the same as the dictation button in ChatGPT, so do not assume the two behave alike.

Streaming is the main limit. The guide says that file streaming works with the gpt-4o transcription models, and that whisper-1 does not support it ². If you need text while a recording is still arriving, choose one of the newer models or use a different tool. The guide to real-time speech recognition covers that case.

You need a secret API key, and it must stay private

To call the API, you send your key in an HTTP header. The guide's example uses the form Authorization: Bearer followed by the key, which is read from an environment variable named OPENAI_API_KEY ³. Store the key in that variable or in a secrets manager, and keep it out of code you share or publish. Anyone who holds the key can make requests that are billed to your account.

OpenAI says that data sent to its API is not used to train its models unless you opt in ⁴. That is a statement about training, not a promise about every log, so read the current data controls before you send sensitive audio.

The hosted API bills by the minute and caps each file at 25 MB

OpenAI's price list shows gpt-transcribe at $0.0045 per minute, and gpt-4o-transcribe at $0.006 per minute ⁵. Those figures are for one date and exclude any editing. Multiply your audio minutes by the rate to estimate the cost, and check the price list before you budget, because the model you call decides the rate.

Each request also has a size limit. The Transcriptions API accepts files up to 25 MB. For larger recordings, the guide says to use a compressed audio format or split the file into chunks of 25 MB or less ⁶. For the accepted formats and how to fix an upload that fails, read which audio formats and file sizes Whisper accepts.

It depends: the open-source model is free, and many apps call the API for you

The API is one of three ways to use Whisper. The other two are the free open-source model and an app built on top of it.

• You transcribe recordings and want no bill. Run the open-source model yourself. The Whisper README says the Python package requires the ffmpeg command-line tool to be installed on your system ⁷. The guide to running Whisper locally has the setup steps.
• You want dictation without setup. Some Whisper apps send audio to OpenAI with your own key. One keyboard app, for example, uses either the OpenAI API with your key or a Whisper server you host yourself ⁸. Check each app's privacy policy, because the audio follows the app's route, not the API's.
• You want a one-off transcript. A free online audio-to-text tool runs in your browser, so you do not need a key.

Whether you need the API at all is also a question of cost. Is Whisper AI free? answers that for each path, and the price of OpenAI's speech-to-text API compares the rates.

What to do: choose a path, then test one short file

1. Decide between the free model and the API. If you only need transcripts and you have a suitable computer, the open-source model has no per-minute fee. If you are building a product or lack the hardware, the API is the simpler route.
2. For the API, create a key in your OpenAI account and store it as OPENAI_API_KEY. Do not paste the key into a document, a chat or a public repository.
3. Send a short test file first. Use a file under 25 MB and the model name from the price list. The guide's curl example posts the file to /v1/audio/transcriptions with the Bearer header.
4. Check the transcript and the bill. Compare the output against the audio for names and numbers, and confirm the per-minute charge matches your estimate before you send a long batch.

The overview of speech recognition places the API among the other options, and the guide to what Whisper AI is explains the model behind it. For the underlying term, see automatic speech recognition.