MMOOZZGG.com

Glossary

What is OpenAI Whisper?

In short

OpenAI Whisper is a model for speech recognition, speech translation, and language detection; its code and weights are open under the MIT license.

Models and products · Whisper

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine an attentive classmate who writes down every word the teacher says, even when the class is noisy and the teacher speaks fast. Then the classmate gives you neat notes. Sometimes they hear the wrong word, especially in names and rare terms. So you check the important parts of such notes against the textbook.

In plain words

It is a program that turns speech into text. You give it an audio recording, and it returns a transcript; it can also detect the language and translate what was said. The model learned from a huge amount of recordings from the internet, so it handles noise, accents, and different topics fairly well. You can download it for free and run it on your own computer or server, or call the cloud version through the OpenAI API. Example: a sales team transcribes call recordings to find customer objections in them.

For business

The main benefit for a business in Russia is that the code and weights of Whisper are open under the MIT license: you deploy the model on your own server, and call recordings never leave the company. This matters for calls with personal data. The cloud whisper-1 from the OpenAI API is not officially available in Russia: the country is not on the list of supported countries. Also, on 26 August 2026, OpenAI announced that it will retire whisper-1 from the API on 26 February 2027. The cost of your own deployment depends on the volume of recordings, the speed you need, and the model size you choose. Test how transcription connects to your phone system and CRM on one stream of calls.

For a specialist

The model is described in the paper by Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision” (2022). It was trained on 680,000 hours of multilingual, multitask labeled data collected from the internet (weak supervision), and it transfers to new datasets without fine-tuning (zero-shot transfer). Design: an encoder-decoder transformer; one model handles recognition, speech translation, language detection, and voice activity detection. Sizes in the repository: tiny with 39 million parameters, base with 74 million, small with 244 million, medium with 769 million, large with 1,550 million, and turbo with 809 million. Turbo is an optimized large-v3: it is faster with a small loss of accuracy, but it is not trained for translation. In the API, the model is available as whisper-1 with the transcriptions and translations methods; the deprecations table names gpt-transcribe or gpt-live-transcribe as the replacement.

Example

A clinic installs the model on its own server and transcribes doctors’ audio notes after appointments. Recordings with patient data do not go to an outside service, and the doctor edits the text before adding it to the record.

Related terms

Questions and answers

Can Whisper run without the internet? Yes. The code and weights of the model are open under the MIT license, so you install it on your own computer or server.

The repository has six sizes, from tiny with 39 million parameters to large with 1,550 million.

How is open Whisper different from whisper-1 in the API? You run the open model on your own machine, and the recordings go nowhere. The whisper-1 model is the cloud version in the OpenAI API, and the company will retire it on 26 February 2027.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us