In plain words
Whisper is an AI model, not an app. It does one job: speech recognition. You give it a recording, such as a voice note, a lecture or a podcast, and it gives back the words as text.
OpenAI released it in September 2022 and gave away the model itself. Anyone can download it and run it on their own computer for free, so developers can build it into their own apps for subtitles, meeting notes or transcripts.
It learned from 680,000 hours of audio collected from the web, paired with text, in many languages. Hearing that much messy real-world speech is what makes it fairly good with accents, background noise and technical words. It can also listen to another language and write the meaning in English.
Everyday example
A stenographer in the classroom
Picture a stenographer sitting at the back of a lecture. They listen to half a minute at a time, write it down neatly with punctuation, then move to the next stretch. If the speaker switches to Tamil, the stenographer can either note the Tamil or write the English meaning instead. Whisper does the same with recordings: listen in chunks, write clean text.
Where the example breaks: A human stenographer would never write a sentence nobody said. Whisper sometimes invents words during silence or noise, so important transcripts still need checking.
How it actually works
Whisper is a transformer with two halves: one listens, the other writes.
Cut the audio into 30-second pieces
The recording is split into 30-second chunks. Each chunk becomes a log-Mel spectrogram: a picture of which pitches are loud at each moment, which the model reads instead of the raw sound wave.
The encoder listens
The first half, the encoder, reads the whole spectrogram and builds a rich summary of the sounds. This is the encoder-decoder design also used in machine translation: understand the input fully first, then write.
The decoder writes, token by token
The second half writes the text one token at a time, looking back at the encoder's summary. Special starting tokens tell it the task: which language, and whether to transcribe or translate into English.
Trained on loosely labelled data
Instead of small, carefully checked datasets, OpenAI used a huge amount of audio with transcripts found online. The paper calls this large-scale weak supervision: noisier labels, far more examples.
- Made by
- OpenAI, released 21 September 2022
- Training data
- 680,000 hours of speech paired with text, in many languages
- Sizes
- From tiny (39 million parameters) to large (1.55 billion), plus a faster turbo version (809 million)
- Licence
- Code and model weights under the MIT licence: free to use, change and build on, including commercially
The reverse job, turning text into a spoken voice, is text to speech.
Questions people ask
Is Whisper free?
The downloadable model is free. OpenAI released the code and weights under the MIT licence, so anyone can run Whisper on their own computer. OpenAI also offered it as a paid online service through its API, a version it announced in August 2026 it will retire on 26 February 2027.
Is Whisper the same as ChatGPT?
No. ChatGPT is a chatbot app that writes replies. Whisper is a separate model that only turns speech into text; it does not answer questions. Both come from OpenAI, and developers build Whisper into their own apps.
Does Whisper work with Hindi and other Indian languages?
Yes, but accuracy varies by language. OpenAI publishes word error rates for each language it tested, and the gaps between languages are large. For Hindi, Tamil or Bengali, check those figures and test on your own recordings before relying on it.
Can Whisper run offline?
Yes. Because the model weights are public, you can download Whisper and run it entirely on your own computer, with no internet. Smaller sizes run on ordinary laptops; the large version needs about 10 GB of graphics card memory.
Why does Whisper sometimes add words nobody said?
Its writing half predicts likely next words, so it can write plausible text that was never spoken, a kind of hallucination. A 2024 Cornell study found invented phrases in about 1% of transcripts, more often with long pauses. Check important transcripts.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change