Whisper
More info
- Creator
- OpenAI
- Released
- Type
- Automatic Speech Recognition (ASR)
- License
- MIT (open weights)
- Source
- github.com/openai/whisper
OpenAI's speech recognition model, released in September 2022 with open weights under the MIT licence. It made high-quality automatic transcription something anyone can run locally.
An encoder–decoder transformer trained on 680,000 hours of weakly labelled audio from the web — not carefully curated material, but enormous quantities of real speech with varying quality, noise and accent. That's why it's robust where earlier systems fell over. It handles some ninety languages, including strong accuracy on smaller ones, and can translate directly into English. Available in sizes from tiny to large-v3. The known weakness is hallucination during silence: the model invents sentences when nothing is being said, which investigations in 2024 showed occurred even in medical transcription. Faster variants: faster-whisper and whisper.cpp.