Speech recognition and models
Whisper
An open speech recognition model from OpenAI that can be run on your own machine.
Whisper is a speech recognition model that OpenAI released openly in 2022. It was trained on very large amounts of audio in many languages, and it can be downloaded and run on your own machine without asking anyone’s permission.
That last part is the reason it matters so much.
Why it changed the field
Before Whisper, usable speech recognition was in practice a cloud service you bought access to. After Whisper, anyone could run a good model locally, with no cost per minute and without the audio leaving the machine.
That opened transcription up for everything that cannot be uploaded: patient conversations, client meetings, research data, personnel cases.
A model, not a service
Whisper is a model, that is, a file of weights. It has no user interface, it does not handle files for you, and it does not tell speakers apart. The software around it does that. When two tools both “use Whisper”, they can still give quite different results, depending on which version and size they run, and what they do before and after.
Smaller languages
Whisper is trained broadly and internationally, and each smaller language is one among many. It handles standard speech in the Nordic languages well, but does less well on broad dialect than models that have been further trained on one language specifically. NB-Whisper is exactly such a model for Norwegian.