Volunteers recorded and checked thousands of hours of speech to train the model, which runs on an ordinary laptop.

Early users in newsrooms and schools say it cuts transcription time from hours to minutes, though accents outside the training data still cause errors.